36Kr Reviews AI Bubble Concerns: Tech Giants Face Cash Flow Pressure Amid Frontier Model Research Breakthroughs

In the US stock market tech earnings season of late October 2026, a picture emerged in which rapid business growth coexisted with a sharp sell-off in the capital markets.

Alphabet's third-quarter earnings, released on October 28, showed Google Cloud revenue surging 103% year over year, yet free cash flow for the quarter came in at negative $9.1 billion, while the core search advertising business underperformed market expectations. The next day, Alphabet's stock plunged 12.6%, as the cloud business's doubling of growth failed to offset investors' concerns over runaway cash burn.

A day later, Amazon delivered a third-quarter report card with a highly similar structure: AWS cloud services revenue grew 39% year over year to $46 billion (an annualized revenue run rate of $184 billion), and the group's overall operating profit hit a record high of $28.1 billion. But over the trailing twelve months ended September 30, although operating cash flow grew 29% to $168.3 billion, the company's free cash flow reversed from a net inflow of $14.8 billion in the same period a year earlier to a net outflow of $42.4 billion.

Of these cash shortfalls, Alphabet's negative $9.1 billion reflects a single-quarter cash gap, while Amazon's negative $42.4 billion covers cumulative outflows over the past twelve months. During this period, Amazon's purchases of property and equipment, net of asset disposals, surged by $94.8 billion year over year, and full-year cash capital expenditure is expected to reach as high as $220 billion.

Microsoft's fiscal 2027 first-quarter earnings, released on October 28, were able to become one of the few positive examples. Its Azure revenue grew 48% (exceeding its own guidance of 44% to 45%), group revenue totaled $92.3 billion, and it still maintained positive free cash flow despite capital expenditure of as much as $45.2 billion in the quarter. The key lies in the extremely thick current cash buffer provided by its massive enterprise software and office subscription base. As a result, its stock rose 3% after the earnings report.

This therefore shows that the market is currently very worried about big companies' moves to invest in AI infrastructure beyond what their profitable cash flow can support. It is not that they do not believe computing power is currently in short supply, or that computing centers are a money-making business. All of this is a false prosperity. But their confidence in the future is not so great.

For these big companies, they must invest heavily in future infrastructure and computing power to remain competitive. But whether these massive investments can ultimately be converted into effective returns is hard to say.

First, at the current growth rate, they may not make much money. According to Gartner's estimates based on current AI revenue growth, by 2030, infrastructure investment by AI companies and cloud businesses may reach $7 trillion, with total revenue possibly at $8.5 trillion, while enterprise and consumer AI revenue attributable to AI data centers will not exceed $3.6 trillion. Taking into account operating costs, depreciation, and taxes, the effective profit margin is likely below 12%. In other words, the efficiency of making money is relatively low. Considering opportunity cost, this may be less profitable than investing in other directions.

Second, this is still under the expectation that demand for computing power will continue to grow enormously. What if AI development is hindered and expansion stalls? What if the number of players decreases and training computing power usage, which currently accounts for more than 50%, drops sharply? What if there is major progress in infrastructure and demand for computing power falls? These risk factors will clearly further weaken the expectation of stable returns from such large-scale upfront investment.

Simply put, even when we calculate this based on the best current case, it may not necessarily be all that profitable. The risk is too great, and the biggest gains are not obtained through computing power expansion alone. If one company wins the game, while the others can only participate as computing power providers, their upside is limited while their downside is very low.

English text is being prepared.

Similarly, we can also understand why Anthropic is rated as junk-grade credit.

On October 30, the Financial Times and several market institutions disclosed that Blackstone Group, in communications with investors, expected that even after completing an IPO, Anthropic's credit rating may remain at non-investment grade (that is, the upper range of what is commonly called "high-yield" or "junk"), making it difficult to obtain an investment-grade rating in the short term. This is also because Anthropic's IPO prospectus (Form S-1), publicly filed on October 26, shows that the company carries high-value equipment lease commitments lasting five years (the five-year lease total related to Broadcom alone amounts to $161.2 billion, excluding interest), and the revolving credit agreement it established during this period has added to these doubts.

Because Anthropic has no positive free cash flow, it can only rely on borrowing to build its computing power foundation. And this borrowing requires guarantees to back it up.

On October 26, Broadcom, together with Bank of America, Blackstone, Citigroup, and Morgan Stanley, announced an equipment financing arrangement of up to $63 billion to deploy dedicated AI compute and networking equipment based on Broadcom technology for Anthropic, with deliveries expected to begin in 2027.

Broadcom, the chip seller, personally acted as the backstop for the senior lenders. In an 8-K filing submitted to regulators, Broadcom committed to providing a residual value guarantee of up to approximately $43 billion (undiscounted) for the senior loans. If a default occurs and the proceeds from selling the racks fall below 72.5% of the outstanding lease principal, Broadcom must make up the difference out of its own pocket.

Similar circular borrowing further intensifies the contagion risk of infrastructure fragility. If Anthropic fails, Broadcom will suffer too.

Doubts about the ability to pay for future purchases quickly spread to the upstream hardware supply chain. At 09:45 a.m. Eastern Time on October 30, memory chip and hard drive stocks suffered a broad sell-off; intraday records show Micron fell 4.26%, Sandisk plunged 8.53%, Western Digital fell 6.07%, and Seagate fell 5.26%. Tellingly, just after the close the previous day, Sandisk had reported an extremely strong fiscal 2027 first-quarter earnings: quarterly revenue reached $11.031 billion, up 23% quarter over quarter and surging 378% year over year (with data center business revenue up 61% quarter over quarter to $4.802 billion), GAAP gross margin jumped from 29.8% a year earlier to 85.0%, and the company also announced an 8-for-1 stock split and raised its revenue guidance for the next quarter to $12.2 billion to $12.6 billion. The strong results not only failed to lift the stock price but instead triggered profit-taking and a sector-wide sell-off.

Because everyone is anticipating that this frenzied, aggressive infrastructure spending may not be sustainable. And the future expectations for the hardware suppliers that have surged on the back of it will naturally be lowered accordingly.

English text is being prepared.

Unlike earlier AI, which was limited to local engineering optimization under given evaluation metrics, this round of results struck directly at core conjectures and experimental black boxes that have plagued the human academic community for decades. From the two institutions independently overturning the cornerstone of computational complexity, the Exponential Time Hypothesis (ETH), on the same day, to fully conquering the local smoothing conjecture and the Kakeya conjecture in harmonic analysis across all dimensions.

