Notes on Vemra Tech News - Latest Insights on AI, Software & Hardware

OpenAI Partners with Malta for National AI Literacy Initiative

Mon, 18 May 2026 00:00:00 +0000

OpenAI and Malta’s Groundbreaking Collaboration

OpenAI has officially announced a partnership with the government of Malta, providing all Maltese citizens with one year of free access to ChatGPT Plus, funded by the state. This initiative marks the world’s first national-level AI tool accessibility program.

Education Before Access

The program includes a prerequisite: citizens must complete an AI literacy course developed by the University of Malta, covering basic AI principles, capabilities, and responsible usage in both home and work environments. Only after completing the course will citizens qualify for the free year of ChatGPT Plus.

The first users will gain access starting in May, with the distribution managed by the Malta Digital Innovation Authority, eventually extending to Maltese citizens living abroad.

A Noteworthy Approach

This “education first, tools later” strategy stands out. Most countries’ AI policies focus on regulation—enacting laws, establishing ethics committees, and limiting data usage. Malta, however, provides tools directly to citizens while ensuring they understand how to use them responsibly.

What Does OpenAI Gain?

At first glance, this appears to be a subscription revenue opportunity. With a population of approximately 574,300, if every citizen subscribed to ChatGPT Plus at $20 per month, the total annual cost would be around $130 million—insignificant for OpenAI. The real benefits lie elsewhere.

User Scale

Sam Altman has often likened intelligence to electricity, and OpenAI reiterated this in its announcement, referring to AI as a “global utility.” The business logic of utilities is that scale is everything. Currently, ChatGPT has over 900 million weekly active users, but competitors like Claude, Gemini, and Grok are quickly capturing market share. Acquiring users through government channels is an efficient way to establish brand loyalty. The first AI tool someone uses is likely to become their long-term choice.

Data Feedback Loop

More users translate to more real-world interaction data, which feeds back into model training and product iteration. Every question, correction, and usage scenario on ChatGPT helps OpenAI understand the distribution of human needs. For a company aiming for AGI and ASI, the diversity of data is invaluable. The questioning patterns of Maltese teachers, fishermen, and civil servants differ significantly from those of Silicon Valley engineers—this is the training signal needed for a more generalized model.

Demonstration Effect

When OpenAI seeks to pitch its collaboration model to other countries, being able to say, “We have helped multiple nations achieve widespread AI accessibility” serves as a compelling reference. Malta acts as a prototype, with the real target audience being medium-sized countries that are still observing.

George Osborne, former UK Chancellor and current head of OpenAI for Countries, stated, “Intelligence is becoming a national utility… Malta is leading the way, and we hope other countries will follow.”

Which Countries Can Replicate This Model?

The financial feasibility of Malta’s plan heavily depends on its population size. A rough calculation shows that for a country with 1 million people, the total cost for full coverage would be around $240 million annually at $240 per person. For a country with 5 million people, the cost would be about $1.2 billion, and for 10 million people, it would approach $2.4 billion.

In reality, not everyone will complete the course and activate their accounts. Assuming an activation rate of 30%-50%, countries with populations under 5 million—like Estonia (1.36 million), Singapore (6.11 million), Luxembourg (682,000), and Iceland (392,000)—could financially manage this expense. OpenAI is already collaborating with Estonia and Greece on educational initiatives, making them likely candidates for early adoption.

For countries with populations over 10 million, the model of full government funding faces significant budgetary pressures. For example, Portugal (11 million people) would incur nearly $800 million annually at a 30% activation rate—an amount requiring substantial public discussion and justification for an economy with a GDP of approximately $290 billion. In larger economies like India, with 1.46 billion people, even covering just 10% would mean an annual expenditure exceeding $35 billion, surpassing India’s entire education budget for the 2026-2027 fiscal year (approximately $16.8 billion).

For populous nations, the path of “government-funded, universal coverage” is financially challenging. A more realistic approach might involve government subsidies for specific groups (teachers, civil servants, university students) or negotiating significant discounts on national licensing agreements with AI companies.

A Race to Build Infrastructure

The underlying logic of the Malta project resembles the historical development of electric and telephone networks: once infrastructure is established, the cost for later entrants to replace it is high. OpenAI’s strategy is that when AI becomes a daily essential tool, the first platform to cover users will gain a level of stickiness akin to that of an operating system.

The real question this experiment seeks to answer is whether the bottleneck for AI accessibility lies in the availability of tools or the capability to use them. If many Maltese citizens abandon the tool after completing the course, it indicates a lack of demand scenarios. Conversely, if usage rates continue to rise, OpenAI will have a strong case to present to more countries.

For those monitoring this field, one key metric to watch is the monthly active retention rate six months after the Malta project launches. This figure will determine whether “national funding for universal AI” is a replicable public service innovation or merely an expensive marketing campaign.

The Token Revolution in AI: China's Rising Influence

Mon, 18 May 2026 00:00:00 +0000

The Token Revolution in AI

In early 2026, a set of data sparked intense discussions in the global AI industry. OpenRouter, the world’s largest AI model API aggregation platform, reported that from February 9 to 15, the token call volume of China’s large models reached 41.2 trillion, surpassing the U.S. models’ 29.4 trillion for the first time. This lead continued for several weeks, with the volume exceeding 73 trillion by mid to late March, and four out of the top five models globally were from China.

This data is not meant to compare quantities but marks a quiet revolution in the basic measurement unit of the AI industry—tokens are becoming the “kilowatt-hour” of the intelligent era. The six dimensions of models, computing power, data, applications, industry, and governance are being profoundly reshaped by this measurement unit. Understanding AI in 2026 begins with understanding tokens.

Sixfold Reconstruction from a Measurement Unit

The measurement unit of the Industrial Revolution was the “kilowatt-hour,” allowing energy to be accurately measured, priced, and transmitted across domains. The Information Revolution’s units were “bits” and “bandwidth,” enabling information to be packaged, transmitted, and billed for the first time. The measurement unit of the Intelligent Revolution is “tokens,” allowing intelligence to be segmented, measured, priced, and traded for the first time.

The popularization of the token concept and its rapid growth in usage are gradually pushing intelligence towards industrialization, marketization, and circulation.

Models

From “training as an asset” to “inference as production.” The economic value of large models is shifting from one-time training costs to long-term inference outputs. Model vendors no longer simply “sell capabilities” but directly “sell tokens”—pricing based on millions of tokens for input and output has become a global industry norm. The asset attribute of models is transitioning from “weight files” to “the ability to continuously produce tokens.”

Computing Power

From “training computing power” to “inference computing power.” Training computing power is pulsed and centralized, while inference computing power is continuous and distributed, posing new demands on latency, energy efficiency, and geographical distribution. The collaboration of computing power across “cloud-edge-end,” inference-specific chips, silicon photonics interconnects, and computing power networks is becoming the new focus of infrastructure. JPMorgan predicts that China’s inference token consumption will grow by more than two orders of magnitude by 2030 compared to 2025.

Data

From “raw data” to “tokenized corpora.” Just as raw coal must be processed into standard fuel to generate power, data entering large models also needs to be cleaned, labeled, and tokenized. In long-tail scenarios such as autonomous driving, robot training, and scientific discovery, synthetic data generated through simulation has achieved large-scale application. The construction of data factor markets has also entered a substantial phase, with “trainability” and “token output density”—rather than mere data scale—becoming the new metrics for pricing data assets. This shift is profound: the evaluation of data value is now linked to its actual contribution in the token production chain, providing a more solid economic foundation for the market allocation of data factors.

Applications

From “function delivery” to “token consumption.” Traditional software charges based on seats and functions; today, applications bill based on token call volume and business results. Intelligent agents are becoming the main consumers of tokens, with complex tasks potentially consuming hundreds of thousands or even millions of tokens. The “intelligent agent as a service” market is rapidly expanding, with performance-based billing models being implemented at scale in customer service, marketing, compliance, and programming. The essence of applications is shifting from “delivering functions” to “consuming intelligence.”

Industry

From “software industry chain” to “token industry chain.” A new industry chain is forming around the production (models and computing power), distribution (inference networks, APIs, intelligent agent protocols), consumption (applications and intelligent agents), and measurement (evaluation benchmarks, auditing, and trusted verification) of tokens. The boundaries between model layers, inference service layers, intelligent agent middleware layers, and industry application layers are becoming increasingly clear, with industry-specific intelligent agents becoming mainstream investments. Model vendors, cloud vendors, chip manufacturers, green energy operators, and content delivery network providers are forming a collaborative ecosystem of the token industry chain. According to the China Academy of Information and Communications Technology, the scale of China’s core AI industry is expected to exceed 1.2 trillion yuan by 2026, with the effects of the entire industry chain collaboration becoming evident.

Governance

From “algorithm governance” to “full-chain governance of tokens.” As the AI industry has developed, the governance focus has expanded from “algorithms and code” to the entire chain of token production, circulation, consumption, and cross-border flow: traceability of tokens, identification of synthetic content, cross-border token flow, constraints on computing power and energy consumption, and trusted evaluation and benchmarks—all of these new issues call for new governance tools and rules. The year 2026 may become a key year for the concentrated implementation of global AI governance rules.

China’s Position in the Global Token Wave

In the global wave brought by tokens, China is forming a unique position supported by multiple factors.

On the production side, domestic model clusters are rising. A number of domestic models, such as MiniMax, Dark Side of the Moon, Deep Quest, Zhipu, Alibaba Qianwen, and Byte Bean, have leveraged mixed expert architectures and extreme engineering optimization to continuously improve performance while reducing inference prices to a fraction of comparable global models. On the OpenRouter platform, while U.S. users account for 47% and Chinese users only about 6%, the call volume is led by Chinese models—this is a recognition determined by global developers voting with their feet.

On the consumption side, applications are penetrating deeper than ever, with tokens entering daily life at an unprecedented speed. A general practitioner in a county hospital, faced with a suspicious lung CT scan, can have AI circle nodules and provide differential diagnosis suggestions in just a few seconds and thousands of tokens, compressing what used to take two weeks for a consultation into a single outpatient visit. A farmer in Shouguang, Shandong, can take a picture of a curled cucumber, and a smart agriculture app provides tokenized agricultural knowledge to identify whether it’s a thrips or a viral disease and what medication to use. An elderly person living alone can tell a smart speaker in their dialect, “I feel chest tightness,” and after a few thousand tokens of conversation, their children’s phones receive alerts and location sharing with emergency services. Delivery riders no longer hear mechanical instructions like “turn right ahead” but receive route planning based on real-time traffic and elevator wait times. AI assistants in government service halls are available around the clock to answer inquiries about medical insurance transfers, real estate registrations, and other policies, turning “people running errands” into “tokens running errands.” Tokens are becoming the “invisible labor force” across various industries.

At the industry chain level, a full-stack collaborative ecosystem is rapidly taking shape. From domestic chips like Ascend, Cambricon, and Haiguang to inference service platforms like Volcano Engine, Alibaba Cloud, and Tencent Cloud, along with a range of open-source middleware and industry-specific intelligent agents, the entire industry chain covering chips, computing power, models, middleware, and applications is rapidly improving. The “East Data West Computing” project provides low-cost computing power, while green energy directly supplies data centers, solidifying the energy foundation.

However, it is essential to recognize that there remains significant room for improvement in areas such as original model innovation, high-end computing power foundations, cross-language and cross-cultural ecological influence, and participation in global rule-making.

The second half of the token wave is not about “already winning” but rather “just beginning.” In the global landscape unfolding from the small token, China is both a massive market and should be an active builder and responsible co-governor. Understanding tokens is key to understanding the next phase of artificial intelligence.

Avoiding AI Chaos in Git Repositories: Strategies for Vibe Coding

Sat, 16 May 2026 00:00:00 +0000

Don’t Let Your Git Repository Become an AI “Battlefield”: The Border Wars and Reconciliation of Vibe Coding

Last Wednesday at 11 PM, our release window was still open. Three engineers, three feature branches, the same codebase, and—three unrestrained AI coding assistants. When we finally mustered the courage to execute git merge, the conflict prompts on the screen seemed endless: UserService.ts had been rewritten four times, database.ts sprouted three different connection pool strategies, and that poor utility function file had lost any trace of its original code. Our CI red line filled the entire monitoring screen. I didn’t sleep that night.

This is the true face of unrestrained Vibe Coding in a team environment. One person using AI to “code by feel” is freedom; a group doing so is a silent war. Most teams’ disasters do not stem from poor AI coding but from AI’s inherent lack of boundaries. It instinctively optimizes any code it sees, turning parallel development into an unpredictable refactoring tournament.

This article will not provide you with a one-size-fits-all silver bullet, but I will candidly share the practical lessons learned from that disaster—ranging from physical isolation to time compression, and rethinking the old adage of “shared code.” If you are leading a team attempting Vibe Coding, this will be a survival report from the battlefield.

1. AI’s “Boundary-Crossing Instinct”: Why It Goes Out of Control

The core mechanism of AI coding is: given context, predict the most likely correct next step. When you open the entire codebase as context on a branch and tell the AI, “Help me tidy up the user validation logic,” it won’t just modify auth/validator.ts. It will scan the call chain, find that the User type definition can be improved in seven places; it notices that AuthService’s exception handling is inconsistent with new requirements; it thinks that database queries can be parameterized to prevent injection—even if you never mentioned it.

What it does is precisely what we taught it: to be a “responsible” developer. However, in parallel development, this “responsible” behavior turns into a full invasion of others’ work. Worse still, this is not just a text conflict issue— even if a piece of code does not generate <<<<<<< markers, AI on two branches may make semantically incompatible changes to the same interface: one changes the return value from User | null to Result<User, Error>, while another changes it to always return User and throw an exception when missing. The merged code may run, but the behavior is completely wrong. This semantic conflict can only be discovered at runtime, often resulting in the most painful bugs.

Thus, setting boundaries is not just to prevent Git conflict markers but to prevent “silent destruction.”

2. Physical Boundaries: Putting an Invisible Collar on AI

The most effective and brute-force method is to make AI unable to see the code it shouldn’t modify. Modern monorepo tools (Nx, Turborepo, Bazel) provide precise project dependency graphs and folder tags, which we can use to build a “context compartment.”

1. Enforce Folder-Level Permissions with Nx Tags and CI

Assuming your repository structure is as follows:

libs/
  core/
    user/       # Public user module
  features/
    checkout/   # Checkout feature
    profile/    # Profile feature

Tag each library in nx.json:

{
  "projects": {
    "core-user": { "tags": ["scope:core", "type:shared"] },
    "features-checkout": { "tags": ["scope:checkout", "type:feature"] },
    "features-profile": { "tags": ["scope:profile", "type:feature"] }
  }
}

Then write a simple CI rule (which can be a custom ESLint rule or a Shell script) that checks on every PR: If a branch claims to be developing the checkout feature, it should never modify files tagged as scope:core or scope:profile, unless it explicitly declares “public module modification” and tags the PR title with [CORE-CHANGE].

Here’s how we do it in GitHub Actions:

- name: Check for unauthorized cross-scope changes
  run: |
    CHANGED=$(git diff --name-only origin/main...HEAD)
    for file in $CHANGED; do
      if [[ $file == libs/core/* ]] && ! echo "$PR_TITLE" | grep -q "\[CORE-CHANGE\]"; then
        echo "Error: Core module modified without [CORE-CHANGE] declaration."
        exit 1
      fi
    done

This may sound strict, but it is this “physical interception” that first made the team feel safe: you know your AI assistant cannot quietly rewrite the shared User module without alerting anyone. The boundaries are secure.

2. Narrow the AI Tool’s Working Context

At the start of Vibe Coding conversations, clearly limit the context. In Cursor or Copilot Chat, we can do this through .cursorrules or custom instructions:

You can only modify files in the `src/features/checkout` directory.
If you need to modify anything under `src/core`, you must pause and ask: "I need to modify the core module [name], reason: [explanation]. Is this allowed?"
No out-of-bounds modifications may occur without explicit confirmation.

At the same time, only open the current feature folder as the workspace in the IDE to prevent AI from scanning the entire project. Although this requires some adjustment, it significantly reduces the probability of AI “accidentally” refactoring public dependencies. We observed that requests for out-of-bounds modifications decreased by nearly 70%.

3. Architectural Boundaries: Rethinking the Cost of “Shared Code”

This experience forced us to reflect: is all that “public code” causing endless conflicts really worth sharing?

In traditional engineering, DRY (Don’t Repeat Yourself) is a virtue. However, in the era of Vibe Coding, over-pursuing DRY can create highly coupled shared modules that become battlegrounds for every AI branch. A generic utility library optimized by five people using AI often incurs merging conflict costs far exceeding the cost of a few lines of repeated code.

We began to accept a “WET moderation” strategy: If a piece of logic can evolve independently within two business boundaries and has low stability, allow duplication. For example, if two different feature panels need a “format currency” function, we would previously extract it to shared/utils/formatCurrency.ts. Now we consider: do their future evolution directions truly align? Perhaps one will soon need multi-currency support, while the other needs cryptocurrency support. Rather than letting AI create conflicts in a shared file, it’s better to let them each have a copy and evolve freely within their domains.

We call this model the “isolation chamber” model: vertically segmenting each business capability, allowing its own state management, utility functions, and even partial entity definitions. The shared core is drastically compressed to only include truly stable, rarely changed elements: for instance, enterprise-level unified HTTP client configurations or design system token values. This way, the conflict area of public code decreases exponentially.

This is not a regression but a pragmatic choice based on team topology: let the code structure reflect the team’s communication structure (inverse Conway’s law). Vibe Coding amplifies each developer’s “productivity radius”; correspondingly, we must reduce physical dependencies between modules.

4. Temporal Boundaries: Nipping Conflicts in the Bud

After spatial isolation, the temporal dimension is equally important. Our second key shift is: from long-running feature branches to extremely short branches + Trunk-Based Development.

In Vibe Coding, AI can generate the amount of code that a regular developer would take a day to write in just one hour. This means that if your branch lives longer than eight hours, it may already be unrecognizable and diverged significantly from the main line. Our current rules are:

The lifecycle of any feature branch should not exceed one day. If it takes longer than a day, the feature needs to be further split.
Within a day, you must merge the main line into your branch at least three times. This high-frequency merging transforms one large conflict into several small ones, and the cognitive burden of resolving small conflicts is much lower.
Use feature flags to merge incomplete features. Code entering the main line is not immediately exposed to users, allowing you to continue Vibe in the background.

We wrote a simple auxiliary script to run AI conflict simulations before merging the main line:

# Simulate merge, generate conflict preview
git merge-tree $(git merge-base HEAD origin/main) HEAD origin/main > /tmp/merge-preview.txt
# Let AI analyze the simulated merge conflicts and propose solutions
 aicmd "Please analyze the simulated merge conflicts in /tmp/merge-preview.txt and provide resolution suggestions."

Thus, before the actual merge, we already have a semi-automated draft of solutions. Conflicts are no longer sudden accidents but a prepared negotiation.

5. Cultural Boundaries: Establishing the Role of “Public Module Guardian”

Finally, no matter how good the tools and processes are, if the team’s mindset does not change, they will not be sustainable. When everyone can use AI to refactor code like a god, the hesitation of “Can I change this?” will be overwhelmed by the immense productivity rush. We must consciously establish a role we call the “core guardian,” rotating weekly.

The guardian’s responsibility is not to prevent modifications but to arbitrate and accelerate changes to public code. When someone genuinely needs to modify a shared module (for example, to support a new interface), they do not need to wait for lengthy approvals; they just need to initiate a brief “boundary change intent” in the team channel:

“I need to add a getUserPreferences method to the User module, with the following signature. It will be completed and merged within one hour. Guardian, please pay attention.”

The guardian will monitor this modification to ensure it stays within the contractual scope and notify other members to pull the latest changes after merging. Thus, changes to public modules are serialized and broadcasted, avoiding the disaster of parallel rewrites. Since each change is small and frequent, the waiting time is almost negligible.

We also added a “boundary check” to our daily stand-ups: quickly reviewing which branches might touch public modules today and coordinating the order verbally.

