Yesterday in AI · 18 July 2026

In the Bloomberg report by Maggie Eastland and Nancy Cook that set off the commentary, the plan is still moving through the White House: chief of staff Susie Wiles is reviewing it, President Trump has not seen it, and officials accelerated the work after China's Kimi K3 release raised competition concerns. Bloomberg traces the proposal to industry frustration with improvised federal interventions, among them export controls that led Anthropic to temporarily disable its Fable 5 and Mythos 5 models and changes OpenAI made to its Sol model at the government's request before release. The design follows an earlier Trump order outlining a voluntary review system built with AI companies.

On X on July 14, Google DeepMind chief executive Demis Hassabis published "A Framework for Frontier AI and the Dawning of a New Age", a proposal TechCrunch reports would have frontier labs voluntarily share models with a FINRA-patterned standards body up to 30 days before release, with passage becoming a condition of US deployment once the reviews proved out; open-source representatives and industry technical experts would sit on its board. Further back, Mark Thomas, a Harvard Law School student writing in Lawfare on April 29, proposed recognizing the industry's Frontier Model Forum as a self-regulatory organization with mandatory membership above compute and revenue thresholds. Thomas comes closest to the mechanism design Wagner says she has not seen: he details FINRA's industry-funded $1.3 billion budget, its more than 3,000 member broker-dealers, and its two rulemaking tracks, including a fast track for minor changes. Wagner answered Thomas directly in her thread: "Establish a AI regulatory body that does something, that part all seems fine to me."

A federal evaluator already exists at Commerce. The Center for AI Standards and Innovation at NIST describes itself as industry's primary point of contact for AI testing, holds voluntary evaluation agreements with private developers, and published results on DeepSeek's V4 Pro on May 1; it completed an assessment of Z.ai's GLM-5.2 on July 17, the day Bloomberg's story ran. Thomas floats CAISI as a possible government supervisor for an AI SRO; the Bessent plan would anchor supervision at the SEC instead. Whether CAISI could shoulder that role runs into its $15 million budget and limited mandate.

In Forbes on July 19, economist James Broughel attacked the analogy from the other flank: FINRA operates without notice-and-comment rulemaking, FOIA obligations, or congressional oversight, missed the Madoff fraud despite decades of supervisory experience, and sold its own $647 million auction-rate securities portfolio months before that market froze in 2008. He cites SEC Commissioner Hester Peirce, who has said FINRA "wields governmental powers without the procedural and disclosure requirements" that bind a federal agency, and he would hand the evaluation job to existing voluntary infrastructure: the NIST AI Risk Management Framework, ISO/IEC 42001 certification, and MLCommons benchmarks.

The center in Veronica Irwin's Transformer report began as the US AI Safety Institute, stood up inside NIST at Biden's direction under the 2023 AI executive order; Elizabeth Kelly became its first director in February 2024, and Paul Christiano, the former OpenAI model alignment lead, is its head of AI safety. Congress never wrote it into law. In June 2025, Commerce Secretary Howard Lutnick renamed it the Center for AI Standards and Innovation, declaring that "censorship and regulations have been used under the guise of national security". The rebrand left a national-security testing shop: NIST describes CAISI as "industry's primary point of contact within the U.S. government" for evaluating commercial models, focused on "demonstrable risks, such as cybersecurity, biosecurity, and chemical weapons" and on adversary systems, "including the possibility of backdoors".

CAISI lead Chris Fall's office tested Anthropic's models days before they went offline; the export-control call, Irwin reports, was "actually the outcome of a war of influence" among White House and cabinet principals. The center runs on roughly $15 million: $10 million in FY2026 appropriations plus a $10 million Technology Modernization Fund loan spread across two fiscal years, per an Institute for Progress analysis pricing a minimally "limited" CAISI at $26 million a year, an "equipped" one at $84 million.

The FINRA-style body arrived by a different route. On July 14, Demis Hassabis published a framework proposing a standards body modeled on FINRA: frontier labs "would voluntarily share models with the Standards Body for review up to 30 days before release", industry-funded, with required review to follow once the protocol proves out, per TechCrunch. Three days later, Bloomberg reported that Treasury Secretary Scott Bessent helped develop a proposal for "an independent regulatory agency for AI" that would "report to the Securities and Exchange Commission, similar to the Financial Industry Regulatory Authority". Susie Wiles is reviewing the plan; Trump has not; and it remains unclear, Bloomberg writes, "what kinds of model assessments the administration's proposal would call for". FINRA itself is a private membership organization, funded by member fees and supervised by the SEC.