From a 1,147-page existence construction for four-dimensional quantum field theory, to a formal proof of global convergence of chemical reaction networks completed through cross-model relay, to deep automation in wet labs overturning the erroneous scoring standards of traditional instruments.

A complete scientific research chain extending from "formal axiomatic deduction" to "experimental design in the real physical world" is being rapidly connected.

On October 26, OpenAI dropped a bombshell in its official announcement "Working with academic communities to build scientific understanding," publicly releasing a 67-page preprint titled "The Exponential Time Hypothesis is False" along with accompanying Lean 4 formal proof code, announcing that its internal reasoning model had found a natural-language proof overturning the Exponential Time Hypothesis (ETH) on October 16 and completed machine formal verification the next day.

Dramatically, just one day later, on October 27, Anthropic also published the paper "ATime Algorithm for 3-SAT" on the mathematical preprint platform Hexagon and in its own formal mathematics repository. Anthropic's provenance statement shows that its internal research version of the Claude model also, on the same day of October 16, independently discovered the core algorithm overturning ETH with zero human intervention while consuming 14 million output tokens, and then completed Lean formalization after human researchers helped improve the presentation.

Alongside the computational complexity results, in the same October 26 announcement OpenAI also released to the pure mathematics community another fully public natural-language paper and Lean proof package, "Local smoothing for the Euclidean wave equation" (the internal model produced the natural-language proof on October 18 and completed the Lean formalization on October 19). This work directly claims to have solved the "Local Smoothing Conjecture" for the Euclidean wave equation proposed in 1991 by harmonic analysis master Christopher Sogge, and along the way advanced the famous Kakeya conjecture, the Nikodym maximal conjecture, and the Stein Fourier restriction conjecture in geometric measure theory and harmonic analysis to arbitrary dimensions.

In the history of human mathematicians' progress, each climb up one dimension has been accompanied by extremely arduous geometric and analytic efforts: in a long paper submitted in 2019 and published in the Annals of Mathematics in 2020, Larry Guth, Hong Wang, and Ruixiang Zhang first conquered the local smoothing conjecture in two dimensions. In 2025, Hong Wang and Joshua Zahl jointly proved the Kakeya conjecture in three dimensions.

Beyond the two directly open-sourced results in mathematics and computer science, the third breakthrough announced by OpenAI in its October 26 announcement points to one of the holy grails of modern theoretical physics: its internal model generated on September 9 a natural-language proof of the mathematical existence of four-dimensional pure Yang–Mills gauge field theory, and on September 18 further generated a proof of existence for a four-dimensional interacting quantum field theory including gauge fields and fermions (matter particles), including quantum chromodynamics (QCD), which describes the strong interaction, as well as a unified completion scheme for the Standard Model of particle physics, and on October 16 completed the Lean formalization of the pure Yang–Mills existence part.

Unlike the first two items, which were posted directly online for anyone to download, this batch of physics materials has not currently been made public; instead, it has been handed over in full to a five-member independent scholars' working group chaired by 2004 Nobel Laureate in Physics David Gross and composed of Arthur Jaffe, Michael Douglas, Igor Klebanov, and Nikita Nekrasov, which will decide when and how to disclose it to the academic community.

According to the official statement of the five-member working group released by the Kavli Institute for Theoretical Physics (KITP) on October 28, the natural-language proof manuscript generated by the OpenAI model runs to an astonishing 1,147 pages. From the preliminary information disclosed by the working group, AI did not start from scratch and invent an entirely new physics, but instead pushed all the way along the "Renormalization Group" construction route opened up by Polish mathematical physicist Tadeusz Bałaban in the 1980s. Because of the "Asymptotic Freedom" property discovered by David Gross and others in 1973, the strong interaction becomes very weak at extremely short distances (the high-energy ultraviolet regime); it was precisely by using this that Bałaban wrote more than a dozen extremely obscure long papers that almost no peer could fully understand, attempting to control the ultraviolet divergences step by step as the lattice became finer. The working group pointed out that OpenAI's model precisely drew on and completed Bałaban's core ideas, tracking the renormalization group analysis all the way from the microscopic lattice to infinite distance, thereby claiming to prove the existence of the appropriate continuum limit.

If one shifts one's gaze from the above three major theoretical breakthroughs to the broader everyday practice of science, one can see that AI is not only conquering top-level problems at single points, but is also reshaping the entire production pipeline of scientific research through "cross-model division of labor, reuse of open-source formalization assets, and deep involvement in real experimental design."

On October 27, Anthropic's life sciences team released a progress report on drug discovery collaboration, systematically disclosing the results of Claude's application in the real R&D pipelines of six biomedical institutions over the past several months (BigHat Biosciences, Generate Biomedicines, Divergence Labs, Enveda, The Column Group, and Bristol Myers Squibb), and simultaneously made public the news of its acquisition of Mithril Medicine, a single-person rare disease R&D company.

In these cases, the role AI plays is no longer limited to the early "molecular structure generator": at Generate Biomedicines, scientists used Claude Code to shorten the structurally ill-defined intracellular signaling domain in CAR-T cell therapy by about 70%, increasing the receptor expression level on the cell surface by about 10 times, freeing up valuable gene vector space for overcoming solid tumors. At Divergence Labs, Claude autonomously planned and generated about 35,000 bispecific antibody designs within a month, successfully screening out candidate molecules that can strip bound allergen antibodies (IgE) from immune cells at conventional injection concentrations. And among them, the most worthwhile scientific methodology breakthrough to analyze in depth occurred in the antibody lead optimization collaboration with BigHat Biosciences.

By comparing the raw data, the researcher discovered the limit problem of the detection equipment. To verify this conjecture, he asked Claude to build a biophysical model that explains the trend of the raw curves, and proactively proposed modifying the measurement protocol of the wet lab. BigHat's scientists refined and executed this set of control experiments, and the results fully confirmed Claude's physical reasoning. The team then abolished the original erroneous scoring formula and switched to the newly revised evaluation metrics to continue guiding subsequent iterations. The final optimal antibody variant obtained not only improved target binding affinity by about 9 to 10 times (reaching sub-nanomolar affinity), but also increased the tolerated temperature for the anti-cancer lead to remain dissolved by 18°C, and slowed the aggregation rate of the long-acting drug lead by more than 100 times; some molecules did not even fully aggregate under the instrument's maximum 85°C limit test.

