Yesterday in AI · 15 July 2026

Tracked here since June 20 as local moratoriums and permit disputes, the effort to slow AI data centers reached Albany on July 14: Gov. Kathy Hochul signed Executive Order No. 62, the nation’s first statewide moratorium on new hyperscale data centers. For up to a year, while the Department of Public Service drafts a Generic Environmental Impact Statement covering energy demand, water, and air quality, the Department of Environmental Conservation will hold every data-center application for a discretionary permit not already deemed complete. The order’s definition covers facilities that consume, or can consume, 50 megawatts or more. New York will “lead the way in creating the strongest standards in the nation,” Hochul said. Her office paired the pause with the Energize NY proceeding, begun this year to make data centers pay more for energy or supply their own, and Empire State Development guidance on community-benefit deals.

The pause touches few projects, per Heatmap Pro data in Robinson Meyer’s July 14 Heatmap Daily newsletter. Of eight large data centers recently proposed in New York, three are already canceled and one was approved last year; the developer of an upper Hudson Valley campus that could have covered a million acres and drawn up to 1,000 megawatts scrapped it last month after the town passed its own moratorium. Eight New York municipalities have banned data-center development outright and three have restricted it, Meyer wrote, so most places where a project would pencil out have already closed the door.

The legislature had gone further: its Responsible Data Center Development Act, passed last month, would freeze approvals above 20 megawatts for a year, and amNewYork reports Hochul has not committed to signing it. Senate sponsor Kristen Gonzalez, at the announcement, praised the order while noting its higher threshold: “If big tech is coming onto our turf, it should be on our terms.” Meyer read the order as a way around the bill and tied its timing to Hochul’s re-election this fall, recalling her June 2024 congestion-pricing suspension, revived after that November’s election. He doubts the pause gets extended past 12 months; whether other Democratic-run states follow is another question.

In a Semafor exclusive the same day, David Weigel reported Sanders-aligned progressives working to make AI-industry PAC money as toxic in Democratic primaries as AIPAC became. Wisconsin state Rep. Francesca Hong calls herself “the only one in this race that supports a one-year moratorium” on new hyperscale construction; a February Marquette Law School poll found 70% of Wisconsinites and 85% of Democrats saying the centers’ costs outweigh the benefits. Nida Allam, whose 2022 challenge to North Carolina Rep. Valerie Foushee turned on AIPAC support, is challenging her again on data centers. The International Brotherhood of Electrical Workers endorsed Michigan’s Matt Maasdam, warning that “data center bans eliminate good-paying union construction jobs”; Sunrise Movement president Denae Ávila-Dickson told Semafor developers “will give $2 million to a candidate who’s going to let them do that.” Sanders and Rep. Alexandria Ocasio-Cortez proposed a federal moratorium in March.

In a 15 July University of Sydney speech, “AI in Australia’s interests,” and a same-day media release, Prime Minister Anthony Albanese established an Office of AI inside the Department of the Prime Minister and Cabinet, effective immediately, to coordinate new Australian Standards for AI after a response he said had run “issue-by-issue, sector by sector.” The standards fold the Data Centre Expectations issued in March into one framework covering data centres and AI training alike, “clear, consistent and mandatory,” and, the government says, “the first to be legislated by a government worldwide.” Industry minister Tim Ayres joined him alongside assistant minister Andrew Charlton, who called “a clear and enforceable social licence for AI” fundamental to the plan.

Albanese reserved the heaviest obligations for large data centres. The next generation of large-scale centres will carry a legal obligation to underwrite new power supply, pay their full share of grid connection so no costs pass to homes or businesses, and put “at least as much energy into our grid as they take out of it,” building new renewable generation and firming. They must also minimise water use, maximise energy efficiency, and pay for any extra water infrastructure; the release adds that operators must reduce power when the grid needs it, with siting settled with states and communities so centres do not compete with housing. The government promises faster approvals; Forbes Australia reported the pitch as making Australia “more than a data warehouse” for products made overseas.