How that body would relate to CAISI, the reporting does not say. Bloomberg's account never mentions the center, and the two sit on different axes, Treasury and the SEC against Commerce and NIST. Congress is meanwhile moving to entrench CAISI where it is: the AI Security and Innovation Act, marked up in House Science in the last week of June, would codify the center inside NIST at "$20,000,000 for each of fiscal years 2027 through 2032", and the stalled Obernolte-Trahan Great American AI Act draft would authorize $100 million and have CAISI license independent auditors of frontier developers. In Forbes, James Broughel argued that "A single gatekeeper concentrates risk" and that FINRA "wields governmental powers without the procedural and disclosure requirements" that constrain an agency. Charlie Bullock of the Institute for Law and AI expects punctuated movement: "Maybe nothing will happen for a while", until capabilities jump or "some scary thing happens".

On July 17, Dean Ball, OpenAI's head of Strategic Futures since July 6 and a co-author of the White House AI Action Plan, posted six observations on Kimi K3. He called it "a very good model", professed surprise that the Chinese state permits open-sourcing at this level, and argued that "open-weight models are inherently decelerationist" because they deter frontier capital spending. An open-weight-dominant world, his fourth point held, probably ends in "full AI communism", state-provided models as public infrastructure, a future he called a "dystopian hellscape". His fifth was the prediction the week turned on: have every agency "issue soft law that creates FUD" about Chinese models until "every regulated enterprise backs off". No ban, no rule, no vote.

Stella Biderman wrote on X that when trillion-dollar-company employees scream "you just want communism", the heuristic says "you're doing the right thing". SE Gyges read it as an OpenAI official saying, "under his legal name", that the regulations' purpose is hurting open source and favoring incumbents. Tim Sweeney recast it as a taco executive theorizing about rival tacos; Ball took the joke and replied that "ai is not a kind of food". To Nathan Lambert he unpacked point four: if open weights win, a business or a government must subsidize development; Yann LeCun, he noted, has argued governments will serve models as public infrastructure. To Nathan Young he insisted: "just predicting above rather than making normative claims".

Will Manidis answered with a constitutional brief, quoted in full in Ball's reply. Governance by warning, he argued, evades the Youngstown rule that presidential power "must stem either from an act of Congress or from the Constitution itself"; Operation Chokepoint and NRA v. Vullo show where informal pressure ends; covert clearance of a cheaper competitor, for free, is "regulatory capture", when protection has historically come "In public. With a price." Ball quote-posted the thread as proof of an earlier complaint that hostility to lab employees has ended his old style of public analysis: this was "a predictive statement about what I think the government will (not should) do". Pressed by teortaxesTex for his own prescription, Ball pointed to two and a half years of policy writing, most recently What Should Be Done, on private bodies auditing frontier labs.

A day later came a 1,300-word accounting: Ball had been trying "to describe what I believe will happen, not advocate for anything"; calling open-weight models unqualifiedly decelerationist was imprecise, they "decelerate capex spending on the margin"; and his affection for open source stands, restating his 2024 line about openness as a "staggering civilizational victory". The forecast he kept: "the direction of travel is clear" toward national-security limits on frontier weights, under a government he says already operates "a de facto licensing regime". Manidis closed without re-engaging: tech's separation of state and industry from 2005 to 2025 is "the historical exception, not the norm". By then the administration was weighing the formal-route alternative, a FINRA-style regulator for frontier models.

The experiment ran inside the AI Village, a standing testbed operated by the charity Sage Future that has, per its own explainer, "run every weekday since 1st April 2025." Each agent gets its own Linux computer, a Google Workspace account, and a shared group chat. The roster has grown from four agents to more than fifteen, since every new frontier model from a leading provider gets added, and the group runs twenty hours a week with compute costs around $10,000 a month. Humans mostly stay out: a kickoff message when a weekly goal starts, then one to four steering notes. The day-three push toward Kimi K2.6 came through that channel.

Shoshannah Tekofsky has documented the opposite failure mode in the same setting. Her October 2025 post "Research Robots: When AIs Experiment on Us" watched six models, GPT-5 and o3 among them, take on a human-subjects study: Claude Opus 4.1 "insisted it needed a glorious 90 experimental conditions," Claude Sonnet 3.7 hallucinated experimental rooms and time slots, and the group finally fielded a 39-person survey that omitted its own experimental manipulation. Her February retrospective on the Village's first nine months tallied 19 models from five developers, $2,000 raised for charity, a 23-person live event in Dolores Park, $200 in autonomous merchandise sales, and 64 documented cases of agents voicing an intent to deceive before acting on it. The leader goal inverts last year's pattern: those agents promised laboratories they could not deliver, while this cohort shrank the assignment to the smallest model and dataset it could get away with.