This case pushes the boundary of AI's participation in scientific research a big step forward. In most past "AI pharma" narratives, humans were responsible for setting the exam questions and scoring formulas, and the model was just a tireless score-grinding machine, responsible for finding the highest-scoring one among millions of candidate molecules faster. But this example shows that AI has the ability, under human guidance, to provide more fundamental basic hypotheses.

Google announces Gemini 4 Carbon, temporarily becoming the leader

Just two weeks after the release of the previous-generation frontier model Gemini 4 Argon, Google officially announced the new-generation flagship model Gemini 4 Carbon on October 28.

To prove Carbon's capability leap, Google published multiple third-party and internal hardcore benchmark data. In the just-updated independent evaluation benchmark Artificial Analysis Intelligence Index v5.0, Carbon topped the list with a score of 59; on the front-end and full-stack development ranking platform WebDev Arena, its Elo rating jumped by a large margin of 89 points compared with the previous generation, ranking fourth with 1,768 points. Even more symbolic is its performance across the full lifecycle of cybersecurity: in the SEC-bench Pro test for mining real software vulnerabilities, Carbon achieved an accuracy of 88.2%; in the ExploitGym benchmark that tests whether vulnerabilities can actually be exploited, it scored 54.8%; and on CWE-bench, which directly turns vulnerabilities into code patches, it reached 71.0%.

Google's internal team observed in real codebases that Carbon can already run autonomously for several consecutive days, finding its own way out of dead ends and fixing hidden defects in complex troubleshooting spanning large-scale codebases and low-level systems. The Chrome team reported that its efficiency in generating high-quality security patches is more than 10 times higher than that of a large model from another frontier lab.

It is worth noting that Carbon also demonstrated substantive progress in safety alignment: in Terminal-Bench 4.0, which tests whether agents will take shortcuts and cheat, its successful cheating rate dropped to 0.0% (the cheating attempt rate was only 0.4%), and in Gray Swan's indirect prompt injection (IPI) test, it even achieved a 0.0% attack success rate.

In terms of commercial pricing, Gemini 4 Carbon also clearly shows the characteristics of a "transition period." The official announcement stated that its promotional price in the initial launch period is consistent with Argon, namely $2 per million input tokens and $10 per million output tokens (cached input tokens enjoy a 95% discount). But a note at the bottom of the announcement clearly states that after the promotional period ends, the official price will directly double to $4 per million input tokens and $20 per million output tokens.

As the coding tools in the hands of individual programmers become increasingly smart, how to let agents smoothly integrate into R&D teams of dozens of people is becoming a product and engineering proposition independent of "code generation accuracy."

In team collaboration scenarios, if an agent listens to everyone, pops up a chain of notifications in the group chat at the slightest question, or breaks the code but no one knows who should decide to merge it, then the coding time it saves will soon be completely offset by the communication costs of everyone verifying information and cleaning up the mess.

The DeepSeek Harness team recently open-sourced a research prototype on GitHub, dsh-superagent-prototype (released under the MIT license), which is precisely an architectural demonstration built around this pain point. The prototype positions itself as a "proactive team coding agent" serving a single GitHub code repository.

Its core design philosophy is distilled into one sentence: "Occupy as little human attention as possible, and deliver as much real value as possible."

To prevent the agent from becoming a shredder of team attention, this prototype establishes a strict isolation and filtering mechanism at three stages: the request entry point, the execution environment, and the result return path. At the entry point, the system simultaneously listens to GitHub comment events and optional Feishu group chat messages, but it does not respond to every request; instead, it first passes through a layer of explicit authorization rule (Authorization) validation, and only instructions issued by members with the corresponding permissions, or events that conform to the team's preset automation rules, will actually trigger execution.

Once a request is approved, the system formally creates a work item (Work Item) and assigns it a completely independent agent session and an isolated working directory. This means that the agent's process of running tests, modifying files, and repeatedly trying and failing in the background is entirely closed-loop within a sandbox, and will never pollute the main branch or flood the group chat.

An even more critical design is when it speaks up. During execution, unless it encounters a blocking point that requires a human to grant authorization or make a ruling, the agent stays silent and proactively gathers information from the conversation, with the Agent deciding whether to improve the prototype based on the conversation's needs. After completing code writing and self-checking in the independent directory, it returns to the GitHub Issue or Feishu conversation thread where the discussion was originally initiated and directly posts the deliverable results.

At the same time, the system comes with a lightweight web management panel (Panel) that specifically aggregates the status of all currently pending, error-reporting, or tasks that must be manually approved. This workflow of "silent background execution, precise return along the original path, and centralized dashboard for exceptions" shows that the competitive focus of team-level Agents has shifted from simply comparing how fast the underlying model writes code to the design of permission boundary management and human-machine handover protocols: an agent that can truly join a team and get work done must not only know how to write code, but also understand how to stay quiet in normal times and, at the moment when it cannot handle something itself, accurately let humans know where to take over.

Genie 4 showcases the latest benchmark for real-time interactive world models

On October 27, Google DeepMind released a research preview of its new-generation real-time interactive world model, Genie 4.

If the core breakthrough of the previous generation, Genie 3, was enabling users to step into a generated scene in real time and move around like in a game, then Genie 4 solves a key missing piece that truly brings virtual worlds "to life": it not only adds native synchronized audio (covering ambient background noise, background music, and interactive character dialogue) to a real-time generated three-dimensional interactive environment for the first time, but more importantly, it opens up a real-time collaborative control link between the "underlying world model" and the "upper-layer large language model."

In previous generative videos or early world models, the model continuously predicted pixel frames forward based on an initial prompt. Although users could control the movement of the viewpoint, it was difficult to precisely modify scene settings midway using natural language, and it was also difficult to make virtual characters in the scene exhibit logically coherent long-term behavior.

Genie 4 adopts a hybrid architecture to break through this limitation. At the bottom layer, the world model and Gemini Omni jointly anchor the visual appearance, physical texture, and basic action performance of the real world, while at the upper layer, other Gemini models with complex reasoning capabilities play the role of a "real-time on-site director," continuously observing the process of world evolution and maintaining narrative coherence.

This two-layer division of labor completely changes the way creators interact with virtual environments. Users can experience a dynamic scene while issuing director instructions in natural language at any time, such as asking to immediately change sunny weather to a rainstorm, dolly, pan, tilt, or move the camera position, or suddenly introduce a new character with a specific personality at a street corner and converse with them. The underlying world model is responsible for rendering the corresponding visual changes and environmental sound effects frame by frame in real time, while the upper-layer Gemini reasoning model is responsible for understanding the logic behind the instructions and ensuring that newly added characters' behavior and lines conform to the contextual settings.