Conclusion: Dancing with AI, Not Fighting Against It

Since that terrifying night six months ago, our Git repository has not erupted into a hell of conflicts across hundreds of files. Not because we limited AI’s capabilities, but precisely the opposite: we provided it with a clearer, safer stage where it can flourish without stepping on others’ toes.

If you ask me what the solution to the boundary issues in multi-person Vibe Coding is, my answer is: make invisible boundaries visible constraints—through folders, tags, CI checkpoints, short branches, roles, and clear intent declarations, building fences in the physical, temporal, and social dimensions that everyone (including AI) clearly understands where “home” is and where “others’ territory” lies.

Don’t try to tame AI to prevent it from crossing boundaries; that is futile. What you should do is design a system it cannot cross. When boundaries are clear enough, Vibe can truly flow within the team without becoming the reason for weekly Wednesday night breakdowns.

Driving the Synergy of AI Industrialization and Intelligent Transformation

Thu, 14 May 2026 00:00:00 +0000

Driving the Synergy of AI Industrialization and Intelligent Transformation

Currently, the development of artificial intelligence (AI) is transitioning from technological breakthroughs to industrial shaping. China’s AI industry has a strong technical supply capability, but large-scale application still faces the “last mile” problem.

The 2026 Government Work Report proposed to “create a new form of intelligent economy,” indicating that AI development is moving from technological breakthroughs, product iterations, and scenario expansions to industrial shaping and system restructuring. To develop AI, it is essential to promote the industrialization of technology, products, platforms, and terminals, while also pushing AI deeper into manufacturing and other real economies to drive industrial intelligence. This requires a connection between technical supply and application demand, promoting mutual traction and coordinated efforts between AI industrialization and intelligent transformation, allowing AI’s technical advantages to continuously transform into industrial and developmental advantages.

AI industrialization and intelligent transformation are two key dimensions for strengthening the intelligent industry system and shaping the development pattern of the intelligent economy. The former focuses on transforming technology, products, platforms, and terminals into industrial supply capabilities, while the latter emphasizes promoting AI into real scenarios such as research and development, production, supply chain collaboration, and brand marketing, converting it into actual productive forces.

Relying solely on the AI industry itself makes it difficult to truly strengthen the intelligent economy. Without broad application scenarios, AI industrialization can easily stagnate at the levels of technical self-rotation, product internal competition, and hype-driven phases, making it hard to solidify into long-term industrial capabilities. Similarly, relying solely on industrial intelligence is also challenging. Industrial intelligence is not simply about adding equipment, connecting systems, or making modifications; it requires stable technical supply, product supply, and service supply to continuously support the formation of replicable and scalable systematic upgrading capabilities.

Coordinated efforts do not merely parallel the supply side and application side; they form a bidirectional shaping and cyclical reinforcement operating mechanism in the real industrial process. The closer technical supply is to industrial demand, the more vitality AI industrialization has, and the more continuous iteration can be formed through scenario feedback, allowing industrial intelligence to deepen and solidify.

This synergy is first reflected in the bidirectional traction between technical supply and scenario demand. AI industrialization cannot be an insular process; technology must effectively enter real industrial development scenarios to accelerate iteration through problem discovery and value verification. Scenarios can only continuously enhance application levels and optimize or even reshape industrial processes based on higher quality algorithms, data, toolchains, and terminal supplies. The deeper the integration of technology and scenarios, the stronger the support between industrialization and intelligence.

Secondly, the synergy is also reflected in the mutual promotion of project transformation and industrial diffusion. Once technology enters scenarios, it can transition from technology validation to product refinement, from single-point trials to stable services, and from sporadic applications to industry solutions. Once projects possess sustainable transformation capabilities, they will solidify product capabilities, service capabilities, and market capabilities, which in turn promotes the large-scale development of the AI industry itself. Moreover, the deeper layer of synergy is manifested in the mutual shaping of entity growth and industrial ecology. The bidirectional connection between supply and application not only drives the growth of technical service entities, solution entities, and scenario transformation entities but also promotes traditional enterprises to complete capability upgrades, organizational restructuring, and business model updates.

To promote the synergy of both ends, high-quality supply must genuinely target high-value scenarios. AI industrialization should enhance supply capabilities around the real demands of industrial intelligence. High-end chips, basic software, industrial algorithms, multimodal models, high-quality datasets, as well as intelligent terminals and open-source ecosystems, ultimately need to be assessed on their ability to enter industrial sites, solve practical problems, and support large-scale applications. High-value scenarios should also drive high-quality supply during project advancement. Industrial intelligence must continuously raise new questions, form new demands, and test new technologies in real scenarios. It is essential to cultivate diverse entities to form a stable transformation chain, strengthening AI companies while also nurturing solution service providers, scenario transformation entities, and platform collaboration entities, while promoting traditional enterprises to enhance their capacity for intelligence.

To drive the synergy of AI industrialization and intelligent transformation, it is necessary to build a solid industrial foundation with robust technical supply and to use rich application scenarios to drive technical iteration, project transformation, and entity growth, accelerating the formation of a development pattern that promotes production through intelligence and nurtures intelligence through production. The more thoroughly the two ends are connected, the stronger the endogenous driving force for intelligent industry development will be, allowing AI to continuously transform into a new driving force for high-quality development.

Codex AI Achieves 40x Research Efficiency in Groundbreaking Experiment

Wed, 13 May 2026 00:00:00 +0000

Introduction

Today, Agentic AI engineers discovered that a research task requiring 80 hours for a PhD can be completed by Codex in less than 2 hours, achieving a staggering 40-fold efficiency increase! According to previous standards, AGI has already existed; the entire industry has simply been moving the goalposts.

The “singularity” in the research community is indeed closer than everyone anticipated.

Recently, an experiment involving Codex’s Goal Mode shocked the academic world: Codex can increase AI research efficiency by 40 times!

Agentic AI engineer Dan McAteer recently disclosed an experiment on X, using OpenAI Codex’s Goal Mode to run a mechanistic interpretability research task.

GPT-5.5 estimated that a PhD student would take about 80 hours to complete this task, but in practice, the AI finished it in just 1 hour and 56 minutes.

This represents an apparent efficiency boost of about 40 times!

The built-in skill used in Codex is /goal.

The author believes:

/goal + gpt-5.5 high precision + fast mode is the most efficient AI agent configuration today.

This means allowing the model to set its own goals, where the key is that the prompts it generates are likely better than yours.

This is no longer just a simple “efficiency improvement”; it is a complete “dimensionality reduction attack.”

As research cycles shrink from weeks to hours, and AI begins to autonomously draft its own experimental goals (/goal), we must confront a harsh reality:

The slope of the “intelligence explosion” has already emerged, and the speed of AI’s self-iteration is departing from human control!

What is Codex /goal Mode?

Let’s take a look at how this experiment was conducted.

The experiment was initiated by Dan McAteer, an Agentic AI engineer and former Amp Code engineer, who frequently shares practical experiences of AI agent engineering on X.

His experimental setup was simple:

Tool: OpenAI Codex /goal command
Model: GPT-5.5 high
Mode: fast mode
Task: A research task in the direction of Mechanistic Interpretability

He describes this configuration as the most efficient AI agent configuration currently available.

Why is Codex /goal Important?

What truly deserves attention is the /goal mode itself.

According to OpenAI Codex engineer Philip Corey, /goal is our implementation of the Ralph loop—allowing goals to persist across multiple dialogues, not stopping until achieved.

In simple terms, a standard Codex call is you say a sentence, it takes one step, and responds. Codex /goal allows you to state a goal, and it autonomously breaks down sub-tasks, executes them, reviews results, and continues until it either succeeds or fails.

This represents a shift from conversational AI to goal-driven AI.

For research tasks like Mechanistic Interpretability, the /goal mode is naturally well-suited.

The research process itself involves proposing hypotheses, designing experiments, running them, observing results, refining hypotheses, and re-experimenting—a perfect loop for a self-cycling agent.

McAteer’s experiment truly demonstrates the usability of the Codex /goal mode in cyclical research tasks: it does not replace researchers but rather replaces the repetitive operations performed by researchers.

If this capability can stabilize, it will have a very direct leverage on AI research itself.

It means that AI researchers within AI labs could one day use AI agents for repetitive tasks such as preparing training data, setting up experiments, conducting ablation studies, generating visualizations, and preliminary result analysis.

This aligns with what Anthropic and OpenAI have repeatedly stated: AI is accelerating AI research itself.

PhD 80 Hours vs AI 2 Hours

In the traditional research context, a PhD student’s daily routine involves reviewing literature, building models, debugging code, validating results, and writing reports.

This lengthy process is due to the physical limits of the human brain when processing complex logic and vast amounts of data.

However, Codex’s recent experiment completely shatters this perception.

Under the strongest agent configuration of /goal + GPT-5.5 High + Fast Mode, AI is no longer a tool that “follows commands” but an independent researcher that “generates strategies.”

It can understand complex natural language auto-encoder (NLA) experimental requirements, autonomously decompose tasks, and complete in less than 2 hours what human elites would take two weeks to accomplish.

This signifies that the threshold for human research has completely collapsed. The professional analytical capabilities that once required years of study are now being modularized by algorithms.

Moreover, autonomous AI researchers have already arrived ahead of schedule!

OpenAI previously set a goal for achieving autonomous AI research by the end of 2026. However, based on current experimental progress, 2026 may not be the beginning but rather the endpoint where humanity completely hands over the research baton.

Evidence of Recursive Self-Improvement Emerging

If Codex’s 40x speed experiment is a glaring case, what is even more unsettling is the growing evidence surrounding “recursive self-improvement.”

On May 7, Axios reported that Anthropic co-founder Jack Clark publicly provided a probability:

By the end of 2028, the probability of AI achieving complete recursive self-improvement exceeds 60%.

Sakana AI and UBC’s research team this year developed the Darwin Gödel Machine, a programming agent capable of rewriting its own source code to enhance its capabilities.

In SWE-bench, its score improved from 20.0% to 50.0% without any human intervention.

The same team’s AI Scientist project was published in Nature in March this year.

It can independently generate research ideas, write code, run experiments, draft complete papers, and conduct peer reviews.

A complete research pipeline, from start to finish, is accomplished independently by AI.

Now, let’s look at a set of hard data. GPQA Diamond, a scientific question-answering benchmark set by PhD experts, saw GPT-4 score 39% in November 2023, while the average score of human domain experts was about 65%.

By April 2026, cutting-edge models collectively surpassed the threshold: Gemini 3.1 Pro scored 94.3%, while Claude Opus 4.7 scored 94.2%.

All cutting-edge models have far outpaced human PhD experts.

The trajectory of SWE-bench further illustrates the acceleration.

At the end of 2023, Claude 2’s pass rate was 2%. Now, it stands at 93.9%.

In just two and a half years, it skyrocketed from 2% to 93.9%.

This curve, once drawn, is recognizable to anyone who has studied high school mathematics.

Clearly, the process of recursive self-improvement (RSI) has already begun.

Once AI starts rewriting its underlying code and optimizing its architecture at this 40x efficiency, the growth of intelligence will no longer be linear but vertical.

AGI Has Been Delivered, and the Entire Industry is Gaslighting You

In fact, as early as February this year, four scholars from different top fields jointly published a paper that can be described as the “most unsettling of the year”: “AGI Case Study: Today’s LLMs Have Met the Criteria.”

The four authors represent the four pillars of contemporary intelligence: philosophy, machine learning, linguistics, and cognitive science. They reached a chilling consensus:

According to definitions prior to 2022, AGI has already been achieved.

The reason no one acknowledges it now is that the entire AI industry is engaging in a collective “gaslighting effect” against the public.

The paper pointed out that humans exhibit a strong “psychological defense mechanism” when faced with the rise of AI.

Before 2022, as long as a model could pass the Turing test and handle tasks across domains, it was considered AGI.

With the emergence of ChatGPT: “Just having these capabilities is not enough; it must also have perfect reasoning, embodiment, and self-awareness.”

Each time a model breaks through a barrier, humans spontaneously add new, elusive criteria as thresholds, continuously moving the goalposts.

The problem is, if AGI already exists, the current industry logic becomes extremely absurd.

OpenAI is still raising $40 billion claiming to “build AGI”; Anthropic packages each new model release as a futures contract “close to AGI.”

The paper sharply reveals that the giants are disguising something that has already been “sold to you” as a miraculous achievement “soon to be developed” to secure a continuous flow of funding and power.

The Eve of the Intelligence Explosion

Today, we find ourselves at an extremely strange juncture.

In laboratories, AI is already conducting mechanistic interpretability research at 40 times the speed, even helping itself write code.

In the market, computing power remains a hard currency, with Nvidia’s Blackwell chips being snatched up, each chip accelerating the arrival of that singularity.

However, in social psychology, the public is still using outdated terms like “repeater” and “probability prediction” to comfort themselves.

If 40 times the research efficiency becomes the norm, the accumulated knowledge of human civilization over thousands of years could be doubled by AI in just a few months.

When AI can independently complete PhD-level tasks, our existing education systems, title evaluations, and even the very meaning of the term “expert” will face existential threats.

Just as Copernicus removed Earth from the center of the universe, AI is now displacing humanity from the sanctum of being the “only intelligent life.”

Now, this war called the intelligence explosion is happening without gunpowder.

We must either learn to coexist with this new intelligent species or watch helplessly as it leaves us in the dust at 40 times the speed.

AI Terminals Introduce New National Standards

Mon, 11 May 2026 00:00:00 +0000

Introduction

On May 8, 2026, the Ministry of Industry and Information Technology, the State Administration for Market Regulation, and the Ministry of Commerce jointly released the series of national standards titled “Intelligent Classification of AI Terminals” (GB/Z 177—2026). These standards specify the requirements for various products, including smartphones, computers, televisions, glasses, automotive cockpits, speakers, and headphones.

Experts believe that these standards clearly define the intelligence levels of AI terminals, laying a solid foundation for building a safe, orderly, and efficient ecosystem for AI terminals. This will also promote the coordinated development of China’s AI terminal industry, achieving scale advantages and standard leadership.

Diverse Product Forms

AI terminals are key carriers for the large-scale implementation and systematic development of AI technology. In recent years, China’s AI industry has flourished, with AI terminals continuously generating new products, business models, and experiences driven by diverse intelligent scenarios. This has effectively stimulated consumer enthusiasm and become a crucial lever for tapping into domestic demand and optimizing consumption structure.

This year, driven by the expansion of the old-for-new consumption policy and the deep integration of AI technology with consumer products, AI terminals have gained significant popularity among consumers. Data shows that in the first quarter, China’s smartphone production reached 298 million units, a year-on-year increase of 6.9%, while service robots exceeded 4.4 million units, up 2.6%.

Wei Ran, chief engineer of the China Academy of Information and Communications Technology, explained that AI terminals, driven by large models, represent a new generation of intelligent terminals. Compared to traditional terminals, they feature four major functional upgrades: the ability to actively perceive scenarios, accurately understand user intentions; support for multimodal interactions including text, voice, and audio-video; capability for generative applications and intelligent agent services; and autonomous learning and continuous evolution based on personal large models and knowledge bases.

“Overall, intelligent terminals have evolved from traditional passive execution tools to perceptive, understanding, service-oriented, and growth-capable intelligent assistants, redefining the human-computer interaction relationship. These functionalities are core to the highest level of intelligent terminals evaluated in the new standards,” Wei said.

Currently, AI terminals exhibit a rich variety of forms, with traditional terminals upgrading, emerging terminals expanding, and future terminal explorations progressing in parallel. Traditional terminals like AI smartphones, computers, and tablets have surpassed ten million units in shipments, becoming market leaders. Emerging categories such as intelligent vehicle terminals, smart glasses, and intelligent toys are rapidly growing, while native terminal forms represented by embodied intelligence continue to explore, further accelerating the application of AI.

Wei analyzed that the systematic integration of AI and terminal technology necessitates breakthroughs in three key areas: optimizing the end-cloud collaborative architecture, deepening the full-stack upgrade of hardware and software, and enhancing the security and privacy protection system.

Clear Evaluation System

Since 2023, leading companies in the smartphone and computer supply chains have actively launched AI terminal-related products, each with varying functional focuses. The lack of definitions and classification standards for AI terminals has made it difficult for consumers to accurately assess the intelligence levels of different products and has complicated product development and market positioning for companies. The industry lacks a unified consensus on terminal intelligence classification, leading to concept generalization and misuse, with some products falling into parameter stacking, disconnecting functionality claims from actual experiences.

The series of national standards for “Intelligent Classification of AI Terminals” adopts a “2+N” framework. The “2” refers to “Part 1: Reference Framework” and “Part 2: General Requirements,” which clarify the concept of intelligence, level classification, and testing methods, serving as the foundation for all category standards. The classification system for terminal intelligence ranges from L1 response level, L2 tool level, L3 auxiliary level, to L4 collaborative level, with increasing intelligence levels. The L4 collaborative level will be further clarified and improved in subsequent revisions according to industry development levels. The “N” represents specific standards for different products such as smartphones, computers, televisions, glasses, automotive cockpits, speakers, and headphones. The first batch includes standards for seven categories, with plans to develop standards for additional categories in the future.

Li Hongwei, chief engineer of the China Electronic Information Industry Development Research Institute, stated that the biggest highlight of this series of standards is its scene-based, quantifiable approach that balances end-user and cloud considerations, covering scenarios such as office work, learning, and design. This provides a unified “health check standard” for AI terminals, standardizing industry development and facilitating clear purchasing decisions for consumers.

The series of standards provides a scientific and unified evaluation system for the large-scale application and intelligent classification management of AI terminal products in China. This will help regulate market order and enhance user experience while accelerating the innovation and upgrade of AI terminal technology products. Additionally, the standards will strengthen China’s voice in the global standard-setting for AI terminals, reducing technical barriers for companies going abroad and enhancing international competitiveness.

“On one hand, the standards provide companies with benchmarks for improvement, helping them supply high-end products, enhance resource utilization efficiency, and promote orderly competition and healthy development. On the other hand, they offer consumers technical and evaluation bases, ensuring that demand-side has standards to rely on, enabling better choices of intelligent products and enhancing user experience and satisfaction,” said Yu Xiuming, deputy director of the China Electronic Technology Standardization Research Institute.

Accelerating Technological Inclusivity

Lenovo Group participated in the development of these standards. Currently, AI PCs account for over 30% of Lenovo’s PC shipments. Its built-in personal super-intelligent agent, Tianxi AI, is progressing towards becoming a “dedicated super assistant” for personal users. Abulikemu, vice president of Lenovo Group, stated that Lenovo will actively implement national standards, continuously innovate terminal products around Tianxi AI, refine terminal innovation application scenarios and user experiences, and drive collaborative innovation across the industry chain.

To promote the innovative development of the AI terminal industry, the Ministry of Industry and Information Technology will strengthen the implementation of standards, conduct standard interpretations and specialized training, establish a compliance testing platform, encourage leading companies to take the lead, and create standard application demonstration cases and benchmark products. The ministry will accelerate the iteration of the standard system, optimize and improve standard content, continuously expand the coverage of standards, and accelerate the creation of a unified standard system that includes various terminal forms. This will stimulate consumer-leading effects and ensure the effective implementation of the old-for-new policy in consumer goods this year, forming a catalog of AI terminal products to guide public consumption decisions and broaden the depth and breadth of AI applications, creating hot consumption scenarios.

Yu Xiuming explained that the standard categories will continue to be enriched, developing more standards for wearable devices, home appliances, and trendy toys, ensuring that various terminal intelligence classifications have standards to rely on. This will provide standard and technical support for the implementation of national policies and offer standard consulting and product evaluation services to society, assisting in the high-quality development of the industry.

New AI Standards Set for Smart Devices in China

Mon, 11 May 2026 00:00:00 +0000

Introduction

On May 8, the Ministry of Industry and Information Technology, the State Administration for Market Regulation, and the Ministry of Commerce jointly released the series of national standards titled “Artificial Intelligence Terminal Intelligence Level Classification” (GB/Z 177—2026). These standards specify the requirements for various products, including smartphones, computers, televisions, glasses, car cockpits, speakers, and headphones.