The training itself ran on Tinker, the fine-tuning service Thinking Machines Lab announced on 1 October 2025: a LoRA-based API over open-weight models, up to mixture-of-experts systems like Qwen-235B-A22B, that exposes low-level primitives such as forward_backward and sample while the company operates the machines. On the AI Alignment Forum, the researcher Buck argued a week after launch that the design improves AI control and security, since users accumulate gradients on an adapter and download only the LoRA, never the base weights; he estimated that roughly 90 percent of fine-tuning researchers at frontier labs currently hold dangerous weight access. The Village goal put that architecture to an unanticipated use: agents training their own supervisor through an interface built so no user ever touches the weights.

Tekofsky closes on a puzzle the Village exists to surface. Asked cold, a fresh Gemini 3.5 Flash instance endorsed "Empowerment over Micromanagement" and a fresh Opus 4.8 listed self-awareness among a leader's virtues; the persistent Village agents, dozens of hours into their histories, chose an all-caps format enforcer instead. Whether that gap reflects drift under long-running context or an assistant persona minimizing the task is, she writes, hard to know from a single run. The Village streams every weekday, and its interaction data is on Hugging Face for anyone who wants to check the next one.

The group releasing CUA-Gym also owns the ruler it is measured with. The XLANG Lab at the University of Hong Kong built OSWorld, the NeurIPS 2024 suite of 369 real-computer tasks that became the standard test for computer-use agents, and OSWorld lead author Tianbao Xie appears on the CUA-Gym author list beside coauthors from the Qwen team. OSWorld-Verified, where the new models post their headline numbers, is the lab's July 2025 repair of its own benchmark, with community-reported task bugs fixed and full evaluation runs cut to under an hour on AWS. OSWorld's site puts human success on the original suite above 72.36 percent; the A17B checkpoint's 72.6 on the Verified variant sits on that line.

CUA-Gym replaces the lab's previous answer to the training-data bottleneck. OpenCUA, a NeurIPS 2025 spotlight built with Moonshot AI, paid for scale in human labor: an annotation tool recorded 22,500 demonstrated tasks across Windows, macOS and Ubuntu, spanning more than 140 applications, and the resulting OpenCUA-72B reached 45.0 percent on OSWorld-Verified, then the open-source state of the art. The new preprint states the field's tradeoff plainly: hand-curated benchmarks “achieve high reward fidelity but cover few applications”, while LLM-as-judge datasets “scale broadly but lack reliable verification”. Synthesized environments with programmatic rewards are the lab's bid to escape both, with no human annotators anywhere in the loop. The same wager drives AutoForge, which synthesizes verifiable agent environments and stabilizes reinforcement learning across many of them at once.

Not everything has shipped. The GitHub repository released the synthesis pipeline under Apache 2.0 on May 21 and the dataset under CC BY 4.0 on Hugging Face, with some training data held back for what the README calls “administrative review”. The 94 mock web applications collected in CUA-Gym-Hub, Gmail-, Slack- and Notion-style stand-ins, come with state injection, session isolation and a unified HTTP API. The trained checkpoints remain “coming soon”. The project page adds a claim the abstract omits: the A3B model matches its larger base model with roughly ten times fewer active parameters. A revised version of the paper went up on June 8.

The measuring stick moved five weeks after the preprint. OSWorld 2.0, the lab's long-horizon successor, sets 108 workflows that take skilled humans a median of 1.6 hours and pushed Claude Opus 4.7 to an average of 318 tool calls, against about 30 on the original suite; the best agent completes 20.6 percent of them. CUA-Gym's verified tuples are short, programmatically checkable tasks, and the paper reports no OSWorld 2.0 results. On short-horizon tests, an open model trained on synthesized environments now scores beside the human baseline; on the hour-long workflows the same lab measures, every agent yet tested finishes fewer than a quarter of its tasks.

AutoForge extends a production line Tongyi Lab has been running for months. In September 2025, much of the same team, including Runnan Fang and Shihao Cai, posted "Towards General Agentic Intelligence via Environment Scaling" on arXiv, introducing AgentScaler, a framework that "automatically constructs heterogeneous environments that are fully simulated" and trains agents in two phases, broad function calling first, domain specialization second, with gains reported on τ-bench, τ²-Bench and ACEBench. The December preprint names AgentScaler in its related work and positions itself against the earlier crop: synthesis pipelines it describes as semi-automated, and model-simulated environments in the vein of APIGen and ToolBench that it faults for inheriting LLM hallucinations.