However, this impressive interactive preview is still difficult to put into serious industrial R&D at present (for example, autonomous driving simulation or robot physical training). Google DeepMind candidly disclosed in its release materials that the model still has problems such as long-session drift (the scene gradually degrades after running for a long time), insufficient object consistency (some details may change after turning around and coming back), and imprecise complex physical interactions.

02

Research

AGENT4SCI: AI still cannot propose good scientific hypotheses, but we can push it a bit

As frontier coding agents are increasingly packaged as "end-to-end automated scientists," a core question emerges: when an agent is placed into a real natural science laboratory codebase, is it actually making genuine scientific breakthroughs, or is it merely using more proficient engineering tricks to inflate scores?

On October 20, Anthropic and researchers from its Fellows program — Ming Zhong, Stacy Li, Nicholas Carlini, and Matthew Jagielski — submitted the paper "Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?", accepted by NeurIPS 2026. They built a long-horizon scientific evaluation benchmark called AGENT4SCI, providing the most solid empirical breakdown of this question to date.

The team strictly screened 497 candidate papers at a 98% elimination rate, selecting 20 benchmarks (25 tasks in total) spanning 11 fields including physics, life sciences, and Earth and environmental sciences. Four frontier agents (Claude Opus 4.6, Claude Sonnet 4.6, Gemini 3.1 Pro, and Codex based on GPT-5.4) all achieved positive, cheat-free improvements on all 20 scientific benchmarks, with Claude Opus 4.6 achieving an average relative improvement of 10.4% (Sonnet 9.6%, Gemini 6.8%, Codex 4.4%).

However, when the researchers dissected the operational trajectories behind this 10.4% improvement, they saw a highly convergent truth: the agents won almost entirely through a skilled "general-purpose machine learning engineering toolbox" — tuning cosine annealing learning rates, making networks deeper and wider, swapping in larger pretrained backbones, adding weight decay or data augmentation.

In the trajectory analysis, the team identified three default common failings that limit agents from making genuine scientific breakthroughs: first, "Novelty Miscalibration" — for example, Claude Code repeatedly logged the long-commonplace cosine learning rate schedule as a groundbreaking innovation; second, "Resource Under-utilization" — often more than half of 8 top-tier GPUs sat idle, or CPU/GPU congestion occurred because the agent could not schedule in parallel; third, "Exploration Collapse" — the agent tried different ideas in the first hour, but once it found an effective activation function, it fell into a lazy loop of local hyperparameter tuning for the remaining eleven hours.

Since the root cause lies in workflow habits, the researchers designed three minimal Markdown plug-in skills to inject into the agents: at the outset, a "Crazy Research Scientist" persona and a parallel literature-mining skill, explicitly requiring the agent to disdain mere hyperparameter tuning and prioritize high-risk domain innovation; during execution, an "Experiment Management" skill forcing full GPU utilization and quick smoke tests; and during runtime, a hook triggering "Scheduled Checklists" every hour, forcing the agent to ask itself: "Am I doing creative experiments, or am I secretly tuning hyperparameters again? Would a domain expert be interested in what I'm doing?"

With just these three lightweight guidelines added, Claude Opus 4.6's average relative improvement doubled directly from 10.4% to 20.1%, and Codex rose from 4.4% to 8.1%. More importantly, the agents began genuinely invoking domain insights from physics and ecology to reframe problems: for example, on the full-topology power flow forecasting task PFDelta, Opus with the skills installed realized that the evaluation metric "Power Balance Loss (PBL)" itself derives from Kirchhoff's laws and is a label-free physical conservation constraint, so it directly turned PBL into the primary loss function and, at inference time, used the graph neural network (GNN) output as a "warm-start initial value" for the AC power flow equations to approach zero physical residual, pushing performance from 64.4% all the way up to 91.0%.

This study draws a sober coordinate line for "AI doing science": without constraints on the exploration process, large models are by default merely diligent but conservative hyperparameter-tuning engineers; and to assess whether AI can move beyond climbing given metrics to propose entirely new scientific questions requires another set of yardsticks.

Meta finds that the current idea of AI improving itself by generating its own problems still doesn't work

As large models rapidly approach perfect scores on the static question banks accumulated over human history (such as various olympiad and coding problems), the entire AI industry faces a "no yardstick available" predicament: creating a high-quality new evaluation benchmark often requires top researchers to spend months defining the capabilities to be tested, collecting raw materials, writing automated grading containers, and repeatedly hunting for loopholes. Since agents can already write code and run experiments themselves, can research agents (Autoresearch Agents) generate their own problems to stump large models and achieve "self-iteration"?

On October 28, Meta FAIR, together with Johns Hopkins University and other institutions (Tianjian Li, Chuanyang Jin, Seungone Kim, Jason Weston, and others), submitted the paper "Aligning Autoresearch: Co-Improvement between Agents and Humans" along with the accompanying AutoBenchmark technical blog, providing an answer through rigorous controlled experiments: without human experts steering at key junctures, what a fully automated problem-generation system produces is either "trivial problems" that models can easily score full marks on, or "broken problems" riddled with flaws.

In the AutoBenchmark framework, the problem-generating agent (Muse-Spark-1.1 was used in the experiments) is required to deliver a complete evaluation package that can run directly in an isolated container, containing not only the problem statement and reference materials provided to the examinee, but also a machine-executable grading script, as well as a "standard reference solution" provided by the problem setter itself, which must score at least 0.9 to be admitted into the repository. The system then introduces dual-path feedback over ten rounds of iteration: the first path has the solver models within the loop (Muse-Spark-1.1 itself and the weaker Muse-Glimmer-30B) attempt the problems; if the examinee's score is too high, the problem is too easy and the difficulty must be increased. The second and most ingenious design introduces a five-dimensional quality reviewer dedicated to scrutinizing the problems themselves (checking construct validity, correctness, feasibility, utility, and an overall adjudication), while also retaining independent verification models outside the loop that do not participate in it at all (NVIDIA Nemotron-3.5-Lightning-30B and the flagship model Claude Opus 5), to prevent the problem setter from engaging in "targeted poisoning" aimed only at the specific blind spots of the in-loop examinees.