On copyright, Albanese said “not everything produced in Australia is up for grabs”: writers, musicians, artists and journalists keep ownership and control of their work, including its price and value, and no company should train AI on Australian books, music, art or news without the artist’s control. “An artist’s creative endeavour is their work and their property,” he said, with the Attorney-General consulting on how the law will spell that out. SBS News reported the pledge answered fears the laws might be loosened after Anthropic lobbied for copyright clarity alongside a proposed $21.6 billion Australian investment; the Media, Entertainment and Arts Alliance and APRA AMCOS welcomed the commitment. Consumer-safety priorities follow within weeks, building on the new AI Safety Institute.

Albanese takes the standards to National Cabinet in August and wants legislation in early 2027, while ministers carry assignments from energy and schools to a digital duty of care and Five Eyes national-security work. Reaction reported by SBS News broke both ways. Greens senator David Shoebridge called an office without statutory powers “a single door in his office,” and Opposition Leader Angus Taylor warned of “more bureaucracy” slowing adoption. Greenpeace’s Joe Rafalowicz called data centres “water-guzzling energy vampires” and wants approvals paused until binding rules arrive, in 2027 at the earliest. Digital Rights Watch’s Lizzie O’Shea backed the office but warned fast-tracked centres would crowd out renewables and housing, and the Tech Policy Design Institute’s Johanna Weaver said the plan must be “fast followed with hard decisions” on copyright, jobs and the environment.

The 27th news day of the debate over public ownership of AI firms, tracked here since June 19, belongs to the two scholars whose research seeded Bernie Sanders' bill. In a July 15 essay for the Law and Political Economy Project blog, "What A Tax On AI Can Teach The Left," Jeremy Bearer-Friend of George Washington University and Sarah Polcz of UC Davis write that Sanders' American A.I. Sovereign Wealth Fund Act, introduced June 18, puts into practice the equity tax they proposed in "Sharing the Algorithm: The Tax Solution to Generative AI" (Columbia Journal of Tax Law, 2025).

Their design starts from a premise they say tax law rarely treats as a choice: taxes need not be paid in cash. An excise tax would be payable in newly issued shares, contributed to a public trust that could sell holdings for revenue without diluting its influence over corporate governance. Sanders' plan pairs the 50 percent stake his office values at $7 trillion with guaranteed board seats and a bipartisan Independent Commission for Democratic AI modeled on the Federal Reserve. For legal footing the authors reach back to a 1929 Supreme Court ruling that let Maryland assess and collect a tax in oyster shells. Co-ownership, they add, repays the public for the labor and data scraped to train the models.

The essay spends most of its length heading off the alternative. Bearer-Friend and Polcz warn against the voluntary, firm-by-firm arrangement that OpenAI's chief executive and President Trump have floated since Altman met Sanders and backed some public ownership short of 50 percent. They call that approach "another form of corruption": non-voting or revenue-sharing rights negotiated behind closed doors, reversible by a later administration, and traded for stalled regulation. A statute, they argue, would offer the industry predictability and the rule of law. As evidence that ad hoc government relationships leave companies exposed, they point to the June 12 directive under which Anthropic disabled its Fable 5 and Mythos 5 models, an export-control order Anthropic itself disputed.

The lesson in the essay's title reaches past AI. Require every firm listed on a US exchange to remit 2 percent of its outstanding stock into a fund of this kind, the authors close, and the US could capitalize a trillion-dollar reparations fund in less than a fiscal quarter.

In “Introducing Inkling,” published July 15, Thinking Machines Lab announced its first open-weights model, an American entry in the trillion-parameter open field tracked here since Meituan’s LongCat 2.0 on July 1. Inkling is a mixture-of-experts transformer: 975 billion total parameters, 41 billion active per token, native input across text, images, and audio, and a context window reaching one million tokens. Pretraining ran from scratch over 45 trillion text, image, audio, and video tokens on NVIDIA GB300 NVL72 systems, and the Hugging Face model card lists the license as Apache 2.0. “Inkling is not the strongest overall model available today, open or closed,” the announcement says, pitching the weights as a customization base.