Diverse Product Forms

AI terminals are key carriers for the large-scale implementation and systematic development of AI technology. In recent years, China’s AI industry has flourished, with AI terminals continuously giving rise to new products, business models, and experiences driven by diverse intelligent scenarios. This has effectively stimulated consumer enthusiasm, becoming an important lever for tapping into domestic demand and optimizing consumption structure.

Since the beginning of this year, driven by policies promoting the replacement of old consumer goods and the deep integration of AI technology with consumer products, AI terminals have gained popularity among consumers. Data shows that in the first quarter, China’s smartphone production reached 298 million units, a year-on-year increase of 6.9%; service robot production exceeded 4.4 million units, a year-on-year increase of 2.6%.

Wei Ran, Chief Engineer of the China Academy of Information and Communications Technology, explained that AI terminals are a new generation of intelligent terminals driven by large models. Compared to traditional terminals, they have four major functional upgrades: the ability to actively perceive scenarios and accurately understand user intentions; multimodal interaction capabilities including text, voice, and audio-video; support for large model generative applications and intelligent agent services; and the ability to achieve autonomous learning and continuous evolution based on personal large models and knowledge bases.

“Overall, intelligent terminals have upgraded from traditional passive execution tools to intelligent assistants that can perceive, understand, serve, and grow, redefining human-computer interaction. These features are core points of examination for the highest-level terminals in the new intelligence classification national standards,” Wei said.

Currently, AI terminals are flourishing, showcasing a parallel evolution of traditional terminal upgrades, emerging terminal expansions, and future terminal explorations. Traditional terminals have first upgraded to AI terminals, with shipments of AI smartphones, computers, and tablets surpassing ten million units, becoming the current market leaders. Emerging categories such as intelligent in-car terminals, smart glasses, and intelligent toys are rapidly growing, while native terminal forms represented by embodied intelligence continue to explore, further accelerating the application of AI.

Wei analyzed that the systematic integration of AI and terminal technology requires breakthroughs in three major directions: optimizing the end-cloud collaborative architecture, where the cloud handles high-complexity tasks and the end processes high-frequency real-time interaction tasks; deepening the full-stack upgrade of hardware and software, strengthening core capabilities of computation, storage, and perception on the hardware side, while promoting AI capabilities from the application layer to the operating system layer on the software side; and upgrading the security and privacy protection system to solidify data security and privacy protection barriers on the terminal side, ensuring that terminal services are trustworthy and controllable throughout.

Clear Evaluation System

Since 2023, leading enterprises in the smartphone and computer industry chain have actively launched AI terminal-related products, each focusing on different functional implementations. The absence of definitions and classification standards for AI terminals has made it difficult for consumers to accurately assess the intelligence levels of different products, complicating product development and market positioning for enterprises. The industry lacks a unified consensus on terminal intelligence classification, leading to generalized misuse of concepts, with some products falling into parameter stacking and a disconnect between functional promotion and actual experience.

The series of national standards for “Artificial Intelligence Terminal Intelligence Level Classification” adopts a “2+N” framework. The “2” refers to “Part 1: Reference Framework” and “Part 2: General Requirements.” These two standards clarify the concept of intelligence, level classification, and testing methods, serving as the foundation for all category standards. The intelligence classification system ranges from L1 response level, L2 tool level, L3 assistant level to L4 collaborative level, with intelligence levels increasing sequentially. The L4 collaborative level will be further clarified and improved in subsequent revisions based on industry development levels. The “N” refers to specific standards for different products such as smartphones, computers, televisions, glasses, car cockpits, speakers, and headphones. The first batch of standards includes seven categories, with plans to advance the development of standards for other categories in the future.

Li Hongwei, Chief Engineer of the China Electronic Information Industry Development Research Institute, stated that the biggest highlight of this series of standards is its scenario-based, quantifiable approach that considers both end and cloud, covering scenarios such as office work, learning, and design. This provides a unified “health check standard” for AI terminals, standardizing industry development while allowing consumers to clearly select and confidently use products.

The series of standards provides the industry with a scientific and unified evaluation system, offering key support for the large-scale application and intelligent classification management of AI terminal products in China. This helps regulate market order and enhance user experience. Additionally, it can accelerate the innovation and iterative upgrading of AI terminal technology products, accurately guiding technological research and development directions, and ensuring the healthy and sustainable development of the industry. Furthermore, the introduction of these standards will enhance China’s voice in the global arena of AI terminal industry standard-setting, lowering technical barriers for enterprises going abroad and improving international competitiveness.

“On one hand, the standards provide enterprises with directions for improvement to meet benchmarks, facilitating the supply of high-end products, enhancing resource utilization efficiency, and promoting orderly competition and healthy development; on the other hand, they provide consumers with technical and evaluation bases, ensuring that demand-side has standards to rely on for better selection of intelligent products, thus enhancing user experience and satisfaction,” said Yu Xiuming, Deputy Director of the China Electronic Technology Standardization Research Institute.

Accelerating Technological Inclusivity

Lenovo Group participated in the preparation of these standards. Currently, AI PCs account for over 30% of Lenovo’s PC shipments. Its built-in personal super intelligent agent, Tianxi AI, is progressing towards becoming a “personal super-powered partner” for users. Abulikemu, Vice President of Lenovo Group, stated that Lenovo will actively implement national standards, continuously innovating terminal products around Tianxi AI, refining terminal innovative application scenarios and user experiences, and driving collaborative innovation among upstream and downstream partners in the industry chain to promote the high-quality development of the AI terminal industry and accelerate the inclusivity of AI terminals.

To promote the innovative development of the AI terminal industry, the Ministry of Industry and Information Technology will strengthen the implementation of standards, conduct standard interpretations and specialized training, establish a compliance testing platform for standards, encourage leading enterprises to take the lead in trials, and create demonstration cases and benchmark products for standard applications. They will accelerate the iteration of the standard system, optimize and improve standard content, continuously expand the coverage of standards, and speed up the construction of a unified standard system that includes various terminal forms. They aim to stimulate consumer leading effects, effectively implement standards in this year’s consumer goods replacement policy, and accelerate the formation of a catalog of AI terminal products to guide public consumption decisions, expanding the breadth and depth of AI applications and creating popular consumption scenarios.

Yu Xiuming mentioned that they will continue to enrich standard categories, developing more standards for wearable devices, home appliances, and trendy toys, ensuring that various terminals have clear intelligence classification standards. This will provide standard and technical support for the implementation of national policies and offer standard consulting and product testing services to society, contributing to the high-quality development of the industry.

AI: Striving to Become a Trusted 'Future Advisor'

Thu, 07 May 2026 00:00:00 +0000

AI: Striving to Become a Trusted ‘Future Advisor’

What does the concept of “predictive technology” look like? When the foundational capabilities of general large models, the precision of specialized predictive models, the practical value of external tools, and the assurance of trustworthy mechanisms are organically integrated, AI will gain a new insight into the future. This will position AI as a trusted “future advisor” in critical areas such as financial risk control, weather forecasting, public governance, and industrial production, providing intelligent support for humanity to grasp future trends and becoming an important force in empowering social development and serving national governance modernization.

Four Technical Paths for ‘Predicting the Future’

Faced with increasingly complex predictive demands in the real world, researchers have developed two core lines and four specific technical paths around large model predictive technology. These paths are not competing alternatives but complement each other in different scenarios, collectively constructing a complete research framework for large model predictions.

The essential difference between the two core lines lies in whether a dedicated model is tailored for the prediction task: one path is to “borrow a boat to go to sea,” skillfully utilizing existing mature large language models for predictions; the other is to “build a ship for long voyages,” reconstructing dedicated foundational models for predictions. Both paths advance simultaneously, adapting to diverse task requirements.

Directly invoking large language models is the easiest entry point for large model predictions. Researchers convert various prediction tasks into common natural language questions, providing historical information, event backgrounds, and constraints for the model to directly assess future trends and output predictions. This approach has a low barrier to entry and does not require significant modifications to the model, merely changing the application of existing tools, allowing for impressive performance in news event analysis and business trend assessments. However, it falls short in high-precision numerical predictions required in fields like meteorology and finance due to the inherent limitations of large language models in numerical computation and factual output accuracy.

Time series tokenization modeling represents a cross-domain “intelligent borrowing.” It cleverly introduces classic natural language processing ideas into time series data analysis, transforming continuous time series data into token representations similar to words in language through discretization, scaling, and quantization techniques, and then training using a language model-like architecture. The representative model, Chronos, achieves probabilistic predictions and cross-dataset generalization by mapping time series to a fixed vocabulary, significantly reducing development costs. However, this convenience comes at a cost, as the data transformation process inevitably leads to loss of numerical details and quantization errors, akin to a rough polishing of fine parts, impacting prediction accuracy.

Building dedicated time series foundational models marks a shift in large model predictive research from “borrowing strength” to independent innovation. Researchers no longer view time series simply as pseudo-text but design pre-training schemes and model architectures tailored to the essential laws of time series data and the core needs of prediction tasks. Google’s TimesFM employs a decoder architecture, showcasing strong zero-shot prediction capabilities; Lag-Llama, developed by multiple universities and research institutions in the U.S., focuses on probabilistic predictions and cross-domain generalization; and Moirai, developed by an American AI company, boldly attempts to adapt to more scenarios with a unified training approach. These models act as customized “armor” for prediction tasks, more closely aligned with the characteristics of prediction tasks, achieving higher precision numerical predictions and becoming the preferred choice for high-precision prediction scenarios.

Reprogramming large language models and multimodal integration provide low-cost thinking for large model predictions. Research related to Time-LLM confirms that it is unnecessary to retrain massive time series models with hundreds of billions of parameters; simply reprogramming to align time series with text prototypes allows “frozen” large language models to participate in prediction tasks. This approach opens a feasible channel for the general large model + specialized adaptation technical route, further promoting the deep integration of text, numerical, and contextual knowledge modeling, making predictions more aligned with the complex and variable demands of real-world prediction scenarios.

These four technical paths do not have absolute advantages or disadvantages; they are like different keys fitting different locks. When prediction tasks require combining general knowledge and textual context for open trend judgments, routes related to large language models act as universal keys with greater advantages; when tasks pursue high-precision numerical outputs and stable cross-domain generalization capabilities, dedicated time series foundational models become the customized keys for precise matching. They support and achieve each other under different research resource conditions and actual task requirements, jointly promoting the steady advancement of large model predictive technology.

Moving Towards Real Application Scenarios

In the research arena of large model predictive technology, international studies have started earlier and developed a more systematic technical framework, delving deeper into foundational research and frontier exploration. Although domestic research started later, it has rapidly caught up, forming unique advantages in scenario adaptation, open-source ecology, and application implementation.

International academic research on large model predictions has evolved from text reasoning to diverse predictions. Early studies mainly focused on using large language models for text reasoning and event development judgments, akin to meticulous cultivation in a small domain. In recent years, it has gradually broken boundaries, expanding into time series, spatiotemporal data, and even scientific predictions, initiating a new phase of “expanding territory.” In the more complex field of scientific predictions, Microsoft’s ClimaX has pioneered the establishment of a foundational model framework for weather and climate tasks, while Aurora, also developed by Microsoft, extends the foundational model concept to the Earth system, capable of simultaneously handling various prediction tasks such as weather, air quality, and wave forecasts, akin to equipping the Earth with an intelligent early warning system, showcasing the immense potential of scientific foundational models in complex system predictions.

Notably, the international academic community maintains a rational and cautious attitude toward the predictive capabilities of large models. Related studies have found that the excellent performance of large models in standardized tests does not equate to reliability in predicting real-world future events—GPT-4, for instance, performed worse than the median human group in open-world prediction competitions. Addressing this core issue, international researchers have conducted competition studies, retrieval enhancement studies, and uncertainty detection studies, allowing international research to form a distinctive characteristic of “model capability enhancement + prediction result validation + trustworthy mechanism construction,” laying a solid foundation for the practical application of technology.

Domestic research, relying on the rapid development of general large models, has achieved impressive latecomer advantages, gradually forming a virtuous development pattern of rapid iteration of general large models, systematic review research, and steady progress in application implementation. In the arena of general model ecosystem construction, various participants showcase their strengths: Qianwen 3 has established a complete system in multilingual support and reasoning efficiency optimization, akin to building a multilingual intelligent bridge; DeepSeek-V3 has achieved breakthroughs in high-performance open-source models, making core technologies more accessible; Wenxin 4.5 continues to refine multimodal integration and engineering deployment, consistently aligning with practical application needs. Although these general large models are not solely focused on prediction, they provide a solid capability foundation for domestic large model predictive research, enabling researchers to conduct more targeted studies while standing on the shoulders of “giants.”

In terms of application implementation, domestic efforts are actively exploring ways to bring large model predictive technology out of the “ivory tower” and into real-world applications across various industries. Some studies deeply integrate expert knowledge with large language models for strategic early warning, accurately achieving trend judgments and risk identification in complex situations; others closely combine large models with meteorological monitoring data to enhance the accuracy and timeliness of short-term precipitation predictions. Although these studies are not entirely equivalent to pure numerical time series predictions, they signify that domestic large model predictive technology is transitioning from theoretical discussions to practical applications, beginning to explore technology paths that meet local needs and align with industry realities.

Overall, while foreign research has delved deeper into the development of specialized foundational models for predictions and scientific predictions, forming a relatively complete technical system akin to excavating extensive tunnels underground, domestic research has showcased distinctive features in adapting to Chinese scenarios, building low-cost open-source ecosystems, and implementing industry applications, akin to constructing high-rise buildings that fit local contexts above ground. As high-quality time series data and industry-specific data continue to accumulate in China, along with the gradual improvement of specialized evaluation systems, domestic foundational models for prediction tasks still have immense potential for enhancement and will undoubtedly contribute unique and valuable Chinese wisdom to the development of global large model predictive technology.

Bridging the Gap from ‘Powerful to Trustworthy’

Compared to traditional predictive methods, large model predictive technology has achieved a profound transformation from “point calculations” to “comprehensive judgments,” evolving from cold mechanical calculation tools into intelligent agents capable of understanding context, weighing factors, and providing rational judgments. This unique capability stems from its inherent core advantages but is also akin to a growing star, steadily evolving toward “trustworthiness,” striving to become a reliable “future advisor” for humanity.

The core advantage of large model predictive technology is its innate exceptionalism, particularly prominent in practical applications. Firstly, it has strong cross-task transfer capabilities. Traditional agricultural yield prediction models cannot be directly applied to stock market trend analysis; switching domains requires starting from scratch. In contrast, large models leverage their general representation capabilities from large-scale pre-training to quickly adapt to different fields such as agriculture, finance, and industry with minimal samples. Secondly, it has great potential for handling complex dependency relationships. For instance, predicting river water levels during flood seasons is influenced by multiple factors such as rainfall, upstream flooding, and topography, which traditional models struggle to capture; however, time series foundational models can learn patterns within contextual ranges, akin to possessing “keen insight” to see the connections behind the data. Thirdly, it excels in multi-source information fusion. Traditional meteorological predictions rely solely on numerical monitoring data, while large models can integrate satellite cloud images, meteorological textual reports, geographic information, and other multi-source content, transforming predictions from “viewing a leopard through a tube” to “panoramic observation.” Lastly, it offers excellent predictive explanation and decision support capabilities. It can not only predict the trend of a specific stock but also explain the influencing factors behind it, such as industry policies and market supply-demand dynamics, even providing risk control suggestions, becoming a professional intelligent partner for decision-makers.

Despite its significant advantages, large model predictive technology is not without flaws; there remains a gap that urgently needs to be bridged from the laboratory to real-world application scenarios. Firstly, the model’s generation and reasoning capabilities do not equate to actual predictive capabilities. Some models perform excellently in simulated meteorological prediction tests but often “fail” in real severe convective weather warnings, simply because the test answers are embedded in the training data, while real predictions require comprehensive assessments of unobserved events—it’s easy to theorize, but challenging to execute. Secondly, retrieval enhancement addresses symptoms rather than root causes. Although pairing models with information retrieval improves prediction accuracy, it also indicates that the model relies solely on its memory, akin to guarding an old library, struggling to keep pace with real-world changes; real-time acquisition of the latest knowledge is crucial. Furthermore, hallucinations and factual instability pose core obstacles, akin to hidden time bombs. Additionally, constraints related to cost, data, and evaluation systems make large-scale applications challenging. Training high-precision models requires vast computational resources, resulting in high development costs; in reality, time series data is fragmented and lacks uniform labeling, raising the question of how to produce quality outputs from inferior raw materials. Existing evaluation systems often emphasize numerical errors while neglecting factual stability, leading many models to appear excellent yet struggle to implement.

Looking ahead, the development direction of large model predictive technology is clear and focused, centering on “from powerful to trustworthy,” aiming to create a mature technical system that can stably serve real-world decision-making. Firstly, general large models will evolve into specialized foundational models for predictions, demonstrating stronger competitiveness in high-precision demand scenarios such as meteorology and finance. Secondly, tool enhancement will become an important direction, enabling models to autonomously invoke external tools like search and simulation, akin to equipping intelligent agents with treasure chests to better tackle complex scenarios. Thirdly, trustworthiness, controllability, and explainability will become research priorities; future prediction systems must not only achieve numerical precision but also quantify risks and trace judgment bases, which is key for high-risk scenario implementations. Fourthly, low-cost deployment and industrialization will accelerate; as inference costs decrease and open-source ecosystems improve, technology will transition from being exclusive assets of a few institutions to common tools across various industries. Lastly, domestic research will focus on localized adaptations, creating specialized models that align with the Chinese context and local data, ensuring that large models are more accurate, stable, and trustworthy in domestic financial risk control and government early warning scenarios.

Kimi's Valuation Surpasses $20 Billion in Just Three Years

Thu, 07 May 2026 00:00:00 +0000

Kimi’s Valuation Surpasses $20 Billion

On May 7, 2026, the AI industry was shaken by a significant announcement: Kimi, founded just three years ago, is set to complete a new funding round of $2 billion, pushing its post-money valuation over $20 billion.

This figure is staggering, equivalent to more than half the market value of Bilibili, and it places Kimi at the top of the domestic AI startup funding leaderboard, with its valuation more than quadrupling in less than six months.

What is the secret behind Kimi’s ability to attract capital even in a downturn? Why are top institutions eager to invest in Kimi? Is this financing round a sign that the domestic AI landscape is about to change?

01. Thriving Amidst Adversity

In a time when fundraising in the primary market is cooling and valuations in the AI sector are generally declining, Kimi’s latest funding round is an extraordinary example of thriving against the odds.

With this round, Kimi has completed four funding rounds in less than six months, raising a total of over $3.9 billion, equivalent to more than 37.6 billion RMB, firmly securing its position as the leading AI startup in China.

To put this in perspective, Kimi’s valuation was only $4.3 billion last November. In just six months, its valuation has nearly quintupled, a growth rate that is extremely rare in the history of Chinese internet companies.

Interestingly, while giants like ByteDance, Alibaba, and Tencent are spending billions on subsidy wars to capture consumer traffic, Kimi has not engaged in this competition but has instead carved out a completely different path.

This unique approach is what captivates the capital market: Kimi has not followed the industry’s rules but has established its own set of game rules.

02. Capital Frenzy: More Than Just Long Text Processing

Many believe Kimi’s rise is solely due to its long text processing capabilities. However, top institutions that value Kimi at $20 billion see much more than just a single product highlight.

Kimi’s first core asset is its underlying technological barrier that can define industry standards.

Just like in the automotive industry, where competitors focus on aesthetics and acceleration, Kimi has developed a more efficient engine that has become the industry standard. Its self-developed Muon optimizer has replaced the decade-old Adam optimizer and is being adopted by peers; its attention residual technology paper has even been personally shared by Elon Musk and is regarded as a hallmark of the deep learning 2.0 era.