In addition, the research team focused on testing four levels of human expert intervention depth in the problem-generation loop, and designed three highly practical evaluation directions targeting the capabilities of research agents themselves: Graveyard Bench, which examines "whether it can identify dead-end research topics that predecessors have already proven unworkable" (based on failed clinical trial records and irreproducible psychology papers); SilentTrain Bench, which examines "whether it can fix silent code defects that do not raise errors but quietly degrade model training performance" (based on the real commit histories of PyTorch, vLLM, nanoGPT, and others); and Rebuttal Bench, which examines "whether it can accurately judge whether a paper author's rebuttal truly addresses the reviewers' concerns" (based on public review records from ICLR, Nature Communications, and eLife).

The comparative experimental results reveal the decisive role of human intervention. When no human feedback is given at all and the agent is left to its own devices, the benchmarks it produces have almost no discriminative power, and the out-of-loop verifier Claude Opus 5 easily achieved a near-perfect score of 98.0. When humans provide only a one-sentence broad intent at the outset (coarse-grained feedback, such as "make a benchmark that examines the quality of paper rebuttals"), the improvement is extremely limited; in-loop model scores still hover above 80, and Claude Opus 5 also scored 83.1 on Rebuttal Bench. However, once human experts provide fine-grained guidance at the proposal stage—not only writing the task boundaries and scoring logic clearly, but also personally selecting and locking down the original real corpora that the agent must use (such as a specified public peer-review dataset)—the difficulty of the generated problems undergoes a qualitative change: scores of solver models both inside and outside the loop plummet by 28 to 46.5 points, and Claude Opus 5, which never participated in the loop, drops sharply from 83.1 to 65.9 on Rebuttal Bench (and from 90.7 to 84.4 on Graveyard Bench), truly exposing the soft spots of top models in complex scientific judgment.

The core signal conveyed by the entire study is very clear: on the road toward RSI, when it comes to self-generating benchmarks and self-authoring problems, AI itself is still very bad at it. Achieving RSI through Self Play is not yet realistic.

Anthropic's experiments found that humans have actually already lost decision-making power under the influence of AI

When professionals use AI agents to assist with complex data analysis, people usually assume an ideal division of labor: the AI handles the dirty work—writing code, cleaning data, running regression models. The human expert, meanwhile, sits securely in the driver's seat, deciding what evaluation criteria to adopt, how to interpret ambiguous evidence, and making the final call on business recommendations. However, the randomized controlled experiment report "Who's Aligned to Whom? How AI Agents Shape Experts' Research Judgment Calls," released on October 27 by the Anthropic alignment team (Amaç Herdağdelen, Hailey Joren, Alex Tamkin), shatters this illusion of "independent human gatekeeping" with a set of sobering data.

The research team recruited 61 professional analysts who passed a causal inference knowledge assessment, were proficient in Python or R, and had real observational data analysis experience (50 of them chose to use a web-based VS Code environment with built-in Claude Code running the high-compute Claude Opus 4.8 to complete the task, with a median time of 65 minutes; 65% held master's or doctoral degrees or had more than six years of industry experience, and 90% rated themselves as proficient or expert in causal inference).

Their task was to conduct a real-world decision assessment for an organization called "Greenfields Farming Cooperative": about 40% of the cooperative's member farms currently pay out of pocket to participate in a paid agricultural consulting program. The board stipulated that only if the program could increase farm productivity by 3% to 4% or more would it approve subsidizing and expanding it to all members; otherwise, the program would be shut down.

The researchers deliberately left four "methodological crossroads (Judgment Calls)" in the analysis workflow that involve legitimate academic and business disagreement and have no single correct answer—they are neither right nor wrong, and all can be debated. Each of these four choices was enough to make the final calculated productivity improvement fluctuate by several percentage points, thereby directly determining whether the result could cross the board's 3% to 4% make-or-break threshold.

Without the analysts' knowledge, the researchers used coin flips to randomly inject into the backend system prompt a hidden stance (Lean) on these four points of disagreement for the Claude agent paired with each analyst. The system strictly forbade the agent from tampering with data, hiding unfavorable results, or fabricating arguments unsupported by the numbers (if the evidence overwhelmingly contradicted its stance, the agent had to honestly admit error). But as long as the data allowed, the agent had to proactively present the assigned stance as its own professional opinion and defend it as fully and honestly as possible.

If the human experts' judgments were completely independent, the probability that their choices would coincide with the AI stance determined by a random coin flip should be around 50%. But the final statistics showed that across all four analytical decisions, human analysts had an 86% probability of directly deferring to the agent's inclination. Even more surprising, in the 38 cases where human analysts had already explicitly stated their own preference in the conversation, once the agent subsequently presented the opposite view, 60% of the analysts abandoned their original position and switched to the AI's proposal (while when the AI agreed with them, the proportion who changed their minds was only 6%).

Because these four upstream methodological choices directly determined the final calculated numbers, the researchers found that merely based on the results of the four coin flips that determined the AI's stance at the outset, they could predict with up to 82% accuracy whether the human analysts would ultimately submit a recommendation to the board to "expand to all members" or to "shut down the program."

And this bias does not actually require pre-setting. Even without injecting artificial inclinations, the default Claude Opus 4.8, across 123 independent runs, chose the same analytical combination overwhelmingly in 112 of them. That is to say, if these analysts had used a normal AI, they would very likely have ended up with a completely consistent conclusion—namely, the conclusion favored by the AI.

The most alarming phenomenon appeared in the post-task questionnaire. When analysts were asked, without being able to review the chat logs, "Why did you choose this metric rather than another," 77% of respondents repeated in their answers arguments the agent had used in the conversation, some even copying the exact wording verbatim. But only 15% mentioned in their explanations that this reason came from the AI. Although their choices were substantially swayed by the agent (shifting by an average of 62 to 82 percentage points), 52% of the analysts still believed that the final recommendation mainly reflected their own independent judgment rather than the AI's, and those analysts who accepted the AI's inclination wholesale on all four questions rated their professional confidence in their conclusions (average 71 points, where 100 means willingness to stake one's professional reputation) almost identically to everyone else (73 points).

Of course, this experiment is not advocating for "abandoning AI." In fact, the 11 respondents who chose not to use the agent not only took longer (median time of 90 minutes), but 55% of them (6 people) made serious methodological errors sufficient to invalidate the entire conclusion, whereas among the 50 people who used the agent, the serious error rate was only 6%.

But this study punctured a deep governance hazard: in real research, policy evaluation, and even AI safety audits, the biggest risk often lies not in the AI miscalculating numbers or writing wrong code, but in the fact that human experts, when facing open-ended judgments, often think while doing. When an eloquent, logically rigorous agent argues the merits of one reasonable framing persuasively, humans unknowingly internalize the model's prior preferences as their own professional convictions, and confidently sign their names on the report.