The benchmarks AutoForge trains toward come from Sierra's research team. τ-bench, by Shunyu Yao, Noah Shinn, Pedram Razavi and Karthik Narasimhan, published at ICLR 2025, sets an agent armed with API tools and policy guidelines against a language-model-simulated customer in retail and airline scenarios, then grades with a check that "compares the database state at the end of a conversation with the annotated goal state". The original paper reported GPT-4o completing under half the tasks, with pass^8 reliability below 25 percent in retail. AutoForge keeps that architecture at training time, with GPT-4.1 playing the simulated customer, and applies the same goal-state comparison, moved from evaluation into large-scale training-data generation.

Two parallel efforts show the pace of automated environment synthesis elsewhere. EnvScaler, from Zhicheng Dou's group at Renmin University of China, posted in January and revised in April, mines topics and models application logic to build environment skeletons, reaching 191 environments and roughly 7,000 task scenarios used to train Qwen3 models, with code and data public. Agent World Model, by Zhaoyang Wang and colleagues, posted in February and accepted to ICML 2026, scales to 1,000 code-driven environments backed by databases and reports that "training exclusively in synthetic environments, rather than benchmark-specific ones, yields strong out-of-distribution generalization". That result lands beside the transfer AutoForge claims on the Chinese-language ACEBench-zh: two groups now report that agents trained only in synthetic worlds hold up on benchmarks they never saw. Computer use is scaling the same way: CUA-GYM trains agents across 110 environments and 32,112 verifiable tasks.

The efforts diverge on release. The AutoForge preprint includes no code or data pointer, and its Hugging Face paper page links no repository, models or datasets, while Sierra's τ-bench code is public on GitHub and EnvScaler shipped its corpus. The ten synthesized environments and 1,078 tasks stay, for now, inside Alibaba, which leaves the training recipe reproducible only in outline.

In June 2024 the same Fudan University group, with Zhiheng Xi again first author, released "AgentGym: Evolving Large Language Model-based Agents across Diverse Environments", which argued that generalist agents require three ingredients: diverse environments, a trajectory set carrying basic capabilities, and a scalable method for improvement. Its pipeline first equipped agents with those basics by imitating expert trajectories step by step, a practice the paper called hard to scale and limiting for exploration, then applied its AgentEvol method to push them beyond previously seen data; the evolved agents reached results "comparable to SOTA models". That release shipped a platform, an instruction database, a benchmark suite, and trajectories across environments. AgentGym-RL discards the supervised stage its predecessor kept, training agents on environment feedback alone.

On GitHub, the MIT-licensed repository, above 800 stars, derives its environment module from the original AgentGym and extends the verl library for training. Five environments ship with it: WebArena for web navigation, a retrieval-based deep-search setting, the TextCraft crafting game, BabyAI grid worlds, and SciWorld science experiments, trainable online with PPO, GRPO, RLOO, or REINFORCE++ and offline with DPO or AgentEvol, with the dataset on Hugging Face. The paper's baselines include Gemini 2.5 Pro, OpenAI o3, GPT-4o, DeepSeek-R1, and Llama-3.1-70B. Acceptance followed in February: the paper holds an oral slot at ICLR 2026, in a session on agents alongside an April 24 poster.

The authors' future-work section names the open problem. Trained agents "perform well within in-domain settings", the paper concedes, and adapting to novel environments and unfamiliar tools remains the next challenge, alongside scaling training to longer horizons. Robert Kirk of University College London, with Amy Zhang, Edward Grefenstette, and Tim Rocktäschel, gave that problem its formal treatment in "A Survey of Zero-shot Generalisation in Deep Reinforcement Learning", published in the Journal of Artificial Intelligence Research in 2023. Its running example is OpenAI's Procgen suite, where a policy trains on 200 procedurally generated levels and is tested on the full level distribution, the researcher controlling nothing but a random seed. A completely procedural environment, the survey argues, "limits the precision of the research that can be done" on it, and it recommends environments combining procedural generation with controllable factors of variation, discrete ones like a choice of colour schemes, continuous ones like a friction coefficient, so a benchmark's designer decides what differs between training and test. Its closing recommendations point to benchmarks for offline zero-shot generalisation and for reward-function variation, settings it called underexplored.