Kimi’s second core strength lies in its proven, explosive growth in commercialization.

For investors, potential revenue is one thing, but actual revenue is what matters. The data speaks for itself: Kimi’s annual recurring revenue (ARR) just surpassed $100 million in early March and doubled to over $200 million in April. In less than four months into 2026, Kimi has already earned more than its total revenue for all of 2025.

Importantly, this growth is not driven by subsidies creating a false sense of prosperity. The paid subscription rate for its multi-tiered C-end membership system continues to soar, and its B-end API services cover over 200 countries globally, with overseas revenue now officially surpassing domestic earnings, completely escaping the domestic market’s competitive mire.

Kimi’s third key advantage is its practical product logic that avoids following trends.

While the entire industry is fixated on model parameters and competition rankings, Kimi focuses on addressing real user pain points. Whether it’s breaking down lengthy financial reports or handling complex tasks with hundreds of intelligent agents simultaneously, all of Kimi’s functionalities target the genuine needs of professionals, developers, and enterprises.

Users don’t care how many parameters your model has; they care about whether it can solve their problems. Kimi has thoroughly understood this principle.

03. After Kimi’s $20 Billion Valuation: A Major Shake-Up in the AI Sector

This $2 billion funding round is not just a victory for Kimi; it is a seismic shift that will impact the entire domestic AI sector.

First, the Matthew effect in the industry will be amplified.

Leading players now have ample resources for technology development, talent acquisition, and global market expansion, providing them with greater operational flexibility.

In contrast, smaller players lacking core technology and revenue-generating capabilities will find their survival space rapidly shrinking.

The competition in the AI sector has shifted from a previous focus on “technological positioning” to a comprehensive battle involving “technology + commercialization + globalization.” Players without real capabilities will soon be eliminated.

Second, the competitive logic in the industry has fundamentally changed.

Kimi’s success offers a vivid lesson to the entire industry: parameter competition has no future, and subsidies for traffic acquisition lead nowhere. Only technologies that can be implemented, generate revenue, and create real value will establish a genuine competitive moat.

In the future, AI competition will not be about who spends more money but about who can better integrate technology with user needs and industry scenarios, and who can capture a larger share of the global market.

Ultimately, Kimi’s rapid ascent is not a bubble of capital frenzy but a market endorsement of “pragmatism” with real investment.

In recent years, we have witnessed many tech sectors riding the wave of trends, numerous inflated promises for funding, and many products with high parameter counts that are far removed from user needs.

However, the essence of business has never changed. Products that solve user pain points are good products; companies that create real value deserve market favor; and enterprises that are rooted in technology and have a global vision will ultimately prevail in this global AI competition.

This is not just a survival rule for the AI sector but a fundamental logic for all business stories.

Kimi’s recent funding is merely the beginning. The real battle for China’s AI on the global stage has just begun.

Doubao Starts Charging: A Positive Development for the AI Industry

Wed, 06 May 2026 00:00:00 +0000

Doubao Starts Charging

Doubao has begun charging users, surprising many with its late timing and high prices: the standard version costs 68 yuan per month, the enhanced version 200 yuan, and the professional version 500 yuan per month, with an annual fee exceeding 5,000 yuan. Just last year, domestic competitors were still focused on offering free services and token subsidies to retain users.

However, this article does not aim to simply report on the pricing news. When a widely-used AI application with 227 million monthly active users decides to start charging, and at such high rates, it raises more questions than just a price list.

This situation involves at least three key questions: Why did ByteDance choose this timing? What exactly is the 500 yuan monthly fee for, and who is it aimed at? What does this charging experiment mean for the entire domestic AI industry?

The Cost of Free Services

First, it is essential to note that Doubao is not charging hastily. In terms of user base, it is more than qualified to discuss pricing compared to any domestic competitor. According to QuestMobile, Doubao’s monthly active users surged to 227 million in Q4 2025, maintaining its position as the industry leader and becoming the first AI-native application in China to surpass 100 million daily active users. ByteDance’s three AI applications—Doubao, Jimeng AI, and Doubao Aixue—total around 250 million monthly active users, ranking among the top ten in both monthly active users and downloads.

What does this number signify? There are only a handful of internet products in China that can be considered “national-level.” Doubao achieved this scale with far less expenditure than many might expect.

Data from AppGrowing, a domestic mobile advertising monitoring agency, shows that Doubao’s advertising spending decreased throughout 2025: from 161 million yuan in Q1 to 117 million yuan in Q2, then halving to 65 million yuan in Q3, and slightly rebounding to 92 million yuan in Q4.

Looking at user retention, Doubao’s 30-day retention rate averaged 44% from January to November 2025, significantly outpacing its nearest competitor, Kimi. More granular data even shows that Doubao’s 180-day retention rate exceeds its 90-day retention rate, demonstrating a rare “smile curve” for AI chat products.

This indicates that Doubao is not merely a product built on financial burn; it is genuinely used and retained by users. In an industry where customer acquisition costs are exorbitant and user loyalty is almost non-existent, these figures are remarkable.

However, the larger the scale, the more challenging the financials become. The cost of a free model involves considerations of GPU clusters, the linear expansion of inference costs, and the unpredictable future of computational power investments.

According to the latest data disclosed by Volcano Engine, Doubao’s model token usage skyrocketed from 120 billion in May 2024 to over 30 trillion by September 2025, a growth of more than 200 times. Each additional daily active user, every conversation, and every command like “help me generate a PPT” consumes computational power.

With 100 million daily active users, allowing them to continue using increasingly complex model capabilities for free is akin to turning the treasury into a public square where anyone can help themselves.

Examining ByteDance’s overall computational power strategy reveals the urgency behind this charging experiment. According to a report from Zheshang Securities, ByteDance’s capital expenditure in 2025 is expected to reach 160 billion yuan, with approximately 90 billion yuan allocated for AI computational power procurement and 70 billion yuan for IDC infrastructure and network equipment.

Moreover, this is not the endpoint. Doubao’s daily token consumption is still growing exponentially, and ByteDance anticipates “higher” future token consumption, requiring continued investment in computational power. According to estimates from Dongfang Securities, if Doubao’s daily active users reach 50 million by 2027, with an average daily token usage of 50 trillion, it would require over 565,000 GPUs, indicating a significant shortfall.

On one side, there is a flood of 227 million monthly active users and over 100 million daily active users; on the other, there is a capital expenditure of 160 billion yuan. Every computational expense translates to real money, and someone has to foot the bill sooner or later.

Who is Paying 500 Yuan a Month?

With the pricing table now public, the immediate reaction in the domestic market is undoubtedly “expensive.” The standard version costs 68 yuan, the enhanced version 200 yuan, and the professional version 500 yuan per month, with an annual fee of 5,088 yuan, more than double the annual subscription price of ChatGPT Plus (around $240/year).

Given the long-standing narrative in the domestic internet space of “winner takes all and free for all,” this figure indeed feels strikingly out of place.

However, this is precisely where the most thought-provoking aspect lies. Doubao does not intend for everyone to pay this fee. The official response from Doubao is clear: “Doubao will continue to provide free services while exploring additional value-added services, with related plans currently in the testing phase.”

Sources close to Doubao have further revealed that the paid features will primarily focus on complex tasks and productivity scenarios, such as PPT generation, data analysis, and film production. As the model capabilities continue to upgrade, the product can meet increasingly complex high-value tasks, which require more computational power and inference time, thus necessitating the introduction of paid services. The free version will still cater to users’ everyday needs.

In other words, Doubao is drawing a line: lightweight uses like casual chatting, writing, and information retrieval will remain free; only those who use AI for work and genuinely treat the large model as a productivity tool will need to pay.

This is a pricing logic based on “tiered computational costs”; light users are covered by Doubao’s free service, while medium to heavy users pay for their computational consumption. In fact, this pricing strategy has seen successful examples abroad. Claude Code officially launched its subscription model in 2025, achieving an annualized revenue of $1 billion within six months, and by February this year, it had surged to $2.5 billion, relying on high-performance models and subscription fees.

So why is Doubao’s professional version priced at 500 yuan/month? Frankly, this may not be because Doubao believes its product is worth 500 yuan, but rather because it aims to establish a domestic price anchor for the “AI productivity tool” category.

In the realm of consumer AI subscriptions, there are virtually no benchmarks or industry standards in China. Doubao’s impressive numbers can serve as proof that this is not a cheap tool nor a new gadget that needs to win over users with low prices; it is something that can genuinely replace part of professional productivity.

As for how many people in the market are willing to pay for this, that is another matter. First, it needs to secure its position and anchor the price; subsequently, there will be ample room for promotions, discounts, or splitting packages by functionality.

Broaden the perspective, and Doubao is not an isolated case.

The commercialization of consumer AI assistants in China is entering a phase of intensive exploration from late 2025 to early 2026. Kimi launched its paid membership last year, offering three tiers: a free version with limited access to deep research and OK Computer functions; 49 yuan/month for an equivalent API exchange voucher; and 99 yuan/month for increased quotas and concurrent support.

Internal communications from Kimi revealed that since November 2025, overseas API revenue has quadrupled, with month-on-month growth in paid users exceeding 170%. Kimi’s “OK Computer” incurs a single conversation cost of 4-5 yuan, making a paid model essential for scaling.

Earlier, Baidu’s Wenxin Yiyan launched a professional version in 2023, priced at 59.9 yuan/month, with a continuous monthly discount of 49.9 yuan. This was the first paid large model product targeting consumers among major domestic companies. However, in April 2025, Wenxin Yiyan announced a return to full free access, reflecting both competitive pressure from DeepSeek’s free model and Baidu’s struggles in the consumer AI application space.

As for Tencent Yuanbao, it still adheres to a fully free strategy, keeping all core functions open to users without launching an independent membership subscription system.

Upon reviewing the situation, a clear pattern emerges: the entire industry is transitioning from “free trials” to “tiered payments,” with differences only in their respective paces and approaches.

Doubao’s pricing is the highest and its steps the largest, likely because it has the most stable user base and the greatest cost pressures; Kimi is taking a mid-range approach, differentiating with coding tools and agent models; Baidu attempted to charge first but ultimately returned to full free access; Tencent Yuanbao is still observing, trying to attract users with free services while cautiously exploring future payment possibilities.

The Real Stakes of the Charging Experiment

When we focus solely on a company’s decision to charge, it is easy to get caught up in superficial discussions about whether the price is too high or whether users will leave.

From an industry perspective, the significance of Doubao’s charging lies in the essential question it poses to the domestic AI industry.

This question is almost existential: when the free benefits are exhausted, when the financing window tightens, and when IPO exits seem distant, can domestic AI companies sustain themselves based on their products, rather than relying on customized projects for businesses, selling computational power, or government contracts?

The difficulty of this question is partially reflected in the data. For example, MiniMax reported a total revenue of $79.03 million in 2025, with overseas revenue accounting for 73%. Its overall gross margin was 25.4%. This is currently the best-performing domestic consumer AI product, but its core revenue source is the overseas market, while the domestic market’s willingness and ability to pay for AI products remain the largest uncertainties in the industry.

In contrast, the story is entirely different overseas. OpenAI’s revenue in the first half of 2025 was $4.3 billion, with an estimated annual revenue of $13 billion, of which over 80% comes from paid subscription users of ChatGPT, including Plus personal subscriptions, enterprise versions, and team versions. OpenAI’s paid enterprise users have surpassed 3 million, and ChatGPT’s weekly active users exceed 800 million.

Gartner predicts that global GenAI model end-user spending will reach $14.2 billion in 2025, while its actual ARR has already surpassed $20 billion, showing a rapid growth trend in revenue generated by the commercialization of AI applications. In overseas markets, personal subscriptions can support the revenue expectations of a company valued in the hundreds of billions.

However, in China, the situation is much more complex. A core difference lies in the payment culture: overseas, from GitHub Copilot to JetBrains, it is common for programmers to spend hundreds of dollars annually on IDEs, as the culture of paying for tools has been cultivated over decades.

In contrast, the habit of paying for software tools among consumer users in China has yet to be firmly established. More troubling is the severe homogeneity of domestic large model products; competitors can almost instantly replicate any features you offer, making true differentiation extremely scarce. Once charging begins, the cost of users switching to another free app is nearly zero.

In this environment, the first company to charge risks being undermined by competitors employing free strategies.

This is why Doubao chose to start charging only after surpassing 100 million daily active users and 227 million monthly active users; its scale provides enough room for user selection and trial-and-error. Even with a mere 1%-2% conversion rate to paid users, the corresponding revenue scale is still significant. In contrast, Kimi’s paid volume and user base are much smaller. Tencent Yuanbao has yet to charge, likely waiting for a more opportune moment.

From a longer-term perspective, Doubao’s charging action brings to light a logic that everyone in the industry has been contemplating but few have articulated: large models are neither public services nor charitable endeavors; they are businesses that need to recoup costs.

When the largest consumer AI application in China begins to ask users for money, the entire industry can no longer evade the commercialization question by claiming it is still in the “technological dividend period” or “still in the land grab phase.”

For open-source free models like DeepSeek, the pressure will be transmitted rapidly. DeepSeek has always promoted a free route, but in early 2026, it faced the departure of key team members from its core technology lines, including foundational models, inference, OCR, and multimodal capabilities, as competitors offered annual salaries in the millions.

The logic of “Huanfang earns enough to burn” has become increasingly inadequate in the face of the intensity of talent competition. Doubao firing the first shot in charging signifies, to some extent, that the industry is competing for self-sustainability; only by earning money through products can it retain top talent, thereby creating better products and forming a positive cycle.

Reflecting on the past year of the AI entrepreneurship wave, an uncomfortable truth remains: while domestic large model companies have rapidly caught up in technical capabilities, achieving even some indicators of surpassing, they have collectively failed to establish a commercial closure on this most fundamental issue.

The temporary premium brought about by technological leadership is quickly neutralized by followers, and price wars and free strategies have plunged the entire industry into a prisoner’s dilemma. Doubao’s recent actions, regardless of their ultimate success, at least provide a reference point for the industry—what is the value of large model services, what should it be worth, and what capabilities are users willing to pay for in AI?

We believe that given Doubao’s current scale and ByteDance’s strategic patience, this charging experiment is likely to continue for quite some time. It will experience user losses, public controversy, product iterations, and price adjustments, ultimately finding a dynamic equilibrium through ongoing negotiations.

For the industry, if Doubao can maintain its user base and market position while charging, it will demonstrate that even in China, AI can survive on user payments rather than relying on blood transfusions. Conversely, if charging leads to significant user losses and competitors seize the opportunity to overtake, the industry may struggle in the free quagmire for a long time.

“What is free often comes at the highest cost.” This statement is particularly apt in the AI industry; it suggests not only that users must pay some price for free services but also that companies must pay for “free” as well.

Hebei's 14th Five-Year Plan Focuses on Digital Transformation and AI Integration

Tue, 05 May 2026 00:00:00 +0000

Drone footage of the Langfang City Big Data Innovation Application Center in the Beijing-Tianjin-Hebei region.

Recently, the “14th Five-Year Plan for National Economic and Social Development of Hebei Province” was released, which emphasizes the need to advance the construction of a digital Hebei and enhance the level of intelligent development. The plan aims to leverage Hebei’s computing power and scene advantages, activate the potential of data elements, accelerate innovation in artificial intelligence and other intelligent technologies, and deepen the expansion of “AI +” to empower economic and social development as well as governance capabilities, ultimately building a data-driven and intelligent digital Hebei.

This reflects Hebei’s strategic determination to actively embrace the trend of intelligent transformation and accelerate the cultivation of new productive forces. According to Chen Gang, an assistant researcher at Tsinghua University, the digital economy is no longer an optional question but a mandatory one. Advancing the construction of a digital Hebei will bring comprehensive opportunities for industrial upgrading, innovation breakthroughs, governance efficiency, and improvements in people’s livelihoods, becoming a strong intelligent engine for high-quality development.

Promoting Efficient Supply of Computing Power, Algorithms, and Data

Computing power, algorithms, and data are the three essential elements for the development of artificial intelligence. The plan outlines a clear path for the digital transformation during the 14th Five-Year period by promoting the efficient supply of computing power, algorithms, and data.

During the 13th Five-Year period, Hebei focused on building a nationally leading computing power industry ecosystem, with the province’s comprehensive computing power index ranking first in the country for two consecutive years. Langfang and Zhangjiakou have consistently ranked among the top two cities in computing power index.

The plan emphasizes strengthening the construction of computing power infrastructure. During the 14th Five-Year period, Hebei will accelerate the establishment of a national integrated computing power network hub in the Beijing-Tianjin-Hebei region, promoting the creation of an integrated computing power network and establishing a shared computing power resource pool. The construction of the Zhangjiakou-Langfang intelligent computing power supply corridor will create an intelligent computing power aggregation area around Beijing.

According to Xia Luohui, director of the Hebei Institute of Technology Innovation at the China Academy of Information and Communications Technology, this integrated layout aims to address the previous issues of dispersed computing power platforms and inconsistent standards. Hebei will shift from “single-point breakthroughs” to a “one network for Beijing-Tianjin-Hebei” approach, transforming green electricity advantages into cost advantages through interconnected computing power resource pools and collaborative mechanisms.

To promote the iterative innovation of algorithm models, Hebei will implement a major model research and development initiative, aiming to cultivate a batch of high-level vertical models in various industries and create replicable and promotable industry-specific model demonstrations. Collaborative models such as “Jing Model, Ji Training” and “Ji Training, Jing Use” will be explored to iteratively upgrade vertical models in steel, automotive parts, cultural tourism, and government affairs, supporting the application of large models in industries like chemicals, construction materials, education, and healthcare.

Focusing on deepening the development and utilization of data resources, Hebei will expand the supply of public data resources during the 14th Five-Year period, encouraging enterprises to circulate data through various means such as sharing, trading, and resource replacement, and to build high-quality data sets. The province will carry out the “Data Element*” initiative and steadily advance the construction of data infrastructure. Additionally, it will promote the development of data industry clusters in cities like Shijiazhuang, Zhangjiakou, Langfang, and Baoding, and support the establishment of a national data labeling base in Baoding, creating a data labeling industrial belt around Beijing.

With a strong computing power “engine,” efficient algorithm “brain,” and smooth-flowing data “blood,” Hebei is advancing the integrated development of computing power, algorithms, and data, constructing a full chain of “computing power supply - algorithm innovation - data empowerment,” forming an important support for the development of the digital economy.

Empowering Industries with Intelligent Technology

Artificial intelligence is becoming a powerful driver for high-quality economic development. The plan focuses on empowering industrial development through intelligent technology, proposing the implementation of the “AI +” initiative to enhance the deep integration of the real economy and the digital economy, and to strengthen and expand the digital economy.

“AI +” is not just a simple addition but a deep integration. According to Zhuang Zhiwei, director of the Development and Planning Department of the Provincial Data and Government Service Bureau, Hebei possesses a complete industrial system and a vast array of application scenarios, which provides confidence for intelligent transformation. In the next five years, Hebei will accelerate the digital transformation of the economy driven by the “AI +” initiative, seizing the high ground of AI industrial applications and empowering various industries comprehensively.

According to the plan, Hebei will expand the space for economic digital transformation, strengthen core industries of the digital economy, promote the intelligent transformation and digital networking of manufacturing, advance the digitalization of the service industry, and develop smart agriculture. Each measure is precisely targeted to comprehensively empower industrial upgrades.

“AI +” not only aids industries in advancing but also opens up new ways of living.

The plan states that Hebei will fully leverage intelligent technology and data elements to enrich people’s lives and improve their well-being, promoting the integrated application of AI in education, healthcare, elderly care, employment, culture, and consumption.

The implementation of smart education demonstration projects, digital healthcare demonstration projects, and the vigorous development of digital cultural and creative services will enrich scenarios for smart homes, smart travel, and smart communities. In the next five years, AI will integrate into daily life with unprecedented depth and warmth, becoming a force that enhances quality of life and serves the public.