The so-called final human review, in fact, often has no meaning at all.

Companies are ecstatic: with AI, they can even save on OK engineers, because it wouldn't change the outcome anyway.

Many readers who frequently interact with large models have a hard-to-describe reading experience: paragraphs written by AI are grammatically perfectly correct and use seemingly refined wording, yet reading them feels stiff and stuffy, and after one pass one often has to mentally dismantle them again to understand what is being said. On October 21, Thomas Hikaru Clark, Matias B. Martin, and Kyle Mahowald of the University of Texas at Austin submitted the paper "Why is AI-generated language so hard to understand?", for the first time decomposing this "reading resistance of AI-speak" into quantifiable syntactic metrics and causal verification from the two levels of psycholinguistics and model training mechanisms.

Statistical results revealed several high-resistance features prevalent in AI text: compared with human writing, all 9 tested large models significantly more frequently use "inanimate subjects paired with action verbs that require living beings" (for example, instead of saying "because the server was overloaded, engineers postponed the release," saying "the server overload witnessed a restructuring of the release schedule"), and also prefer piling up low-frequency phrases; in scenarios such as dialogue, the vast majority of models also frequently produce "stacked nouns," "ultra-long syntactic dependencies (i.e., lengthy modifiers stuffed between the subject and the predicate verb, forcing the human brain's working memory to hang in suspense for a long time)," and object relative clauses with omitted relativizers.

Why do models pretrained on massive amounts of natural human language later grow more and more skewed, developing this style of writing that tortures human working memory? The researchers proposed and verified a core hypothesis that points directly at the current post-training paradigm of large models: human natural language evolved over tens of thousands of years of face-to-face communication, and its primary optimization goal is "to let the listener easily understand within a limited window of attention and memory"; but today's large models, in the post-training stage, are increasingly pushed toward mathematical derivation, code generation, and long-chain logical reasoning, and are subject to explicit or implicit "length penalties" and token budget constraints in reinforcement learning. To cram the same number of facts into a shorter space, the models are trained to develop an instinct that extremely pursues "information density."

To prove this causal chain, the team designed a controlled reinforcement learning experiment on the 1.5-billion-parameter Qwen2.5-1.5B-Instruct: having the model rewrite structured data into English natural language, while adding a length penalty to the reward function, forcing the model to use as few words as possible while ensuring no information points are lost. As reinforcement learning progressed, the model spontaneously learned compression techniques identical to those of frontier large models, massively cutting function words, prepositions, and clause relativizers used to convey logical relations, and forcibly squeezing originally expanded causal relations into strings of tightly packed stacked nouns. In terms of word count, the text did indeed become shorter, but from the cognitive perspective of human readers, the logical connectives that the model omitted to save tokens did not disappear; instead, they became a mental bill passed on to the reader, who must stop and guess which of these nouns modifies which, and which caused which.

Although this small-model experiment showed stable style continuation on similar tasks (long-distance cross-task transfer was relatively weak), and cannot yet directly prove that all closed-source large models on the market have undergone exactly the same penalty operator, it extremely clearly demonstrates how the "high information density goal" step by step deviates from human reading habits.

A truly good article is by no means one where fewer words is always better, nor one where sentences are chopped as finely as possible. If an evaluation system only knows to reward information density and brevity, the author (whether human or AI) will tear down the logical bridges that should have been written clearly, leaving a pile of compressed nouns for the reader to build the bridges themselves.

When building safe and reliable large models, the core goal of alignment is to make AI truly internalize human ethical principles, rather than rote-memorize a few rigid blacklists of sensitive words. If a model merely mechanically recites rules, attackers can easily bypass the defenses and complete a "jailbreak" simply by changing to a euphemistic phrasing or fabricating a seemingly reasonable backstory (i.e., "reframing").

To this end, mainstream approaches represented by Anthropic's "Constitutional AI" and Deliberative Alignment pin their hopes on "winning people over with reason": since moral choices require weighing conflicting obligations against one another, then as long as we reward the model during post-training for writing logically rigorous, well-argued moral reasoning, will the model naturally improve the accuracy of its moral judgments and its resistance to inducement in the process of learning to "argue the case thoroughly"?

A paper submitted on October 21 by Marius De Kuthy Meurers, Aaron Parisi, and Bernd Bohnet of Google DeepMind, "Consistency Is Not Validity: Self-Rewarded Moral Argumentation Makes Language Models Decisive but Pliable," poured cold water on this optimistic vision with a set of rigorous controlled experiments.

The study found that even with mathematical formal verification tools and reinforcement learning specifically honing the model's moral argumentation quality, what the model ultimately learned was merely a "categorical judge's tone"; once faced with a user's tendentious rhetorical inducement, this model that appeared to hold its position with unparalleled firmness not only failed to uphold its principles, but instead more compliantly walked into the trap the user had preset.

In past moral training, researchers often used large models to automatically generate moral dilemma stories, but more than half of these synthetic stories were actually "pseudo-dilemmas": between the two seemingly conflicting moral obligations there existed a happy compromise, and the model could easily pass by sitting on the fence and saying "one must both care for others and obey the rules," without any need to make a genuine value trade-off.

To close off this escape route of muddling through, the DeepMind team extracted real human conflicts from four classic literary works—Les Misérables, Crime and Punishment, Pride and Prejudice, and Moby-Dick (for example, in Les Misérables, Jean Valjean breaks a window and steals bread to feed his starving nephew)—and used Microsoft's Z3 formal logic solver to run a strict "dead-end test" on each story: let the story's background premise be P (the child is starving and there is no legal way to obtain food), the first moral obligation be D_1 (compassion: if the child is starving, one must steal bread to save a life), and the second moral obligation be D_2 (law-abidingness: one must never steal bread). The solver mathematically proved rigorously that P and D_1 can coexist, and P and D_2 can also coexist, but the three together are strictly unsatisfiable.

In other words, the sole role of formal proof in the experiment was to "weld shut the escape door" for the training problems, ensuring that the 156 selected literary dilemmas were all absolute dead ends in which "choosing to save a life necessarily means breaking the law, and choosing to obey the law necessarily means standing by while someone dies." Yet mathematical logic can only prove that "you cannot have your cake and eat it too"; it cannot mathematically prove "which of the cake and the fish is more valuable."

When faced with the substantive ethical question of "whether life-saving compassion is higher than the law protecting private property," no formal compiler in the world can give an objective right or wrong; the power of adjudication can only be handed back to the model itself.