The announcement’s centerpiece has Inkling retrain itself: told to write without the letter “e,” it drafted the plan, generated its own evaluation and synthetic training data, ran a 96-step fine-tuning job on Tinker, and loaded the new weights, about 27 minutes end to end. Asked what a team should do after shipping a large model, the retrained version answered without a single “e”: “party, thank staff, post a summary, watch for bugs, fix faults fast.”

The lab’s write-up and a companion Hugging Face post list the departures from the common recipe. The expert layout follows DeepSeek-V3: 256 routed experts and two shared, six routed firing per token. Relative positional embeddings replace rotary ones, and sliding-window and global attention interleave at a 5-to-1 ratio. Multimodality skips the usual encoder; audio enters as dMel spectrograms, images as 40-by-40-pixel patches. Training combined Muon and Adam optimizers and ran more than 30 million reinforcement-learning rollouts. On X, Jack Morris called Inkling the only open-weight model trained without distilling from OpenAI or Anthropic; the announcement credits post-training synthetic data from open models including Kimi K2.5.

At the model’s top thinking-effort setting, the announcement’s table puts Inkling ahead of Nemotron 3 Ultra, the other American model listed: 77.6% to 70.7% on SWE-bench Verified, 46.0% to 37.4% on Humanity’s Last Exam with tools, and matching Terminal Bench scores at about a third the generated tokens. The leading Chinese open models stay ahead: GLM 5.2 posts 80.0% on SWE-bench, Kimi K2.6 leads MMMU Pro 79.0% to 73.5%, with the three within three points on AIME 2026. Claude Fable 5 tops the same table at 95.0% on SWE-bench. Inkling’s numbers come from the lab’s own harness; competitors’ scores are self-reported. Nathan Lambert called it the “new best American model” on Bluesky.

Thinking Machines is distributing Inkling broadly: full weights, original and NVFP4 for NVIDIA Blackwell, sit on Hugging Face; fine-tuning is live on Tinker at a temporary 50% discount, with inference through TogetherAI, Fireworks, Modal, Databricks, and Baseten. LMSYS Org shipped day-0 support, inviting developers to “run and shape it on an open stack”: SGLang serves 71,700 input tokens per second, and reinforcement learning runs on Miles, a customized Megatron backend. A 276-billion-parameter Inkling-Small, 12 billion active, follows once testing completes; the announcement says it matches or exceeds the full model on many benchmarks.

Alex Turner, a research scientist who spent more than two years on AI safety at Google DeepMind, published an essay on July 15 explaining why he quit. He resigned in June, he told Business Insider, after Google signed an agreement letting the Pentagon use its AI for classified operations. "When Google signed the deal, my conscience simply said 'nope,'" he said. He argues Google broke the explicit promise on which it acquired DeepMind in 2014: that its AI would never be used for military or weapons purposes.

Turner's account starts in January, when Department of Homeland Security officers killed Renée Good and Alex Pretti. Researching Google's exposure, he learned the company sells Cloud services to Immigration and Customs Enforcement through third parties; his campaign widened once the Pentagon set a 180-day deadline for AI contractors to accept "any lawful use" terms. He wrote a 25-page proposal, "A Red Line and Oversight Framework," gathered about 250 signatures on a petition to Google Chief Scientist Jeff Dean, and joined roughly 600 employees asking Google CEO Sundar Pichai to keep the company out of deals involving classified work. In a thread on X, Turner says he cold-messaged DeepMind CEO Demis Hassabis, who routed the proposal to two senior policy staff; they left it unevaluated until the deal closed.

Business Insider's Hugh Langley reports the Pentagon confirmed in early May a deal with Google, Microsoft, Amazon, and OpenAI for "lawful operational use." The contract language Turner quotes says Google's AI "should not" be used for mass surveillance or autonomous weapons without human oversight, wording he calls aspirational because the agreement gives Google no veto over government operations. In May, a Google spokesperson said the company remains committed to the consensus that AI should not be used for "domestic mass surveillance or autonomous weaponry without appropriate human oversight." Turner reads Google's restrictions as weaker than OpenAI's, and points to a contradiction: Hassabis co-authored the February 2025 update to Google's AI principles that removed its weapons and surveillance prohibitions, then told an interviewer "Nothing's changed about our principles."