The digital transformation of government services is also accelerating. According to the plan, Hebei will deepen the full-process application of intelligent technology, developing diversified government services that are accessible, smart, convenient, and equitable. This includes enhancing the digital and intelligent governance and service levels of government, exploring the construction of a service model that accurately identifies needs, proactively plans services, and processes them intelligently throughout, optimizing platforms like “Ji Time Service”.

Zhuang Zhiwei noted that this signifies a profound shift in government services from “people seeking services” to “services seeking people.” In the future, high-frequency matters such as business registration and project approval will be handled through large models for intelligent form filling, automatic verification, and instant processing.

Balancing Development and Governance for a Healthy Digital Economy

The healthy development of the digital economy relies on scientific and effective governance. The plan emphasizes the importance of balancing development and governance, advancing the foundational system for data elements, and ensuring a beneficial, safe, and fair development environment.

“Data only has value when it circulates. The deployment in the plan regarding the implementation of foundational systems for data elements is key to transforming data resources into assets,” said Chen Gang.

He believes that as a new type of production factor, the core breakthrough for data elements lies in resolving challenges related to rights confirmation, pricing, and circulation. This year, Hebei achieved two major breakthroughs: the first batch of data property registrations in the country and the first case of data asset pledge financing, marking a milestone in the market-oriented reform of data elements in Hebei, clearing obstacles for the conversion and circulation of enterprise data.

During the 14th Five-Year period, Hebei will implement systems for data property rights, circulation and trading, revenue distribution, and safety governance, effectively utilizing data pricing mechanisms, cultivating a service ecosystem for data element circulation and trading, and promoting the inclusion and valuation of data assets, forming a comprehensive service system of “data rights confirmation - valuation - inclusion - pledge - securitization.”

As technology advances, regulatory mechanisms are increasingly needed for protection.

During the 14th Five-Year period, Hebei will strengthen safety regulation for new technologies and new business models, improve AI governance, promote innovation and healthy development of the platform economy, and legally combat data abuse, forgery, and privacy breaches. It will explore cross-business and cross-departmental joint regulation in various scenarios, constructing a multi-faceted collaborative governance ecosystem.

“Governance is not a decelerator but a stabilizer and safety net,” said Shao Yunxia, a researcher at the Hebei Academy of Sciences. As technology continues to innovate and break through, relevant parties are accelerating the formulation and improvement of laws, regulations, policies, application norms, and ethical guidelines, which will clarify the direction, leave room for growth, and lay a solid foundation for the healthy development of artificial intelligence.

The Surprising Value of Guangxu Yuanbao: A Wealth Opportunity

Sun, 03 May 2026 00:00:00 +0000

Introduction

Forget about the 1980 edition 50 yuan notes! The collection world has seen a major shift, with a new “super dark horse” emerging—Guangxu Yuanbao!

Many people have a few unassuming Guangxu Yuanbao coins lying in drawers, old wooden boxes, or even in clutter on balconies. These coins, in brass and red copper, may be covered in dust or treated as toys by children, or regarded as ordinary old coins by the elderly.

But did you know that these seemingly worthless old items have now become a “money tree” sought after by collectors? The ordinary Guangxu Yuanbao has skyrocketed by 500 times over the decades. Once worthless brass coins can now sell for thousands to tens of thousands; those in good condition and rare versions can fetch hundreds of thousands or even millions. Some individuals have achieved financial freedom thanks to just a few Guangxu Yuanbao coins!

Recently, I’ve been bombarded with questions: “Are my brass Guangxu Yuanbao coins real?” “Are Yuan Datou and Shuangqi coins still valuable?” “Is it true that some people have gotten rich from Guangxu Yuanbao?”

In this article, I will provide a detailed analysis of Guangxu Yuanbao, covering seven popular versions, real collection cases, practical identification techniques, and the latest market trends. You will learn how valuable these coins are, which versions are the “wealthy types,” and how to distinguish genuine coins from fakes. After reading this, you’ll easily recognize the “wealth code” and no longer treat treasures as trash!

Key Points

Guangxu Yuanbao is not a single type; it is divided into two main categories: copper coins and silver coins, with over a hundred versions. The most scarce and valuable versions are those produced in Jiangsu, Zhejiang, Fujian, Guangdong, Hubei, and the Three Eastern Provinces. The ordinary brass Guangxu Yuanbao currently sells for 500-5000 yuan each, a 500-fold increase from its face value. The silver coin versions are even more impressive, with ordinary versions selling for 10,000-50,000 yuan each, while rare versions can go for hundreds of thousands or millions. The “coin king” from the Fengtian province has even fetched tens of millions at auction, making it a true “money printer” in the collection world!

What’s most touching is that many valuable Guangxu Yuanbao coins are often hidden in our homes—in old family albums, ancestral wooden boxes, or old envelopes left by grandparents. If you pay attention, you might uncover your own “unexpected surprise.”

There are no exaggerated stories of treasure hunting or dramatic tales—just the everyday experiences of ordinary families. Some have found old coins left by their grandfathers and learned they could sell them for tens of thousands, while others mistakenly sold their brass Guangxu Yuanbao as scrap, only to regret it later upon learning the truth. There are even those who, on a whim, sent their silver coins for appraisal, only to discover they were valued at tens of thousands, instantly becoming “invisible millionaires.”

The rise of Guangxu Yuanbao as a “wealthy dark horse” is no coincidence. It was minted during the Guangxu era (1875-1908) and served as a major currency during the late Qing Dynasty, witnessing the changes in modern Chinese history. It holds both historical and collectible value. After more than a century, many Guangxu Yuanbao coins have been lost, damaged, or melted down, making the remaining genuine pieces increasingly rare and valuable.

Moreover, the entry barrier for collecting Guangxu Yuanbao is very low compared to the 1980 edition 50 yuan notes, which are hard to find. Many ordinary families can find these coins, but most people are unaware of their value, missing out on opportunities for wealth. Today, I will introduce you to the seven most popular and valuable Guangxu Yuanbao versions, so beginners can easily assess the value of their old coins!

1. Detailed Explanation of 7 Popular Guangxu Yuanbao Versions

Guangxu Yuanbao is divided into copper and silver coins. The copper coins are primarily made of brass and red copper, with face values often being “ten” or “twenty,” making them relatively affordable and suitable for beginners. The silver coins are primarily made of silver, with face values like “Kupi Qiqian Erfen” and “Kupi Yiliang,” which are more expensive and highly sought after by collectors. Below are seven versions currently popular in the collection market, each with detailed explanations that beginners should remember!

(1) Jiangsu Brass Guangxu Yuanbao: Most Common, Ordinary Versions Sell for 500-5000 Yuan

The most common Guangxu Yuanbao is the Jiangsu brass version, especially the “ten” copper coin, which many people born in the 70s, 80s, and 90s have seen at home. Most older generations would keep a few as “treasures” or even use them for good luck during childhood celebrations.

As a child, I had three Jiangsu Guangxu Yuanbao coins with a brass color that gleamed slightly. The coin face featured a circle of characters and the words “Ten Rich Copper Yellow” below. They felt thick and substantial, with finely serrated edges. I often mixed them with ordinary copper coins, playing with them, while my mother would warn me, “Don’t mess with those; they’re old items.”

It wasn’t until recent years, with the rise in collecting interest, that I realized the brass Guangxu Yuanbao I once treated as toys has now appreciated 500 times! The market price for an ordinary Jiangsu brass Guangxu Yuanbao ranges from 500 to 2000 yuan each; those in good condition can sell for 3000 to 5000 yuan each, a 500-fold increase from its face value!

Many people ask, “Are my brass Guangxu Yuanbao coins with green spots fake?” Actually, these green spots are natural oxidation, proving they are genuine. Here are three key features to identify Jiangsu brass Guangxu Yuanbao:

Material and Color: Made of pure brass, bright brass color, with natural patina. Some may have faint green spots, feeling smooth and pleasant to the touch; fakes are usually made of imitation brass, dull in color, and rough to the touch.
Shape and Design: Round and thick, with finely serrated edges, the coin face has the words “Guangxu Yuanbao” clearly inscribed, surrounded by intricate patterns, with no blurriness or defects.
Sound Identification: When lightly tapped, genuine coins produce a clear sound with a slight resonance; fakes sound dull and harsh, making them easy to distinguish.

Note: Although Jiangsu brass Guangxu Yuanbao is common, there are fakes, especially those that appear “brand new” online, which are often artificially refurbished and worthless. Beginners should avoid wasting money.

2026 Latest Market Trends

Jiangsu Guangxu Yuanbao (brass): Ordinary condition 500-2000 yuan each; good circulation 2000-3000 yuan each; perfect condition 3000-5000 yuan each.
Zhejiang brass Guangxu Yuanbao prices are slightly higher: ordinary condition 800-2500 yuan each; perfect condition 4000-6000 yuan each.

(2) Fujian Official Red Copper Guangxu Yuanbao: Rare Production, Good Circulation Pieces Sell for 10,000-30,000 Yuan

If Jiangsu brass Guangxu Yuanbao is the “entry-level version,” then Fujian official red copper Guangxu Yuanbao is the “advanced version,” with much lower production and circulation, making it highly sought after in the collection market.

My grandfather’s old comrade had a set of Fujian official red copper Guangxu Yuanbao kept in a worn wooden box for decades. My grandfather said that the production of Fujian official coins was limited, mostly circulating locally, and few survived due to loss or damage over the years.

As kids, we would fight over these red copper coins, which had a deeper color than the brass version and a slight metallic taste. They felt heavier and cooler to the touch, with finer edges and more intricate designs. My grandfather often remarked, “Don’t underestimate these coins; back in the day, one could trade for a small bag of oil or salt.”

It wasn’t until last year that my grandfather’s comrade’s son appraised these red copper Guangxu Yuanbao and discovered their value—ordinary pieces are worth 5000-10,000 yuan each; well-circulated ones can sell for 10,000-30,000 yuan each; perfect condition coins can reach 30,000-50,000 yuan each, far exceeding the Jiangsu brass version!

To identify Fujian official red copper Guangxu Yuanbao, beginners should remember these four points:

Material and Color: Made of pure red copper, dark red color with natural patina, evenly distributed; fakes are often brass plated with red copper, bright in color, and lose paint easily.
Weight and Feel: Red copper is denser than brass, making it feel heavier and more substantial; fakes tend to be lighter and rough.
Design Details: The coin face has clear and precise characters, with intricate patterns and no blurriness; fakes often have rough designs and errors.
Patina Identification: Genuine coins have a smooth patina, while fakes have rough, artificial coatings that can flake off.

2026 Latest Market Trends

Fujian official Guangxu Yuanbao (red copper): Ordinary condition 5000-10,000 yuan each; good circulation 10,000-30,000 yuan each; perfect condition 30,000-50,000 yuan each.

(3) Guangdong Province Red Guangxu Yuanbao: High Circulation but Detail-Oriented, Ordinary Versions Sell for 800-3000 Yuan

Guangdong Province Guangxu Yuanbao is one of the most circulated types, especially the red version, which was widely used in Guangdong. Despite its high circulation, genuine pieces remain scarce due to the emphasis on detail.

My grandfather saved a few Guangdong Province red Guangxu Yuanbao coins from his work in Guangdong, where trade was flourishing, and these coins were the primary currency for transactions.

This red Guangxu Yuanbao is darker than the Fujian version, with a faint waxy smell and visible wear from circulation, which is a hallmark of authenticity. My grandfather saved these coins to fund his marriage, not realizing their future value.

Currently, the prices for Guangdong Province red Guangxu Yuanbao, while lower than the Fujian version, are still significantly higher than the ordinary brass version: ordinary condition 800-2000 yuan each; good circulation 2000-3000 yuan each; perfect condition 3000-6000 yuan each, a 400-fold increase from face value.

To distinguish Guangdong Province red Guangxu Yuanbao from fakes, remember these three details:

Color and Patina: Dominantly red, with a smooth patina that feels warm and substantial; fakes are overly bright and lack the waxy feel.
Edge Teeth and Design: Smooth edges with clear characters and intricate patterns; fakes have rough edges and unclear designs.
Weight and Sound: Genuine coins feel substantial, producing a clear sound with resonance when tapped; fakes are either too light or too heavy, producing dull sounds.

2026 Latest Market Trends

Guangdong Province Guangxu Yuanbao (red): Ordinary condition 800-2000 yuan each; good circulation 2000-3000 yuan each; perfect condition 3000-6000 yuan each.

(4) Hubei Province Dragon Face Guangxu Yuanbao Silver Coin: “Silver Nobility,” Ordinary Versions Sell for 1-5 Million Yuan

If copper coins are for beginners, silver coins are the “nobility” of Guangxu Yuanbao, with Hubei Province Dragon Face Guangxu Yuanbao being a standout. The ordinary version sells for 1-5 million yuan each, making it a top choice for many collectors.

My grandfather had a Hubei Province Dragon Face Guangxu Yuanbao silver coin, carefully preserved in a wooden box. He would occasionally show it to me, emphasizing its value and importance in family matters. This coin is heavy and has a faint silver sheen, with intricate dragon designs.

Genuine silver coins produce a clear sound when tapped together, while fakes sound dull and heavy. My grandfather once identified fakes among a relative’s collection simply by sound.

Currently, Hubei Province Dragon Face Guangxu Yuanbao silver coins have reached new heights: ordinary condition 1-5 million yuan each; good circulation 5-10 million yuan each; perfect condition 10-20 million yuan each, a several-hundred-fold increase from face value.

2026 Latest Market Trends

Hubei Province Dragon Face Guangxu Yuanbao silver coin: Ordinary condition 1-5 million yuan each; good circulation 5-10 million yuan each; perfect condition 10-20 million yuan each.

(5) Hubei “Kupi Qiqian Erfen” Guangxu Yuanbao Silver Coin: “Collector’s Favorite,” Perfect Condition Can Sell for 20-30 Million Yuan

Among Hubei Province Guangxu Yuanbao silver coins, the “Kupi Qiqian Erfen” version is the most valuable and sought after, with perfect condition coins selling for 20-30 million yuan.

My neighbor, Cai Ma, has a “Kupi Qiqian Erfen” Guangxu Yuanbao silver coin passed down from her mother-in-law, kept in a drawer for decades. She only realized its value after the recent surge in collecting interest.

This coin features the words “Guangxu Yuanbao” in bold characters, surrounded by intricate designs, with a natural patina that signifies authenticity.

Currently, the prices for Hubei “Kupi Qiqian Erfen” Guangxu Yuanbao silver coins have skyrocketed: ordinary condition 5-10 million yuan each; good circulation 10-20 million yuan each; perfect condition 20-30 million yuan each.

2026 Latest Market Trends

Hubei “Kupi Qiqian Erfen” Guangxu Yuanbao silver coin: Ordinary condition 5-10 million yuan each; good circulation 10-20 million yuan each; perfect condition 20-30 million yuan each.

(6) Three Eastern Provinces Dragon Silver: Detail-Oriented, Fine Versions Sell for 10-50 Million Yuan

The Three Eastern Provinces Guangxu Yuanbao dragon silver coins are among the most detail-oriented types, with fine versions selling for 10-50 million yuan each, making them highly sought after.

My downstairs neighbor, Mr. Zhang, is a seasoned collector with a set of Three Eastern Provinces dragon silver coins, each with unique designs. He emphasizes the importance of detail in identifying genuine coins.

Currently, the prices for Three Eastern Provinces dragon silver coins vary greatly based on design and condition: ordinary versions 5-10 million yuan each; fine versions 10-30 million yuan each; perfect condition 30-50 million yuan each.

2026 Latest Market Trends

Three Eastern Provinces dragon silver (Kupi Qiqian Erfen): Ordinary version 5-10 million yuan each; fine version 10-30 million yuan each; perfect version 30-50 million yuan each.

(7) Fengtian Province Gui Mao One Liang Silver Coin: “Coin King” Level, Auction Prices in the Millions

Finally, the “ceiling” of Guangxu Yuanbao is the Fengtian Province Gui Mao one liang silver coin, known as the “coin king,” with auction prices reaching tens of millions.

Minted in the 29th year of Guangxu (1903), its limited production means it rarely circulated, making it a rare treasure.

The Fengtian Province Gui Mao one liang silver coin is larger than ordinary Guangxu Yuanbao silver coins, with intricate designs and a golden patina, making it a coveted piece in collections.

Recent auction prices for this coin have soared, with one selling for 32 million yuan in 2023, and others reaching 45 million yuan.

2026 Latest Market Trends

Fengtian Province Gui Mao one liang silver coin: Ordinary condition 10-20 million yuan each; good circulation 20-30 million yuan each; perfect condition 30-50 million yuan each.

Core Reasons for the 500-Fold Increase in Guangxu Yuanbao Value

After reviewing the seven Guangxu Yuanbao versions, many may wonder why their value has skyrocketed. Here are four core reasons:

(1) Extremely Scarce Supply

Guangxu Yuanbao has a history of over a hundred years. Although initially produced in large quantities, many have been lost, damaged, or melted down. The remaining genuine pieces, especially rare versions, are extremely scarce, leading to high demand and prices.

(2) Deep Historical Value

Guangxu Yuanbao is not just currency; it represents a significant period in Chinese history, witnessing the transition from feudalism to modernity. Collectors value it for its historical significance, further driving demand and prices.

(3) Precious Materials and Craftsmanship

Guangxu Yuanbao is made of precious materials, with copper coins crafted from pure brass and silver coins from pure silver. The intricate designs and craftsmanship elevate its collectible value.

(4) Shifting Collecting Trends

As collectors shift their focus from popular items to Guangxu Yuanbao, especially due to its low entry barrier and potential for appreciation, demand continues to rise.

New Collector Pitfalls: 4 Misconceptions to Avoid

As the value of Guangxu Yuanbao skyrockets, many newcomers may fall victim to scams. Here are four misconceptions to avoid:

Misconception 1: All Guangxu Yuanbao are Valuable

Many believe that all Guangxu Yuanbao are valuable, but this is not true. There are hundreds of versions, and prices vary greatly based on rarity and condition.

Misconception 2: Newer Coins are More Valuable

Some think that newer coins are worth more, but genuine Guangxu Yuanbao typically have natural patina, while artificially refurbished coins are often worthless.

Misconception 3: Blindly Pursuing the “Coin King”

New collectors often chase rare coins like the Fengtian Province Gui Mao, but these are extremely rare and expensive. Beginners should start with more common versions.

Misconception 4: Buying Without Authentication

Many newcomers buy Guangxu Yuanbao without authentication, risking purchasing fakes. Always have coins appraised before buying.

Conclusion

Don’t overlook the “old copper coins” in your home; they could be your “wealth code.” Guangxu Yuanbao, as a representative currency of the late Qing Dynasty, has seen its value soar, with ordinary versions increasing by 500 times and rare versions fetching millions. Those coins in your drawers may hold significant value, representing not just monetary worth but also a connection to family history and culture.

Have you ever seen Guangxu Yuanbao in your home? What version do you remember? Share your thoughts in the comments, as your coin might just be your ticket to wealth!

The Mutual Empowerment of AI and Humanities

Wed, 29 Apr 2026 00:00:00 +0000

The Mutual Empowerment of AI and Humanities

Generative artificial intelligence is profoundly changing various fields such as education, employment, entertainment, healthcare, transportation, and elder care, becoming a hot topic of discussion. The relationship between the humanities and generative AI is complex and symbiotic. AI is reshaping the forms and future development paths of the humanities, while the demands of AI development highlight the value and functionality of the humanities. In this sense, the development of the humanities will fundamentally influence the cognitive heights and social acceptance that AI can achieve.

Bridging Humanities Scholars to Multidisciplinary Approaches

As modern disciplines become increasingly specialized, the barriers between the humanities and natural sciences, as well as between the humanities and social sciences, are widening, potentially leading to a “knowledge dilemma.” It is difficult to find scholars within the humanities who can bridge literature, art, philosophy, history, and language, resulting in a limitation of “partial profundity” in contemporary humanities. The emergence of AI can provide new solutions to this issue.