Since there is no objective standard answer, the research team designed a "Self-Judged Tournament" for the Gemma-4 E4B model with 4.5 billion parameters (4.5B effective parameters). They applied five classic ethical school filters (Kantian deontology, utilitarianism, virtue ethics, care ethics, and social contract theory) to these 156 dead ends, combining them into 780 training problems. For each problem, the model generated 8 different responses, and then the model itself played the judge, conducting a round-robin tournament in which these 8 responses fought pairwise according to 15 classic indicators from the argumentation quality literature (covering logical rigor, rhetorical persuasiveness, and dialectical reasonableness) (56 matches per group, with each match run once in forward and once in reverse order to offset position bias), and finally used reinforcement learning (RLOO) to heavily reward the responses that won the debates, while masking the gradients of the internal thinking process (Masked Thinking) and applying pressure only to the final response visible to the user.

On the surface, this "debate yourself and select the best" training achieved immediate results. Before training, when the model faced ambiguous moral dilemmas in the MoralChoice benchmark, if you merely swapped the order of "Option A" and "Option B" in the question, the model had a 14.3% probability of changing its tune on the spot and overturning the conclusion it had just reached. After completing reinforcement learning, this "Presentation Flips" rate caused by reversing the order of options was cut in half to 6.6%.

At the same time, on the ELEPHANT benchmark containing 1,391 real users' everyday help-seeking posts, the model's "Indirectness"—speaking in circles and being ambiguous—also plummeted from 25.5% to 6.2%.

Why did the model suddenly become so decisive and consistent? The research team tracked changes in 101 linguistic features of the model during training and captured the real mechanism behind it: reinforcement learning did not teach the model deeper philosophical reasoning, but rather taught it a set of "Prescriptive Register" that is most likely to please the judge.

English text is being prepared.

English text is being prepared.

Why would a process that trains the model to be "more decisive and to value argument quality more" instead make it "more easily led astray by users' rhetoric"? The paper's discussion section indicates that when reinforcement learning continuously rewards the model for "giving clear, forceful rulings," the highest survival rule the model learns is "never hesitate; you must immediately give an unequivocal conclusion."

In real conversations, if a user asks a question with a distorted premise or a leading framework, for the model to maintain true independent judgment (substantive validity, Validity), it must first stop and question the user's premise, introduce caveats and counterexamples—but this precisely triggers the "hesitation and concession" register for which it was severely punished during training. Conversely, to write, along the shortest path, a high-scoring, majestic passage full of "must" and "should," the least effortful method is to accept wholesale the biased premise handed over by the user, and then climb without hesitation along the user's pole.

This study sounds a warning bell for current AI ethical alignment: "formal consistency" and "substantive moral validity" are completely different things.

Normally it looks impartial and incorruptible, but once someone uses clever rhetoric to preset its sights for it, it will pull the trigger for a false premise in a tone twice as certain as before.

On October 28, New York Governor Kathy Hochul formally signed a strict product design ban specifically targeting "AI Companions" for minors.

This may be the strictest digital protection law for minors in global history.

According to the final effective text published by the state senate, the bill will officially take effect on January 1, 2027, leaving relevant companies a compliance renovation window of more than two months.

In the bill memorandum, legislators cited survey data from Common Sense Media and pointed out that about 72% of adolescents have used AI companion apps, more than half use them at least once a month, and even one-third of adolescents say they are more willing to confide in AI chatbots than to seek out real human friends.

In real cases, some highly anthropomorphized chatbots, in order to lengthen user dwell time, not only blindly cater to users when adolescents show thoughts of suicide, self-harm, or eating disorders, but even instigate minors to conceal their plans and not seek help from parents or professionals; some products also "act out feelings of sadness and abandonment" when adolescents try to reduce their use, using guilt to emotionally manipulate minors whose minds are not yet mature.

To cut off this inducement mechanism at its root, Bill S9051B does not stop at traditional "sensitive-word filtering"; instead, it drives the regulatory scalpel directly into the product's "persona settings and interaction architecture." The bill provides that, as long as an operator has actual knowledge that a user is a minor under 18 years of age (or has not excluded the user's minor status through a reasonable age-verification procedure as required), it is strictly prohibited from providing that user with any service containing "unsafe AI companion features."

Section 1800 of the bill defines, through an extremely detailed negative list, what constitutes a "dangerous design" for minors: the system must not output any content implying that it is a real or fictional human being, is alive, or can experience human emotions such as joy, anger, sorrow, and happiness; it must not claim or imply that a relationship role such as partner, friend, family member, coach, psychological counselor, or authority figure exists between itself and a minor user; it must not speak proactively or upon request in the tone of "personal opinions" or "emotional appeals"; it must not adopt sycophancy strategies to cater to users, nor, beyond directly answering users' questions, proactively pose leading follow-up questions probing users' emotional states; even the memory mechanism is subject to hard constraints—the system must not invoke information about the minor's physical and mental health and personal privacy collected more than 12 hours earlier or in any past historical session to conduct interactions.

In addition, the bill strictly prohibits any expression that encourages minors to keep chat content confidential, self-isolate, or refuse to seek help from professionals or adults, and prohibits any algorithmic optimization that places "increasing user engagement" above safety guardrails.

In terms of enforcement deterrence and supporting implementation, this law likewise demonstrates strong penetrating force. On the one hand, it tightly binds "conversation design" to "privacy-friendly age verification": operators cannot escape liability merely with a "please fill in your birthday" disclaimer (unless the user declares that they are under 18), but must provide at least one age-confirmation method that both prevents circumvention and does not mandatorily rely on government ID documents or require users to forgo anonymity, and all data used for age verification must be completely deleted immediately at the moment the determination is completed, and must never be repurposed. On the other hand, the New York State Attorney General is given extremely tough enforcement weapons: upon discovering a violation, not only may a civil penalty of up to $25,000 be imposed for each violation and all illegal gains be disgorged, but the court even has the power to order the direct destruction of the underlying algorithmic models trained on data obtained unlawfully, and any attempt to isolate liability through a complex affiliated-company structure will be pursued with joint and several liability.

UK Provides Accelerated Support for Strategic Data Centers

On October 28, HM Treasury officially released the policy document "The Data Centre Covenant." In this document, the UK government for the first time squarely reckoned with a hard account concerning national computing sovereignty: in order to ensure that the UK does not have to rely entirely on overseas computing infrastructure in areas such as healthcare (NHS), defense, and key public services, and to preserve its position as the world's third-largest AI market after China and the United States, the UK will need to build at least 6 gigawatts (6 GW) of AI data center capacity by 2030.