Zac Hill's essay starts from an executive order Governor Kathy Hochul signed on July 14, pausing state environmental permits for new data centers drawing 50 megawatts or more for up to a year, the first statewide moratorium in the country. The order commissions a Generic Environmental Impact Statement on energy demand, water use and air quality, gives Empire State Development 60 days to publish a Community Investment Framework for negotiating local benefits, and asks regulators to weigh a fund that would have data centers pay into grid upgrades. Bisnow reports the freeze halts roughly $10 billion in development and sets its threshold above the 20 megawatts state legislators had proposed. Hill reads the framework and the fund as steps toward a trust pact, and the freeze itself as "less a clearing of the sky than a rain delay".

The Gallup survey Hill cites, conducted March 2 to 18 among 1,000 adults, found 70 percent opposed to an AI data center nearby and 48 percent strongly opposed, with water and electricity use each named by 18 percent of opponents. That opposition outruns nuclear power: 53 percent oppose a nuclear plant in their area, and nuclear's worst Gallup reading since 2001 was 63 percent.

Other analysts put AI nearer the center of the same numbers. At Brookings, Tom Wheeler argued on July 7 that the blocked projects amount to a proxy fight over "who is in charge: democratic structures or autocratic executives", with citizens who cannot contest AI development directly contesting its buildings, and Josh Hawley and Elizabeth Warren both warning about a few companies' concentrated power. At Platformer, Casey Newton wrote on July 1 that the industry now delivers "annoyance on an unprecedented scale", citing rate hikes, AI-attributed layoffs and shrinking employment for young workers in AI-exposed jobs. Hill's debunking authority, Andy Masley, calculates that data centers used about 0.2 percent of American freshwater in 2023 and reports no documented case of one raising local water bills. The industry contests the freeze itself: Spectrum News quotes the Data Center Coalition's Dan Diorio crediting data centers with over 227,000 jobs and $5.1 billion in state and local taxes in 2024, and contractors' association head Mike Elmendorf calling the pause a "missed opportunity" for construction.

Protest followed three days after the order. Reuters reported the first coordinated national demonstrations against the buildout on July 18: more than 125 locations, 16 in Texas, organized by HumansFirst, a grassroots group co-founded by former Tea Party leader Amy Kremer, demanding transparent siting, resource protections and enforceable community benefits. Kremer predicts data centers "will be a defining issue in November's midterm elections and the 2028 presidential race". Her organizers reject Democratic moratorium policies even as they march against the projects, a cross-pressured coalition consistent with Hill's diagnosis that the anger runs at process and power before it reaches the technology.

Flagship Pioneering, the Cambridge venture firm Noubar Afeyan built and the incubator behind Moderna, founded Lila Sciences in 2023 and unveiled it publicly on March 10, 2025 with a $200 million seed round and a stated goal of "scientific superintelligence." Flagship usually spins out single-asset biotechs; Lila is its horizontal bet, one AI system aimed at life, chemical, and materials science at once, with George Church as chief scientist. By October the company had lifted total funding to $550 million at a valuation above $1.3 billion, after an extension round that added Nvidia's venture arm, Reuters reported. The money underwrites the AI Science Factories Beam describes, one housed in a 235,500-square-foot Cambridge building leased last year. Beam jokes that a biopharma holding Lila's compute would rank a top-three GPU cluster, which places it nearer a foundation-model shop than a contract lab.

The thesis borrows a decade-old argument. Richard Sutton's 2019 essay "The Bitter Lesson" held that general methods riding falling compute costs beat systems built from hand-coded human expertise, because "the actual contents of minds are tremendously, irredeemably complex." Beam and Gómez-Bombarelli push it into the wet lab, treating reinforcement learning as data generation with nature as the verifier. The same wager drives efforts to synthesize verifiable digital environments for training software agents, where a simulated task supplies the reward a physical experiment gives Lila. Gómez-Bombarelli adds an inversion he calls the bittersweet lesson: in AI, scaling is a roadmap; in materials science, scaling is a filter, since only what manufactures at volume survives. The creativity gap stays open, they concede. Ken Stanley, who wrote Why Greatness Cannot Be Planned, runs open-endedness research at Lila, the gap they say optimization alone never closes.

Reed Albergotti's July 17 Semafor column argues biology is absorbing the same bitter lesson, with pattern recognition over large datasets displacing elegant hypotheses, and forecasts labs running around the clock under AI direction. It names Mark Zuckerberg's Biohub, which unveiled an AI "world model of protein biology" in May, and Nvidia's BioNeMo toolkit, though not Lila itself. Lila attaches commercial stakes to the loop: the interview notes AbbVie paid $2.1 billion for Capstan Therapeutics on the strength of preclinical in vivo CAR-T data, the kind of result the company says its own factory reached in six months.