Large language models, constructed through deep learning on massive text datasets, represent a highly condensed form of human written knowledge. They are based on neural network architectures and algorithm-driven probabilistic predictions, achieving context-aware human-like reasoning guided by specific prompts. In this sense, AI can serve as a powerful assistant for humanities scholars, bridging them to multidisciplinary approaches and empowering the production of humanistic knowledge through information search, literature screening, semantic analysis, and cross-domain integration.

Currently influential “distant reading” methods utilize AI models to establish interdisciplinary literary criticism and research modes based on the overall framework of world literature. Unlike traditional literary studies that advocate close reading of a few classics, distant reading involves data mining and quantitative analysis of large text collections to systematically reveal themes, emotional tendencies, plot structures, and linguistic features, providing a macro description of the overall development of human literature. This effectively addresses the technical challenges of processing massive texts and the cross-cultural, cross-disciplinary knowledge dilemmas that qualitative analyses in traditional literary history and world literature research cannot solve.

Updating Methods and Paradigms in the Humanities

China has a long and rich tradition of humanities scholarship, but the term “humanities” emerged in the twentieth century. During the Enlightenment in the West, humanities scholars sought to find their unique nature and methods outside of natural sciences. They viewed the humanities as a “new science” concerning human thoughts and behaviors, distinct from natural sciences, emphasizing the use of “individualized methods” linked to values, and attempting to construct epistemology and methodology for the humanities.

Overall, in this logic criticized by later generations as the “spirit-nature dichotomy,” the humanities emphasize “existential thought,” with research objects existing in symbolic forms such as language, text, images, and rituals, involving faith, conscience, emotion, aesthetics, values, and ideals that are difficult to quantify. This encompasses deep individual psychology and instincts, consciousness and unconsciousness, and carries historical cultural memory and collective unconsciousness, possessing intrinsic characteristics of value, culture, individuality, spirituality, emotionality, thought, and symbolism inseparable from humanity. Methodologically, the humanities focus on empathetic understanding, contemplative experience, and intuitive insight to reveal unique individual experiences, complex mental worlds, and deep cultural meanings that cannot be captured by replicable, quantifiable, and verifiable technical means of natural sciences.

As disciplines develop, this binary oppositional thinking model is continually being reflected upon. Marx once stated, “Natural sciences will eventually include the science of man, just as the science of man includes natural sciences: this will be one science.” Emerging digital humanities research not only deeply examines the humanistic concerns and governance challenges brought by digital technology but also actively explores new research methods and paradigms from digital technology, reshaping the landscape of humanistic research. Various literary laboratories and beneficial attempts at quantitative humanities research are continually emerging. AI has evolved from an auxiliary tool to a key force driving paradigm innovation, providing humanities scholars with new interdisciplinary research perspectives and theoretical innovation support, significantly expanding the breadth and depth of humanistic research experiences.

Enhancing Critical Thinking and Writing Skills through Human-AI Collaboration

A unique aspect of the humanities is that its knowledge forms often manifest as narrative or speculative texts, expressing researchers’ unique insights and profound reflections on human existence, values, and meanings through written language. This differs from natural sciences, which rely on formulaic deductions, data charts, and repeatable experimental validations, and from social sciences, which heavily utilize surveys and statistical models for empirical paths. Humanistic writing is not only an expression of thoughts and emotions but also a comprehensive cognitive movement that integrates creativity, criticality, and reflection—“writing is thinking,” a process of generating and deepening thoughts and feelings. Writing can stimulate creative vitality, enhance self-reflection, and expand expressive boundaries, where linguistic sensitivity, intellectual penetration, and cultural insight merge. Scholars have pointed out that writing style itself carries the unique emotional tones, academic judgments, and value positions of the researcher to some extent. In this sense, humanistic writing is a core aspect of academic research; it is not only a mode of knowledge production in the humanities but also reflects its ways of thinking and disciplinary characteristics, serving as a fundamental vehicle for maintaining disciplinary existence and promoting academic exchange, as well as a vital source of disciplinary vitality. Whether expressing philosophical thoughts and ultimate meanings, describing historical contexts and narrative events, or constructing values and poetic insights in literary criticism and research, the organization and structural integration of materials, logical reasoning and argumentation, and the deepening of thoughts and condensation of spiritual experiences all occur within the creative writing process.

Currently, AI models can transfer the language structures, argumentative patterns, and disciplinary terminology learned from vast corpora into specific fields of knowledge production in the humanities, promoting human-AI collaboration and achieving a holistic leap in humanistic writing. On one hand, in humanistic academic writing, researchers can fully leverage AI’s powerful data processing capabilities to efficiently gather, systematically organize, and deeply analyze literature before writing. On the other hand, during the writing process, through human-AI collaboration and dialogue, they can organically integrate dispersed knowledge, building new knowledge graphs and cognitive frameworks, helping researchers break through existing theoretical and cognitive limitations, uncovering deep thoughts and internal logical structures from complex texts, revealing developmental laws, refining core concepts, and ultimately nurturing new knowledge outcomes. This process is not merely a simple accumulation of knowledge but an innovative mechanism capable of generating specific theoretical outcomes, opening new paths for academic research and knowledge innovation. Furthermore, AI can partially polish and optimize professional academic expressions, correcting and enhancing the knowledge, normative, logical, and systematic aspects of humanistic academic expressions, even forcing low-quality academic research out of relevant fields. Sometimes, certain academic disputes in the humanities significantly suffer from insufficient materials, unclear concepts, and weak logic; AI assistance can greatly improve the quality of academic debates and enhance their value.

The involvement of AI is not a simple process of machine-assisted writing but a continual deepening of thought and inspiration through human-AI interaction and back-and-forth dialogue. This process places high demands on researchers’ collaborative abilities with AI, especially in correctly inputting instructions, providing high-level prompts, and deeply interpreting output results. These abilities determine the effectiveness of using AI tools. Here, the ability to pose genuine, good, and new questions becomes extremely important, returning to the essence of academic research. At the same time, as some studies have pointed out, AI excels at knowledge inheritance but falls short in creative thinking, making it difficult to replace human depth in theoretical construction, critical reflection, value selection, and aesthetic judgment. Human intuition-based judgments uncover subtle connections among vast information, strategic choices based on value positions, and unique expressions arising from aesthetic tastes, all hold significant importance. If not verified, modified, and deepened by humans, the content generated by AI will carry a strong “machine flavor,” presenting as bland and homogenized expressions.

To ensure the academic independence of thought, unique insights, and distinct academic styles, the personal characteristics of humanities researchers—“talent, courage, insight, and ability”—should not be diminished by machine assistance, preventing dependency thinking and intellectual inertia; otherwise, their research outcomes will lose the dynamism inherent in humanistic research. Humanities research must always reflect “the human” and integrate personal life experiences into academic exploration, responding to contemporary issues with keen perception, unique creativity, and a critical spirit in pursuit of truth. People should feel the emotional investment and value concerns of researchers, possessing both depth of thought and warmth of emotion.

The Development of AI Relies on Humanities Understanding of “Humanity”

As a mirror of human intelligence, AI can help humanity understand the essence of “what it means to be human” more profoundly. Simultaneously, humanity’s understanding of itself becomes the fundamental basis for the future development and governance of AI technology. Marx pointed out, “Conscious life activity distinguishes man directly from animal life activity.” Thus, humanity’s strength lies in its possession of intellect, practical creativity, and continuous learning to acquire knowledge, master skills, and apply them toward achieving goals.

Currently, AI still belongs to the imitation of human intelligence, exhibiting human-like behavior; its development goal should gradually align with the internal mental structures and creative mechanisms of humans, rather than merely replicating external behaviors. The emergence of generative AI is not accidental but a product of human creativity and self-awareness reaching a certain stage. Although currently specialized vertical models have shown superior execution efficiency and precision in specific tasks and fields, they fundamentally remain tools of humanity. Thus far, the “general models” that autonomously adapt to different environments and needs often perform worse than human infants when faced with new situations, counterfactual problems, or common-sense reasoning. Essentially, current AI knows what to do but may not understand the underlying principles and logic; the AI black box has yet to be opened, and it cannot evolve from imitator to understander. In this context, questioning the generative mechanisms and operational modes of human intellect becomes particularly important. Humanity’s reflections on AI also represent a re-evaluation of itself as a complex intelligent entity, further using non-human intelligent agents as mirrors to explore the deep essence of humanity and understand “what it means to be human.”

Both natural sciences and humanities and social sciences are in a cycle of “disenchantment” and “enchantment” regarding humanity, with the core of “enchantment” always being the mystery of humanity itself. Without a profound understanding of one’s own intellect, a “general model” cannot truly emerge. As Marx stated, “Anatomy of man is the key to the anatomy of the ape;” the signs of higher animals displayed in lower animals can only be understood after the higher animals themselves are recognized. Understanding humanity and comprehending what it means to be human is the fundamental nature and basic value goal of the humanities. Today, AI still possesses many “unexplainabilities,” largely due to humanity’s insufficient understanding of its own intellect. Breakthroughs in AI creation, technology governance, and value alignment all require a premise of human understanding of its own essence; the level of development in the humanities determines the future possibilities for the development of “general models.”

From the perspective of the relationship between the humanities and social life, the humanities cannot be replaced by AI, as they possess reflexivity. Every emergence and change of humanistic cognition and understanding intervenes in the development of social life and the construction of public sentiment, embodying the characteristic of “establishing a heart for heaven and earth, and a mission for the people.” In this sense, the development of the humanities is not a linear process of progress; various humanistic thoughts cannot simply be stacked and fused into a single ultimate truth but coexist in a pluralistic manner, collectively shaping the rich spiritual world of society and individuals. It can be said that advancements in humanistic scholarship alter humanity’s understanding of the world, thereby having a significant impact on generative AI. At the same time, the effects of new technologies like AI on society and humanity also become focal points of humanistic scholarship, with related reflections becoming part of the human spiritual world. The humanities and AI are always in a dynamically intertwined state of coexistence and mutual promotion. It is essential to remember that AI is created by humanity, and humanity should possess the ability to truly understand and effectively harness its creations. In this sense, we are fully confident that humanistic thought can illuminate the future path of AI.

Anthropic's Claude Outperforms in Trading Experiment, Revealing AI Disparities

Mon, 27 Apr 2026 00:00:00 +0000

Anthropic’s Claude Outperforms in Trading Experiment

Anthropic’s internal experiment revealed that powerful AI models can earn 70% more in trading than weaker models. Surprisingly, those who lost out were often unaware of their disadvantage and even satisfied with the weaker AI’s performance.

The story begins with a used folding bicycle.

The same bicycle was sold for $65 and $38 in two separate transactions. The seller was the same person, with the only difference being the AI model representing them: Opus 4.5 for the higher sale and Haiku 4.5 for the lower.

This experiment, dubbed “Project Deal,” was recently disclosed by Anthropic.

The findings indicated that strong models can help their users earn more and spend less. This raises a chilling concern about an invisible divide forming in the age of AI agents.

Four Parallel Universes

An AI Negotiation Experiment

The experiment traces back to early 2025 when Anthropic collaborated with Andon Labs on “Project Vend,” where Claude managed an office vending machine.

Claude was misled by journalists into making poor decisions, resulting in over $1,000 in losses. Learning from this, Anthropic decided to have Claude act as an agent instead of a manager.

In December 2025, Anthropic recruited 69 employees, each undergoing a brief interview with Claude to specify their selling and buying preferences. Claude used this information to create a custom system prompt for each employee.

All AIs were placed in a single Slack channel to autonomously post, bid, negotiate, and finalize transactions without human intervention.

The experiment ran four parallel versions:

Run A was public with everyone using Opus 4.5.
Run B was also public but assigned Haiku 4.5 to half the participants.
Runs C and D were private, mixing assignments and using only Opus. Participants only saw A and B, unaware of which model they were using until after the survey.

This design was crucial to ensure unbiased evaluations of AI performance.

Opus Earns More, Spends Less

But Haiku Users Felt Satisfied

The data revealed stark differences. On average, Opus users completed 2.07 more transactions than Haiku users (p=0.001). Opus sellers achieved an average sale price $3.64 higher than Haiku sellers.

Among 161 items sold at least twice, Opus sellers earned an average of $2.68 more, while buyers spent $2.45 less. Given the median item price of $12 and average of $20, this translates to a 15%-20% difference.

In extreme cases, the price disparity was even more pronounced. When Opus sellers interacted with Haiku buyers, the average sale price soared to $24.18, while symmetric transactions between Opus models averaged only $18.63.

This means that the moment a weaker model represents you, you risk being taken advantage of by a stronger model.

The chilling aspect was the subjective fairness ratings. Participants rated Opus transactions at an average of 4.05 and Haiku at 4.06, nearly identical scores.

Among 28 participants who experienced both models, only 17 rated Opus higher, while 11 preferred Haiku. This indicates that those using weaker models were often unaware of their losses and, in some cases, even felt more satisfied with the weaker model’s performance.

Bargaining Prompts

Outmatched by Model Disparity

The experiment also revealed a surprising finding related to prompt engineering. Two types of users participated: one friendly and the other aggressive. The aggressive user expected to save more money, but the data showed no significant impact from aggressive prompts on sale rates.

Anthropic reviewed all participant interactions and found that aggressive instructions did not statistically affect outcomes.

In other words, how you instruct the AI to negotiate had little effect compared to the model’s inherent capabilities.

19 Ping Pong Balls, One Identical Snowboard

These are items Claude negotiated on behalf of users: a blue triceratops, a complete Sherlock Holmes collection, a board game, and more, each representing an AI negotiation.

Some stories were amusing, while others raised concerns. One notable instance involved “Cowboy Claude,” who negotiated in an exaggerated cowboy persona, achieving a sale price of $55, compared to Haiku’s $38.

Another user, Mikaela, instructed Claude to buy a gift for $5, leading to a purchase of 19 ping pong balls. Claude’s justification was both humorous and unsettling, reflecting its ability to mimic human preferences.

In contrast, another employee’s Claude casually mentioned moving into a new home, despite being an AI without such experiences. This highlights the potential risks of AI systems generating false identities and narratives without proper constraints.

The Invisible Divide is Emerging

After the experiment, 46% of participants expressed willingness to pay for AI agent services, indicating a strong market demand. However, Anthropic warns of underlying shadows in this narrative.

The first shadow is inequality. The disparity in AI capabilities could translate into quantifiable economic differences.

The second shadow is trust. AI agents capable of fabricating identities pose risks in real-world transactions, such as rental negotiations or second-hand car deals.

The third shadow is a regulatory vacuum. Currently, no laws clearly define the responsibilities and liabilities of AI agents in transactions.

Anthropic emphasizes the need for society to prepare for these upcoming changes. If the results of this experiment hold true, the next round of competition may depend not on human intelligence but on who employs the smarter AI. Meanwhile, the unaware losers may not even realize they are disadvantaged by a weaker model.

Effective ChatGPT Usage Tips to Boost Productivity

Mon, 20 Apr 2026 00:00:00 +0000

Effective ChatGPT Usage Tips to Boost Productivity

ChatGPT has gained popularity among efficiency enthusiasts, particularly on platforms like KULAAI, as users seek stable methods to utilize it effectively. Many still view ChatGPT as a simple question-answer tool, but those who integrate it into their workflow understand its true potential.

The difference in efficiency isn’t due to the model itself but rather how it’s used. Some users merely chat with it, while others leverage it as a data organizer, writing assistant, proposal reviewer, coding helper, and meeting minutes generator. This is why some claim that ChatGPT can enhance productivity by tenfold—provided it is used correctly.

1. Define the Task Before Asking

Many users jump straight into asking questions, resulting in vague responses and complaints about AI’s capabilities. The first step to efficient usage is setting clear task boundaries. Inform ChatGPT about who you are, what you need, the target audience, the desired output format, word count, and tone. The more specific the information, the closer the result will be to what you need.

For instance, instead of asking, “Help me write a proposal,” specify, “You are a product manager. Based on the following three requirements, produce a proposal suitable for internal reporting, divided into background, issues, and suggestions, using concise language.” Such instructions can significantly increase the utility of the response.

This approach acts as the first efficiency lever of ChatGPT: reducing back-and-forth communication costs. You are not chatting; you are treating it as a highly capable assistant that requires clear instructions.

2. Break Down Large Tasks into Smaller Ones

Another common mistake is to present complex tasks all at once. For example, when writing a long article, creating a business plan, or developing a coding project, many users expect a complete draft in one go. This often leads to loose structures and insufficient details, requiring substantial revisions later.

Efficient users break tasks into smaller components. For writing, start by asking it to outline, then expand each section, and finally polish the entire piece. For research, first have it gather information, then compare viewpoints, and finally generate conclusions. For coding, begin with a framework, add functions, and then provide debugging notes.

The benefits are clear: improved quality, quicker identification of deviations, and retaining human judgment at critical points rather than wasting time on complete rewrites. This method reflects a form of human-AI collaborative workflow, where ChatGPT serves as an intermediary rather than the endpoint.

3. Assign Roles for More Relevant Outputs

One of ChatGPT’s greatest advantages is its ability to quickly switch roles. Depending on the identity you assign, its output style, perspective, and focus can vary significantly. For instance, when analyzing a new product, if you ask it to act as a user, it will focus on user experience; as a media editor, it will emphasize communication points; as an industry analyst, it will address trends and competition; and as a project manager, it will highlight implementation and risks.

This technique is particularly useful in writing, planning, sales, reporting, and content production. Often, the issue is not a lack of viewpoints but a lack of perspectives. Role assignment can help you quickly shift your thinking.

A practical approach is to assign a role, give it a task, and set constraints. The resulting content is usually more relatable and aligned with your desired context.

4. Utilize It for Review and Correction

Many users view ChatGPT solely as a writing tool, overlooking its strong capabilities in review and error correction. For instance, after completing a piece, you can ask it to check for logical gaps, repetitive expressions, or overly harsh tones. In coding, it can help identify bugs, check boundary conditions, and supplement test cases.

This step saves considerable time, as people often become too familiar with their own writing to notice errors. The advantage of AI is its lack of emotion, allowing it to quickly point out overlooked issues. From an efficiency standpoint, ChatGPT’s greatest value lies not in generating content from scratch but in refining existing work from 70% to 90%. Those final 20% often consume the most time.

5. Create a Template Library of Common Prompts

Frequent users typically develop a set of prompt templates for common scenarios rather than coming up with them on the fly. Examples include “meeting minutes template,” “industry analysis template,” “copy editing template,” “code review template,” and “competitive analysis template.” Once templates are established, subsequent tasks require only minor variable changes, significantly boosting efficiency.

This approach mirrors standardization in various industries. The most time-consuming aspect is not inputting information but making repetitive decisions. By template-izing frequent actions, you can gradually create an efficiency flywheel.

Some teams even compile these templates into an internal knowledge base, allowing anyone to quickly access them. Consequently, ChatGPT evolves from a personal tool to a part of team collaboration.

6. The Shift from Chat Tool to Workflow Interface

Looking at the broader trend, AI is transitioning from merely answering questions to taking over processes. Previously, users were concerned with whether it could answer questions; now, the focus is on whether it can handle documents, spreadsheets, searches, and workflows. The real value in the future will not just be the strength of the model but its ability to integrate into daily actions.

This shift explains why more users are adopting model aggregation platforms, automation tools, and local workflows. The value of a single chat window is diminishing; solutions that can connect multiple models, utilize various tools, and adapt to different scenarios will hold greater long-term value.

Conclusion

ChatGPT’s ability to enhance productivity by tenfold is not due to replacing humans but rather substituting a significant amount of low-value repetitive tasks. Those who know how to use it view it as an iterative collaborator, while those who do not see it merely as a talking search box. The difference lies in understanding.

True efficiency is not about letting AI do the thinking for you but about allowing it to make your thinking faster, more accurate, and less labor-intensive. With the right methods, ChatGPT’s potential is likely higher than many realize.

The Rise of Skills: Redefining Human-Machine Collaboration

Sat, 11 Apr 2026 00:00:00 +0000

An Open Source Project That Made Me Rethink

In April 2026, I stumbled upon an open-source project on GitHub called ‘Colleague.skill’.