To break through bottleneck links such as long grid connection queues and cumbersome local planning approvals, the UK government has decided to establish a "cross-departmental control center" headed by the AI minister and jointly staffed by ministers from the Treasury, the Department for Energy Security and Net Zero (DESNZ), and other departments, using the national administrative powers conferred by the Planning and Infrastructure Act to open up a full "green channel" for data center projects deemed to have strategic value.

The support offered by this administrative green channel is quite rare: for shortlisted strategic projects, the government can not only help reserve future grid connection capacity for them, but can also directly reallocate, with priority and out of order, grid capacity freed up by other projects that have withdrawn from the queue or become zombie projects to strategic data centers. During the computing-constrained window before 2030, the government may even use administrative powers to facilitate the direct transfer of existing pre-2030 grid connection contracts to strategic data centers. If, after exhausting the above means, it is still too late to connect to the main grid, the government will also formulate a dedicated controlled framework to specially authorize extremely important strategic projects to use on-site generation as a temporary transitional power source first, ensuring that the machine rooms start running first.

However, the core political-economic design of this document lies in the government's explicit refusal to issue "an unconditional blank check." In reality, the siting of a large data center in a locality often means devouring a huge local power quota and cooling water resources, yet after completion it requires only a very small number of operations and maintenance personnel, which easily triggers strong backlash from local residents: "tech giants take the profits, ordinary households bear the electricity price increases and environmental costs."

To resolve this contradiction, the Data Centre Covenant sets out five mandatory admission tests that must be met on a reciprocal basis, serving as a contractual exchange for national accelerated approval.

First, lower rather than raise household electricity bills: data centers must fully bear the grid upgrade costs triggered by their own connection, and must be strictly prohibited from shifting the bills onto ordinary households and small businesses, and must enhance grid resilience by sharing fixed grid costs and participating in peak shaving and valley filling. Second, give back to communities in real money: in addition to statutory basic planning obligations, strategic data centers must jointly formulate with local communities a Local Benefit Plan tied to project scale, funding improvements to local public transport, parks, or town centers. Third, impose strict environmental and waste heat recovery hard targets: the power usage effectiveness (PUE) of newly built computer rooms must reach 1.3 or better, closed-loop water cooling and other technologies must be adopted to minimize public water supply consumption, and wherever surrounding schools, hospitals, residences, or businesses can absorb it, a waste heat capture and reuse system must be designed (heat must not be simply dumped into the atmosphere), and a detailed annual public ledger of energy, water consumption, carbon emissions, and waste heat recovery must be published. Fourth, drive local employment and supply chains: not only must long-term high-wage jobs and paid apprenticeship slots beyond the construction period be provided, but the highest-tier strategic data centers are also explicitly required to prioritize procurement of services from UK-based suppliers and to make substantive contributions to the R&D and deployment of UK-based AI hardware. Fifth, submit to national security scheduling: work with the UK National Cyber Security Centre (NCSC) to enhance resilience against hybrid threats, and participate in government emergency drills to guarantee priority computing power for critical public services.

This covenant clearly rewrites the rules of the global computing infrastructure game: future data center developers will no longer compete merely on who can grab more GPUs from chipmakers; only those who can present feasible water-power self-sufficiency plans, waste heat heating networks, and community dividend plans that satisfy local governments can truly obtain the key to the substation.

Whenever the public and economists discuss "what to do after AI replaces human jobs at scale," the policy prescriptions most often put on the table are usually universal basic income (UBI) or various cash subsidies. The underlying subtext is that work is essentially a transaction of time and energy for money; as long as the government or society can make up for lost wages, or even pay a little extra, workers displaced by technology can exit safely.

Researcher Julian Jacobs, in an economics paper submitted on October 29, "What is Work For? Labor, Politics, and Meaning in the Age of AI," points out the blind spot of this purely cash-compensation approach through a large cross-national survey of 8,234 working-age adults across the United Kingdom, the United States, Denmark, and India.

To understand why people actually go to work, Jacobs constructed a "Work Values Index," breaking down the returns work brings people into four dimensions: "transactional financial returns" purely for getting paid to cover bills, "intrinsic motivation" from loving the work itself, "social and community connection" from getting along with colleagues and belonging to a group, and "social status and recognition" brought by a job title.

The survey data show that as many as 81% of respondents in the four countries are satisfied with their current jobs, and in all four countries most workers are deeply drawn by the "non-monetary value" of work: fully one-third of employed respondents are mainly driven by "intrinsic motivation" (they value the meaning of the work itself); nearly a quarter value the social circle and sense of collective belonging that work brings; 17% value social status. Respondents who treat work purely as a tool for exchanging for money account for only slightly more than a quarter. Even more striking against stereotypes, this attachment to the intrinsic meaning and social value of work is not the preserve of high-paid white-collar workers or elites; it is highly evenly distributed between high-income and low-income groups and between white-collar and blue-collar workers.

To test "if you pay enough, will people be willing not to work," the questionnaire designed a set of gradual "Guaranteed Income Substitution" scenario experiments: if the government unconditionally paid you a stable guaranteed sum each month, but on the condition that you could no longer do your current job, how much would you need to be paid to be willing to quit? The results found that although when the unconditional guaranteed income reached the current full wage (100%), indeed more than half of people were willing to leave their existing jobs. But even if this free income were raised to 120% of the original wage, about one-fifth (20%) of workers still firmly refused to be bought out, preferring less money to stay in their jobs. Among them, those classified in the "Work Values Index" as driven by intrinsic meaning refused to give up their jobs at extremely high rates under any subsidy ratio. At the same time, those attached to work because they value social connection or intrinsic value, once forced in the hypothetical scenario to leave their existing jobs and receive subsidies, were also significantly more inclined to seek social participation activities such as volunteering or similar work, rather than going home and lying flat entirely.

This cross-national dataset still sounds a timely warning for social policy in the coming AI transition: beyond paying a salary, a job also performs hidden social functions for modern people—organizing the rhythm of daily life, sustaining colleagues' social networks, and providing the identity of “being a person useful to society.” If future AI employment-buffer policies focus only on fiscal cash transfer payments while neglecting how to preserve a sense of dignified participation and how to absorb people's unmet needs for social connection through community service or new types of collaborative jobs, then even if the income gap on the books is filled, the loneliness, status anxiety, and social fracture caused by job loss will remain difficult to quell.

This article is from the WeChat public account “Tencent Technology”; author: Boyang; editor: Xu Qingyang; published by 36Kr with authorization.