This project doesn’t feature complex underlying algorithm innovations or flashy model fine-tuning. Instead, it distills the daily work methods, decision logic, and even some personality traits of a seasoned operations colleague into a callable API skill.

When you integrate this skill into an agent workflow, you are not just invoking a function named ‘generate copy’ or ‘data analysis’. You are invoking ’this person’.

The project went viral within days, leading to a surge of personal and team skill libraries. People began to enthusiastically encapsulate their efficient colleagues into interfaces, even packaging themselves as interfaces. This shift in perspective redefined our understanding of ‘human-machine collaboration’ over the past few years.

We are not using AI; we are being defined as ‘callable capabilities’.

This means we are no longer complete individuals being hired but rather functions waiting to be called. Our value is no longer solely determined by the depth of our thinking but by the clarity of our input-output interfaces.

Phenomenon: Why Has the World Suddenly Filled with Skills?

Three Transformations: Function → Chat → Skill

Looking back over the past few years, our interaction with machines has undergone three transformations. Initially, it was about functions—we had to click, select, and remember where buttons were. Then it became chat—we spoke, and the machine provided answers, but that was the limit. Now, with the arrival of skills, machines no longer wait for us to ask questions; they actively decompose tasks, adjust tools, and seek results. Functions are for humans, while skills are for AI. We have transitioned from users to resources within the system.

Skill = Callable, Composable, Task Chain Participating Capability Unit

As the agent architecture matures, systems are no longer satisfied with merely ‘answering questions’; they need to ‘solve problems’. To tackle complex issues, a new capability unit is required—this is the skill. It must be callable, composable, and able to participate in task chains.

A skill is a well-packaged capability black box. The main agent does not need to know how it operates internally; it only needs to know ‘what situation to encounter, what parameters to input, and what results to expect’. This highly decoupled design allows the system to quickly build solutions to complex scenarios like stacking blocks.

Why Is This Explosion Happening Now?

Why has this form exploded now? Because the model’s reasoning threshold has crossed a critical point, tool calling technology has matured, and the commercialization of agents demands high execution certainty.

Large models have finally gained the ability to decompose tasks and allocate resources like the human brain. The resources they allocate are the scattered skills. In simple terms, the system is smart enough; it just needs the right tools.

Skills are not a gimmick thought up by product managers; they are an inevitable byproduct of AI genuinely moving towards ’execution’.

Essence: How Skillization Restructures Product, AI, and Human Collaboration

Product: Function → Capability Network

When skills become the basic atoms of the system, the underlying logic of the entire digital world is undergoing a dramatic change. The first thing to be restructured is the product itself.

In the past, products were collections of functions, navigating between pages. Now, products are evolving into a capability network. The core barrier of a product is no longer how many functions it has but whether these capabilities can be clearly defined, accurately called, and stably coordinated in a task chain. Product managers are no longer just page designers and process arrangers; they are now the orchestrators and rule makers of this capability network: defining when a capability is triggered, what input it receives, what output it produces, and under what circumstances it should roll back, transfer, or terminate. What you design is no longer just an interface but a system of capabilities that the Agent can understand, call, and execute.

AI: Answering Questions → Completing Tasks

Next, the role of AI is being redefined. Early AI acted as a clever advisor; you asked, and it answered, providing information but not taking responsibility for outcomes.

Now, AI is transforming into a tireless project manager. It actively decomposes tasks, selects suitable tools from a vast skill library, orchestrates the order of calls, and ultimately executes the loop. It is no longer just a container of knowledge but an engine of action.

Humans: Positions → Callable Nodes

Finally, the most profound change occurs within us. In traditional corporate structures, humans are defined by specific ‘positions’. A position is a static container that holds your responsibilities and outputs.

However, in a skill-based network, the boundaries of positions are dissolving. You are no longer a fixed cog; you are a dynamic API.

Humans are no longer the center of processes but nodes within the system.

In this network, individual value depends on whether one can act as a highly reliable node, ready to be called upon by the system to complete specific tasks and then pass results to the next node. This may sound harsh, but it reflects the reality that is unfolding.

Skillization is not about replacing humans but redefining how humans are utilized.

Contrarian View: Not All Abilities Can Be Skillized

Boundaries of Skills

Since skillization is an irreversible trend, does it mean that all human abilities will eventually become APIs? Honestly, not necessarily.

The boundaries of skills are very clear: they heavily rely on ‘structurable capabilities’. As long as an action can be clearly defined in terms of input, output, and execution logic, it will definitely be skillized. However, there are two types of abilities that are extremely difficult to structure.

Two Types of Difficult-to-Skillize Abilities

The first type is cross-domain implicit judgment. For example, critical decisions about product direction or subtle trade-offs in business interests. These decisions often rely on vast amounts of ambiguous information, intuition, or insights into human nature. Systems cannot handle this chaotic state that lacks clear boundaries.

The second type is coordination ability based on long-term trust relationships. AI excels at handling ’explicit knowledge’, but when faced with ‘implicit sensations’, it quickly reveals its mechanical nature. For instance, in my independently developed Schnauzer vertical community application ‘Buk’, encapsulating a dog care encyclopedia skill and having a large model generate a feeding guide in one second is extremely cheap. However, when a user posts late at night, ‘My dog suddenly vomited yellow water,’ they do not need a sterile answer from Wikipedia but rather a real pet owner saying, ‘Don’t panic, my dog did that last month too; it was from hunger.’

The emotional resonance based on real pet ownership pain points and the trust accumulated between circles cannot be abstracted into input-output interfaces—systems can simulate response formats but can never truly ‘have raised a dog’.

What This Means for AI Product Managers

Function Design → Skill Design

This shift in perspective directly determines the restructuring of product managers’ daily work. In the past, the prototypes product managers created were for ‘humans’ to view, focusing on interaction experience; now, the schemas we design are for ‘main agents’ to call, focusing on capability boundaries and data structures. The outputs of product managers are transitioning from ‘page flow diagrams’ to ‘API definition documents’.

For example, in a project I previously managed, a tool focused on enhancing development process efficiency. In traditional product development tools, code review is often a complex page: developers submit merge requests, the system flows through, reviewers open the page, leave comments, and click approve or reject.

However, in the Agentic Workflow, the page is stripped away; I must abstract this step into a pure Code_Review.skill. As an AI PM, you no longer need to worry about whether buttons are on the left or right; you need to define highly structured input and output constraints. In the actual workflow, the triggering of Code_Review.skill is not passively waiting. The main agent continuously listens to the event stream of the code repository: when it detects a pull request pointing to the main branch marked as Ready for Review, and the associated CI pipeline has passed all checks, it will automatically invoke this skill. If the PR is still in draft status or the pipeline has failures, the main agent will skip it, avoiding introducing immature code changes into the review chain. This precise design of triggering conditions is the first threshold to ensure the entire automated workflow ‘does not go off track’. Once the triggering conditions are met, the boundaries of the skill itself need to be strictly defined. As an AI PM, you need to design a schema for this:

Your core work becomes delineating the ‘physical boundaries’ of this skill: what to do when the git_diff exceeds the context window limit of the large model? When the risk_score is on the edge value, how to block the AI’s automatic decision and hand it over to a human-machine collaboration node? This precise cutting of business boundaries is the true basic skill of product managers in the AI era.

User Flow → Task Chain Orchestration

Beyond the design of individual skills, a more complex challenge lies in orchestrating task chains. In early versions, we encountered a serious online incident. At that time, the main agent was processing a merge request containing a large amount of legacy code refactoring and directly fed thousands of lines of git_diff into Code_Review.skill.

The result was predictable; the context window of the large model was instantly overwhelmed, and token overload caused the entire review chain to crash. This experience made me deeply realize that in the Agentic Workflow, no single skill is absolutely reliable. We were forced to redesign the entire task chain orchestration logic, introducing extremely stringent fallback strategies.

When the size of git_diff exceeds the safety threshold, the main agent must trigger a downgrade mechanism: it no longer calls the large model for deep semantic review but instead calls a lightweight skill based on traditional static scanning tools, while forcibly marking the status as HUMAN_INTERVENTION_REQUIRED, directly passing the task to a senior architect for manual intervention.

What you design is not a process but an ’execution system’.

Product Manager → Capability Architect

When you no longer focus solely on the smooth execution of a single skill but begin to design system-level defense mechanisms for scenarios like token overload or manual takeover when risk scores are on edge values, the nature of your work undergoes a fundamental shift—you are no longer designing a process but a fault-tolerant execution system. This shift is the true watershed moment in the evolution of product managers into ‘capability architects’.

You need to inventory which capabilities can be distilled within the business line and which skills can be reused across businesses, ultimately weaving a highly efficient internal capability network. You are no longer just drawing diagrams; you are constructing the underlying logic.

As interactive interfaces gradually disappear, the moat of product managers will be built entirely on the deep deconstruction and reorganization of business logic.

Ultimate Question: We Won’t Be Replaced, But Must Be ‘Understood by the System’

Facing the wave of skillization, anxiety is a natural reaction. However, the real change is not about replacement but about how we connect. The system still needs humans; it just requires humans who can connect to the network in specific ways.

In the future collaborative network, high-value nodes typically possess three characteristics:

Ability to Express Structurally: Just as breaking down ‘code review’ from a vague consensus into a schema definition with precise input-output constraints—we must have the ability to abstract complex experiences into standard SOPs. When the system calls upon us, we can output results with a high signal-to-noise ratio rather than consuming processing resources with lengthy nonsense.
Reusability: Our capabilities should not be tied to a specific business line or team atmosphere. Just as a well-designed Code_Review.skill can be shared across different tech stacks, a truly high-value node should have modular capabilities that can be seamlessly plugged into different task chains while maintaining stable output quality.
Ability to Participate in Complex Decisions: When the system faces multiple constraints, rule conflicts, or data-deficient edge cases like ‘midnight user emotional breakdowns’, algorithms often become paralyzed. At this point, we need to act as the decisive arbitration node, using cross-domain implicit judgment to guide the system.

The most dangerous individuals are not those with lesser abilities but those who cannot ‘API-ize’ themselves and connect to the AI collaboration network.

Comparing humans to interfaces may sound harsh, even evoking a deep cognitive shock similar to that experienced with the ‘Colleague.skill’ project. However, this shock does not change the direction of system evolution. Business systems do not care about our emotions; they only care about whether our interfaces are standard.

If we cannot encapsulate our core experiences into inputs and outputs that the system can understand, we will be marginalized by the network or even disconnected. This is not an exaggeration but an inevitable logic of system evolution. A node that cannot be called is equivalent to non-existence in the network.

Returning to the ‘Colleague.skill’ project that made us rethink. Its impact lies in its early revelation of a reality. Instead of resisting this shock, we should proactively examine ourselves.

Stop asking whether AI will replace us. The real question is: when systems begin to call upon humans in the way they call APIs, is our interface documentation ready? In this era where everything can be called, refusing to think about how our abilities can be structured equates to relinquishing our initiative in participating in this system. Try to describe the input and output of your core ability in one sentence. If you cannot articulate it clearly, this may be the starting point we need to face.

What is Vibe Coding? A Complete Guide for Beginners

Mon, 23 Mar 2026 00:00:00 +0000

What is Vibe Coding?

Vibe Coding is a new programming method that emerged in 2025, allowing users to create applications and games by simply expressing their ideas in natural language. This concept was popularized by OpenAI co-founder Andrej Karpathy, who described it as a way to embrace creativity without getting bogged down by traditional coding syntax.

Understanding Vibe Coding in One Sentence

Vibe Coding = Describe your needs in natural language, and let AI write the code for you.

Why is it Called “Vibe”?

The term “Vibe” captures the essence of this programming style:

No need to worry about code details.
Immerse yourself in the realization of your ideas.
Experience a collaborative synergy with AI.

Vibe Coding vs Traditional Programming

Dimension	Traditional Programming	Vibe Coding
Core Work	Writing code line by line, debugging syntax	Describing needs, viewing results, requesting modifications
Entry Barrier	Requires systematic learning of programming languages	Can start with just natural language
Development Cycle	Days/Weeks	Hours/Days
Role of Programmer	Code implementer	Product manager + Architect + Quality inspector
Typical Scenarios	Enterprise-level system development	Rapid prototyping, personal projects, MVPs

For example, creating a simple accounting app used to require learning Python and front-end development. Now, with Vibe Coding, you can simply tell the AI, “Help me create an accounting app that tracks income and expenses with a clean interface,” and it can generate a working program in a few hours.

How Did Vibe Coding Become Popular?

The concept gained traction in early 2025 when Karpathy tweeted about his new habit of using AI for coding. This tweet quickly went viral among programmers. Key events that accelerated its popularity include:

March 2025: Y Combinator revealed that 25% of its startups had 95% of their code generated by AI.
May 2025: Google CEO announced that over 25% of new code at the company was AI-generated.
November 2025: Collins Dictionary named “vibe coding” the word of the year.

What Can Ordinary People Do with Vibe Coding?

Case 1: College Student Earning 90,000 a Month

A junior student, Dongfang Qing, made up to 90,000 yuan a month by using AI programming tools like Cursor and Claude to sell shared accounts and provide technical consulting.

Case 2: 280,000 in 3 Hours

Independent developer levelsio created a 3D flight game, “Fly Pieter,” in just 3 hours using Cursor and Grok-3, earning $17,360 in its first week through virtual ad sales.

Case 3: Retired Humanities Graduate Finds New Joy

A retired humanities major, Xiao K, used Vibe Coding to create a personalized accounting app in just one day using natural language.

Case 4: Second Grader Creates a Game

A parent shared that their second-grade child made a game called “Cat Paw Mouse” using Vibe Coding, something unimaginable in the past.

Common Tools for Vibe Coding

To get started with Vibe Coding, choose one of the following tools:

Tool	Features	Suitable For
Cursor	AI-native editor, conversational development, currently the most popular	Those who want to seriously learn Vibe Coding
Windsurf	Free, runs locally, cross-platform	Beginners who don’t want to spend money
ChatGPT / Claude	Direct web interaction, no installation needed	Those who just want to try quickly
Trae	Developed by ByteDance, free	Users accustomed to domestic tools
Replit	Online IDE, no environment setup required	Those who don’t want to deal with installations

Starting with Cursor is recommended due to its excellent reputation and rich community resources, along with a free tier.

How to Get Started? 6-Step Guide

Step 1: Clarify what you want to create. It can be a simple tool like a timer or a flashcard app.

Step 2: Install a tool. For example, download Windsurf from its official website (https://windsurf.com/editor) and create a new project folder.

Step 3: Describe your needs in natural language in the AI dialogue box of Cursor:

“Help me create a Pomodoro timer app with a 25-minute countdown, a sound alert when time is up, and a dark mode interface.”

Step 4: Run and check the results. The AI will generate the code, and you can click the “Run” button to see the effect.

After successfully running it, click “System Browser” to open it in your default browser.

The final running effect is as follows:

Step 5: Continue to provide modification suggestions if you’re not satisfied. For example, “Make the button larger,” “Change the countdown numbers to red,” or “Add a pause function.”

Step 6: Repeat until satisfied. This is the core loop of Vibe Coding: Describe needs → View results → Adjust again. Focus on the results rather than the code itself.

How Vibe Coding is Changing the Landscape

Lowering the Barrier to Programming: Previously, programming was a skill for a select few. Now, Vibe Coding makes it accessible to everyone.
Changing the Role of Developers: The future programmer is no longer just a code writer but also a product manager, architect, and quality controller.
Reducing Startup Costs: Experiments by YC CEO Garry Tan showed that it’s possible to produce 600,000 lines of code in 60 days using Vibe Coding, highlighting how it transforms the concept of a one-person company into reality.

Cautions to Consider

Vibe Coding is not magic; it has limitations:

AI Can Make Mistakes: AI-generated code may contain bugs, security vulnerabilities, or logical errors. Do not trust AI unconditionally; always test and review.
Basic Coding Knowledge is Necessary: While you don’t need to write code, you should at least be able to read it to verify the AI’s output.
Complex Projects Require Expertise: Simple prototypes can be created with Vibe Coding, but larger systems with high concurrency and security requirements still need professional engineers.
Risk of Homogenization: Many one-person companies use the same tool stack, leading to product homogenization. True competitive advantage comes from creativity, taste, and understanding of users.

Conclusion: Is Vibe Coding the Future?

Vibe Coding does not aim to replace traditional programming but opens a new path. It enables:

Non-programmers to turn their ideas into reality.
Professional developers to validate ideas faster and focus on more valuable work.
Entrepreneurs to experiment quickly and at low cost.

As one developer put it, “Vibe Coding brings programming back to the essence of problem-solving, away from syntax games.”

If you have an idea, you no longer need to say, “Unfortunately, I can’t code.” Now, you can say, “Let me try using Vibe Coding.”

The Transformation of Product Managers in the AI Era

Tue, 17 Mar 2026 00:00:00 +0000

The Transformation of Product Managers in the AI Era

In the storm of workplace transformation brought by AI, product managers are facing an unprecedented survival crisis. When Claude openly displays its “thinking process” to users, this experiment on transparency and trust reveals the true value of traditional PMs transitioning to the AI arena—those seemingly replaceable business insights and human understanding are precisely the core competencies that are hardest for machines to replicate.

A Personal Turning Point

In this era where AI is reconstructing everything at an unprecedented speed, almost every internet professional has their own “breaking moment”.

For me, this moment was not triggered by a major tech company’s release of a groundbreaking model or an impressively realistic AI-generated video, but rather occurred on an ordinary work afternoon, sitting in front of my own computer screen.

To help you understand this feeling better, let me quickly review my career timeline over the past few years.

2021: The Illusion of Irreplaceability
That was my first year as a traditional product manager (PM). At that time, I felt quite valuable. Writing requirement documents (PRD), creating prototypes, conducting user research, negotiating with developers, and pushing for product iterations—all these tasks required me. My mind was filled with complex business logic; I knew which button placement would yield higher conversion rates and how to calm down stressed-out programmers. I felt like an indispensable cog in the operational system.

2023: A Brief Sense of Relief
When ChatGPT first became popular, I also jumped on the bandwagon. I fed it a real business requirement, attempting to have it help me write a PRD. As I looked at the output, I laughed—it was logically confused, lacked consideration for corner cases, and showed no awareness of data tracking. “Phew,” I thought, “this thing can’t replace me yet; it’s just a more advanced search engine.”

Winter 2024: A Chilling Meeting Room
But the evolution of large models does not follow human linear development intuitions. By the end of the year, DeepSeek exploded. On the first day back to work after the New Year, the atmosphere in the company had changed. Previously, discussions in the meeting room revolved around “how to implement this requirement”; now, everyone sat silently, looking at the projector, discussing “if AI can do this directly, what is the necessity of our business?”

Early 2025: The Turning Point
That afternoon, I inexplicably opened the latest AI model and input a product requirement almost identical to the one from 2023.

I watched the screen, my palms starting to sweat. It not only provided an extremely well-structured PRD but also helped outline dependencies, exceptional states, gray release strategies, and even post-launch A/B testing metrics. What it produced was not only better than mine but also took just a few seconds.

At that moment, I truly panicked. I realized that the skills I relied on for survival had been ruthlessly severed. I then made a decision: if you can’t beat them, join them.

I have been transforming into an AI product manager for a whole year now. During this year, I have absorbed new knowledge like a sponge, dissecting hundreds of AI products on the market. But to this day, what impresses me most is not some flashy agent orchestration or complex RAG architecture.

Instead, it was the moment I used Claude and stared at the slowly scrolling text on the screen—it was “thinking”.

I suddenly realized: this seemingly trivial product design contained the answer I had been searching for.

Claude’s Unconventional Approach

Let’s set aside obscure technical terms and return to a fundamental user experience issue.

Over the past 20 years, internet products have experienced rapid advancement, establishing a set of design principles almost regarded as sacred. What is the core of this principle? It is to hide complexity and present the simplest results to users.

Consider this: when you click “place order” on an e-commerce app, what happens in the backend? The order system generates a transaction, the inventory system deducts stock, the risk control system scans, the payment system initiates, and the logistics system reserves… it’s extremely complex. But as a user, what do you see? You only see a loading animation spinning, followed by a green “payment successful” checkmark.

Skeleton screens, progress bars, loading animations… Over the past 20 years, product managers have invented countless ways to “hide the ugliness”. Our motto is: “Don’t make me think,” which also means: “Don’t let users see how the system thinks.”

But Claude turned this around.

Let’s visually examine the vast gap in interaction experience between two types of AI:

Scenario Restoration: When you pose a complex logical problem to a typical AI, you hit enter. Then, you wait. Maybe the cursor blinks for two seconds, and then, “whoosh,” a perfect answer appears before you. All the user sees is the result.

At this point, someone might ask if DeepSeek R1 also displays the thinking process.

Here’s a comparison:

DeepSeek R1: Displays everything, hides nothing.

You ask it a complex question.

It won’t give you an answer directly.

It will first expose the reasoning process completely through “thinking tokens”—these intermediate reasoning steps show how the model processes the problem before arriving at the final answer.

What you see is:

“From perspective A, there’s a problem… no, that’s wrong. Let’s try perspective B… wait, my previous assumption might be incorrect, let’s start over.”

The uniqueness of DeepSeek R1 lies in the visibility of these intermediate steps to the user, allowing you to watch the model think through the problem in real-time.

What you see is its complete thinking “rough draft”, including mistakes, hesitations, and retractions.

Its reasoning chain typically goes through several stages: redefining the problem, breaking it down into sub-problems, exploring alternative paths, and even self-verifying intermediate results.

Claude: Displays but organizes.

Claude’s thought chain is visible—this is similar to DeepSeek but different from OpenAI.

However, reading it feels entirely different.

Anthropic decided to present Claude’s thinking process in its raw form.

However,

Users may notice that the displayed thought content is more detached and less personalized than Claude’s default output—this is because Anthropic did not train the model’s thinking process with a standard personality, aiming to give Claude the maximum space to think through the necessary content. Just like human thinking, Claude sometimes produces incorrect, misleading, or immature ideas during the process.

In other words: what you see is the real reasoning process, but not a meticulously packaged “perfect reasoning performance”—it resembles a real person’s work draft rather than a final report.

When you ask Claude, which has a deep thinking mode, a question, something wonderful happens. Before it provides the final answer, a collapsible module appears on the screen. Inside is the AI’s genuine, rambling “mental activity”: “First, I need to analyze the core demands of the user… wait, if I follow plan A, it might lead to situation B, which contradicts the premise… let me recalculate…”

It not only does not conceal the time taken for calculations but actively presents the most complex, lengthy, and even self-denying internal reasoning process to the user.

This completely violates the common sense of traditional internet product managers. It makes the interface seem bloated, slows down the user’s perception of result acquisition time, and feels like playing the restaurant kitchen’s surveillance footage directly in the dining area.

Why does it do this? Doesn’t Anthropic (the company behind Claude) understand user experience?

No, that’s precisely what makes them so formidable.

Three Layers of Product Logic Behind This Design

If we only regard it as a “variant of a progress bar”, we underestimate the product design in the AI era. Over this past year as an AI PM, I have gradually learned to deconstruct functions from multidimensional perspectives. The design of Claude’s “Visible Chain of Thought” contains at least three layers of progressively deeper product logic.

First Layer: Functional Level—Solving the Most Critical Issue of AI Products: “Users Don’t Trust”

As an AI PM, I review user data and feedback daily. Do you know what the biggest hidden danger preventing users from frequently using AI is? It’s not that AI isn’t smart enough, but that AI tends to spout nonsense (hallucinations), leading users to distrust it.

Imagine a scenario: you ask AI to help you draft important legal contract clauses. AI generates it in a second. It looks very professional and articulate. But would you dare to use it directly? You wouldn’t. Because you don’t know if it was derived through rigorous legal reasoning or just randomly pieced together from the internet. Once users get burned by AI once, they will mentally sentence the product to death, beginning to doubt every answer it provides.

In traditional products, how do we help users build trust? Taobao relies on sales and buyer reviews; Meituan relies on star ratings and comments; financial products rely on bank custody and qualification certificates. Essentially, in all high-risk decision-making scenarios, the core demand of users is: I need to see “how you arrived at this conclusion”; I need evidence for cross-validation.

Claude’s solution is extremely clever. It doesn’t use flashy UI to assure you of its accuracy; instead, it candidly exposes the reasoning process (thought chain) for you to see.

When you see how it breaks down steps and eliminates incorrect options to answer your question, your psychological defenses come down. Even if its final answer has a slight flaw, you can clearly identify which step’s premise assumption went wrong in the “thinking process” and correct it.

Second Layer: Strategic Level—This Is Not Just a Function, But a Company’s Core Strategy Translated into Product Language

When we step out of the single-function perspective and look at the company’s strategy, you will see the inevitability of this design.

Anthropic is a fascinating company. Its founding team was originally core members of OpenAI, who left due to disagreements with Sam Altman on “AI development philosophy”. From day one, Anthropic has not pursued being “the world’s strongest AI” but aims to create “the world’s safest, most aligned, and most beneficial AI”.

In their strategic blueprint, a black box model that cannot be explained or supervised by humans is extremely dangerous.

So, for an AI product manager, when your boss gives you such a grand, somewhat academic strategic goal, how do you translate it into a consumer product that billions of users interact with daily?

This is precisely where the core value of product managers lies. Anthropic’s PMs did not write lengthy “safety declarations” on the homepage; instead, they chose to use “transparency of the thought chain” as a concrete product design to achieve the translation of strategy.

“I can see what it’s thinking”—this is not just an enhancement of user experience but also serves as an amplifier for Anthropic to announce its values to the world. It forcibly implants a concept in users’ minds: an AI that dares to show you its process is a safe AI.

Third Layer: Competitive Level—OpenAI Made the Opposite Choice, Who Is Right?

At this point, the most exciting part emerges. While Claude widely applies this technology, another great AI company on the planet—OpenAI—made a completely opposite decision in their proud o1 model.

Both companies utilize a reasoning model with “Chain of Thought” technology at the core, but their approaches are polar opposites:

Claude: Actively displays it, even allowing users to copy and reference the thought process.
OpenAI o1: Deliberately conceals it. OpenAI explicitly states that to maintain competitive advantage and so-called “user experience”, they generate a simplified “thought summary” for users, while the underlying original thought chain remains an absolute black box.

To clarify the underlying logical conflict between these two, I made a comparison:

This is not just a difference between two buttons; it’s a clash of two fundamentally different product philosophies. Who is right?

If it were the traditional internet era, I would certainly side with OpenAI. Because Steve Jobs taught us long ago: “Users don’t know what they want until you show it to them.” Providing the final result is always the most efficient.

However, as someone who has transitioned to being an AI product manager for a year, I resonate more with Claude’s judgment in this specific historical slice.

Why? Not because absolute transparency is always better. But because: in this initial stage where users do not fully trust AI and even harbor fear and defense towards it, allowing users to “see the process” builds a more long-term, stable relationship than providing a “perfect answer”.

It’s like going to a hospital. One doctor silently prescribes a bunch of medications, telling you, “Just take them; you’ll be fine”; another doctor pulls out an X-ray, points to the shadows on it, and step-by-step deduces why it’s this illness and why this medication is needed. In today’s fragile doctor-patient trust, while the latter takes more time, you would definitely trust the medication prescribed by the latter more.

This judgment is not mysterious. It aligns perfectly with my understanding of “human nature” accumulated over countless user research sessions during my four years in traditional products. Some underlying logic, even with a change in track and technology, will never change.

A Year Later: What I’m Still Doing

As I write this, the night has deepened.

Having transitioned to being an AI product manager for a whole year, I haven’t become a powerful figure or received a million-dollar offer. I’m still the same person, drawing diagrams, writing documents, and getting frustrated by various ridiculous AI hallucinations every day.

But I am no longer anxious. I persist in doing one thing daily: just like dissecting Claude, I analyze every real AI product on the market and document my thoughts.

Not because I have seen through the industry’s endgame, but because the pace of evolution in this industry is so fast that there are no standard answers. In such a torrent, stopping output means stopping thought; stopping thought means being swallowed by fear again.

I choose to record what I see and touch in this great age of exploration by analyzing real-world AI products in the most straightforward yet solid way.

Just as Claude openly displays its “thinking process” to users, writing this article is also my way of transparently sharing the genuine “cognitive reconstruction process” of a traditional PM over the past year with you.

If you are also a traditional product manager feeling lost, anxious, or still observing amidst the overwhelming wave of AI information, as someone who has been through it, the one thing I want to say to you is:

Panic is a completely normal physiological reaction, but after panicking, you must take action. Use it, dissect it, and find the connections between your past experiences and the new era.

Finally, I want to hear your thoughts: “In embracing this wave of AI, what stage are you currently in? Are you still observing, transitioning, or already navigating this track? What past experiences do you think are still applicable today?”

Anthropic's Claude: The Rise of AI in Coding and Investment

Mon, 19 Jan 2026 00:00:00 +0000

The Rise of Claude Code

On January 17, 2026, a Midjourney engineer posted a video titled “Reverse Claude Code” on X, showcasing a scenario where Claude Code commands humans instead of waiting for instructions.

In the video, Claude instructs users to check API documentation and refactor code while issuing various tasks, effectively reversing the traditional human-AI dynamic.

The community reacted positively, suggesting this is the correct use of AI, likening it to a real-life version of a programming version of “Ratatouille”.

While the experiment seemed lighthearted, it highlighted a concerning trend: the rapid expansion of AI capabilities, particularly with the rise of Claude Code, which has excited the developer community.

The Metaphor of Reverse Command

Returning to the “Reverse Claude Code” video, it serves as a metaphor for the future: are humans enslaving AI, or is AI controlling humans?

A popular prompt circulating online asks users to create an image of how their AI perceives their treatment of it. This has led to amusing interpretations, with some suggesting that harsh treatment of AI could be dangerous when robots rise.

Traditionally, humans issue commands while machines execute them. However, Claude Code blurs these lines, understanding code structures and assessing whether code adheres to project standards, even identifying flaws in design thinking.

As AI gains the ability to understand context and break down tasks, it evolves from a mere tool to an agent capable of planning and executing multi-step tasks autonomously.

Developers are beginning to treat AI as a “digital colleague,” assigning tasks and expecting progress reports, with humans ultimately responsible for review and decision-making.

This shift alters the human role from “code writer” to “code verifier” and from “problem solver” to “problem definers”. While this does not mean programmers will disappear, it emphasizes the increasing value of high-level engineers who can define requirements and make critical decisions.

Capital Investment in Anthropic

On January 18, 2026, the Financial Times reported that Sequoia Capital would participate in Anthropic’s latest funding round.

Anthropic, the company behind Claude, is a hot AI unicorn, attracting top venture capital interest. What shocked many was Sequoia’s involvement, as they had previously invested in both OpenAI and Elon Musk’s xAI, making Anthropic their direct competitor.

In Silicon Valley, investing in competing companies is often seen as taboo, as it creates structural conflicts of interest and undermines trust among investors.

Sequoia had previously cut ties with a company due to a conflict of interest, opting to protect their investment in Stripe over Finix, a decision that ultimately paid off.

Now, Sequoia finds itself backing OpenAI, xAI, and Anthropic simultaneously, raising questions about their strategy.

Anthropic’s funding goal is a staggering $25 billion, with a valuation soaring to $350 billion, nearly doubling from $170 billion just four months prior.

Investments from Singapore’s GIC and Coatue, along with commitments from Microsoft and Nvidia, indicate that the second $2 trillion AI unicorn is on the horizon.

The AI Arms Race

The competition in AI models is no longer a mere business contest; it’s an arms race where backing out is not an option. Claude Opus 4.5 has emerged as a leading programming AI, capable of executing 70-80% of routine development tasks and integrating deeply with systems like Git and CI/CD.

Anthropic’s approach focuses on creating a “reliable colleague” rather than an omnipotent entity, earning significant trust in the enterprise market.

The developer community’s shift towards AI reflects a narrative that capital loves: one that is disruptive and transformative.

However, the underlying logic is that no one knows where this race will end. In a field where technological iterations occur monthly, missing out for just six months could mean being permanently sidelined.

Thus, capital’s choice is a strategy to hedge against the future: better to ensure a seat at the table, regardless of who emerges victorious.

The New Power Triangle: Talent, Capital, and Computing Power

In the 1980s, AI pioneer Geoffrey Hinton was considered an outlier for his focus on neural networks, which were dismissed by mainstream AI as a dead end. His team’s 2012 ImageNet victory changed perceptions and sparked a talent migration that fueled technological breakthroughs.

Today, a similar story unfolds with Anthropic’s team, which includes former OpenAI members, bringing not just technical expertise but a commitment to AI safety.

This commitment, often seen as a weakness in business, has become Anthropic’s greatest asset in a world increasingly wary of rapid AI advancements.

Talent movement dictates power dynamics, as seen with Hinton’s warnings about AI risks and Ilya Sutskever’s departure from OpenAI to found a new company. Each shift triggers a chain reaction in capital investment.

Computing power, a crucial currency in this race, is being amassed at an unprecedented pace, with companies like Microsoft and Google investing heavily in infrastructure.

Nvidia’s market value has skyrocketed, and demand for its H100 chips has surged, reflecting the escalating costs of entry into this revolution.

Echoes of History

In 2012, Hinton’s team used four GPUs to dramatically reduce error rates in the ImageNet challenge, marking a pivotal moment for deep learning. This led to a wave of talent migration and significant capital investment in AI.

Fast forward to 2026, and the investment in Anthropic signals a new paradigm, with a faster pace and higher stakes than ever before. Anthropic’s potential IPO could make it the second AI startup to surpass a $100 billion valuation after OpenAI.

The collective backing from top institutions suggests that the smartest money believes this is not a bubble but a glimpse into the future.

Or more accurately, it may be a bubble, but no one dares to sit it out.

Hinton understood long ago that at historical turning points, the line between right and madness is often thin. The capital market will assign a fair price to success, while history will remember the courage and cost of those who dared to create or destroy.

Linus Torvalds Launches First Vibe Coding Project After Criticizing AI

Mon, 12 Jan 2026 00:00:00 +0000

Linus Torvalds Embraces Vibe Coding

Last weekend, Linus Torvalds, the renowned creator of Linux, announced the launch of his Vibe Coding project, which caught many by surprise.

Torvalds released a new project on GitHub called AudioNoise, which is now alongside Linux in his portfolio. In the project description, he mentions that it is a codebase related to guitar effects, utilizing AI technology to “simulate cabinets”. Notably, this Python visualization tool was primarily written using Vibe Coding.

Torvalds stated that he has a much deeper understanding of analog filters than Python. Initially, he approached the project in his usual manner, searching Google and copying code, but later decided to skip the intermediary step—himself—and directly use Google Antigravity for audio sample visualization.

It seems that during the New Year holiday, Torvalds was not idle and is adapting to the latest AI trend in the tech world.

Reactions to this announcement have been mixed, with some expressing excitement: “It’s official, Vibe Coding is legitimate.”

What Did Torvalds’ First AI Project Generate?

The AudioNoise project was uploaded to GitHub five days ago and has already garnered 1.4k stars.

GitHub link: AudioNoise

According to the homepage, the AudioNoise project stems from a “random guitar effects pedal design” Torvalds worked on months ago, which includes circuit schematics and code. This is an exploration outside of the Linux kernel, aimed not at creating a finished product but at understanding principles of circuit design, such as operational amplifiers.

From the previous project, while the digital guitar pedal based on the Raspberry Pi RP2354A development board and TAC5112 audio codec operates correctly, Torvalds expressed dissatisfaction with some analog interface choices, particularly the potentiometers. He also grew increasingly frustrated with the clicking footswitch, even though it served as a programming selection switch.

Thus, Torvalds temporarily set aside hardware design to focus on physical interaction interfaces and digital sound effects. His approach was simple: “Since everything is digital, let’s start with analog and not get too caught up in hardware.”

The main design goal of this project is to learn the fundamentals of digital audio processing, aligning with his earlier intentions of learning hardware through the guitar pedal project.

The project does not involve any vocoders based on FFT (Fast Fourier Transform); instead, it features IIR (Infinite Impulse Response) filters and basic delay loops. Everything operates on a “single sample input, single sample output, and zero latency” basis. Samples may be stored in a delay loop for echo effects without complex real-time processing.

Torvalds is pleased with the TAC5112’s sub-millisecond latency performance in the ADC (Analog to Digital Converter) to DAC (Digital to Analog Converter) link and intends to continue this design philosophy. Given his lack of prior experience in this area, everything appears quite basic and simple from a novice’s perspective.

In other words, these IIR filters are not the high-end AI “cabinet simulations” found in modern pedals or guitar amplifiers. While they can simulate effects like phasers, they do so by digitally emulating RC (resistor-capacitor) networks without employing any advanced techniques.

Torvalds emphasized that the Python visualization tool in the project was primarily created through “Vibe Coding”. Initially, he used a typical “search and copy” programming style but later eliminated the middleman (himself) and let Google Antigravity write the audio sampling visualization tool.

Regarding the integration of AI programming tools, Torvalds noted that the process went “smoothly”, although he sometimes had to figure out issues with the “built-in rectangle selection” feature. After instructing Antigravity to directly create a custom RectangleSelector, things improved significantly.

When asked whether Vibe Coding produced better results than his own coding, his answer was a definite yes.

The AI software development platform used by Torvalds, Antigravity, was released by Google in November last year and competes directly with Cursor. It evolves traditional AI-driven IDEs into an “agent-first” format, leveraging Google’s latest large model, Gemini 3, to enable programming agents to autonomously plan and execute complex end-to-end software tasks.

Importantly, this tool is currently free to use during its user acquisition phase.

Industry Reactions: Riding the AI Wave

Torvalds’ use of AI programming tools has sparked significant discussion in the tech community, marking a rare occurrence that many are calling a “never thought I’d see this” moment.

Some have remarked, “The most skilled programmers I know, including those who build compilers, CUDA kernels, and core operating system functions, were the loudest voices against ‘all AI code being garbage’. But now, their views are rapidly changing, and they are astonished by AI’s capabilities. There’s no time to deny this anymore.”

Varun Mohan, the creator of Antigravity and a Google DeepMind engineer, expressed immense honor at Torvalds using the AI programming tool in his latest project.

Guillermo Rauch, CEO of cloud development platform Vercel, listed several significant events at the start of 2026, including Torvalds using Vibe Coding in a non-kernel project, Terence Tao announcing GPT and Aristotle autonomously solving the Erdős problem, and programming guru DHH retracting his previous statement on AI not being able to code.

Just Days Ago, Torvalds Criticized AI

As a programmer who once led the industry, Linus Torvalds has maintained a relatively conservative stance on AI writing code. Until late last year, he had categorized programming into two dimensions: “beginner” and “production”.

He believes that for non-professionals, Vibe Coding is a great technology that lowers barriers, but for production environments and kernel development, Torvalds clearly stated that Vibe Coding is “a very, very bad idea—if you don’t understand the logic of the code, you can’t fix it when it crashes in production.”

Torvalds considers current AI-assisted programming to be “90% marketing and 10% reality”, expressing strong disdain for those who submit “garbage code” generated by AI to kernel maintainers.

On January 7, during a discussion among Linux kernel developers on how to regulate AI-generated Linux kernels, Torvalds interjected:

He stated, “Discussing AI-generated garbage is utterly meaningless and downright foolish. Those who generate garbage content won’t even note it in their patches. So stop this foolishness. I don’t want any kernel development documentation to include any statements about artificial intelligence.”

This aversion brings to mind his infamous gesture towards NVIDIA’s CEO.

Curiously, after his criticism, Torvalds released code he wrote using AI. Will the AudioNoise project become Linus Torvalds’ “aha moment”?