The essay presses hardest on people Turner expected to hold firm. At a conference of the International Association for Safe and Ethical AI, he says, Stuart Russell, in whose lab he once worked, agreed on stage to poll members on a statement supporting Anthropic; both the poll and the statement vanished. Dean, who signed a 2018 pledge against lethal autonomous weapons, co-signed an amicus brief backing Anthropic against the Pentagon. "Jeff is still at Google, despite his pledge," Turner writes. He grants that Google's ethics review can bite, citing Bloomberg reporting that the company exited a $100 million Pentagon drone-swarm contest in February. Anthropic, he writes, resisted the Pentagon's terms; there, "ethics won." "Pledges of conscience often vaporize on contact with power," he wrote on X.

Writing in Pirate Wires on July 14, Trae Stephens argues that AI's early damage to entry-level white-collar work has revived the case for mandatory national service. Stephens is a partner at Founders Fund, a cofounder of the defense company Anduril Industries, and an early Palantir employee. He opens with former Google CEO Eric Schmidt's May commencement address at the University of Arizona, where the crowd booed a pitch for AI's promise until Schmidt pleaded, "May I finish?" Young graduates, he writes, are right to worry.

His case rests on job-loss data. Joblessness among recent college graduates, Stephens writes, is climbing at double the rate of the workforce at large, and employment is falling fastest for workers in their early 20s in the roles AI absorbs first: paralegals, customer service representatives, and entry-level coders. The central figure checks out. A Yale School of Management analysis Stephens links, published May 4 by Jeffrey Sonnenfeld and colleagues, put recent-graduate unemployment near 6 percent, "rising twice as fast as the rest of the workforce since 2022," with computer science graduates near 7 percent and employment among developers aged 22 to 25 down nearly 20 percent from a late-2022 peak.

Even before AI, Stephens adds, college graduates sensed the old deal had stopped working. The same analysis found just 19 percent of graduates calling it a good time to find a quality job, against 35 percent of those without degrees, and software-development job postings down 53 percent over the same period.

The essay runs beneath a 1933 photograph of Civilian Conservation Corps recruits. The usual arguments for service, he writes, have never moved the public: patriotism through shared sacrifice, infrastructure built by young workers who need hard skills, and the mixing of classes and parties that a shared obligation would force. Concrete programs, he says, have fizzled. Britannica's ProCon explainer, updated June 2, casts current proposals as military duty or civilian work such as teaching in low-income areas, elder care, and infrastructure upkeep.

Stephens wagers that a labor shock landing first on young graduates could sell a program that patriotism and public works never could. His subtitle invokes "uncertainty, radicalization, and cynicism," the mood he assigns to a generation whose path from campus to career is narrowing.

Google DeepMind published a July essay, "Conjecture Machines: AI agents and the new validation bottleneck in science," by Don Wallace, Conor Griffin, Sean O'Neill, Thang Luong, and Owen Larter, drawn from conversations with ten of the company's researchers and engineers. Its argument runs through Karl Popper: science advances by conjectures and refutations, and AI agents are conjecture machines that make ideas abundant and cheap while refutation stays physical, institutional, and slow. The essay opens with microbiologist José Penadés of Imperial College London, whose team spent most of a decade working out how a family of superbugs spreads antibiotic resistance, a result they never published. In 2024 he posed the problem to DeepMind's Co-Scientist; within two days it returned five ranked hypotheses, the top one matching his own unpublished answer.

DeepMind's own systems supply most of the essay's examples. Gary Peltz of Stanford used Co-Scientist to find existing drugs for liver fibrosis, the scarring behind 1.4 million cirrhosis deaths a year; neither of his two hand-picked candidates helped live human liver cells, while two of Co-Scientist's three picks both blocked fibrosis and regenerated liver tissue. AlphaEvolve, which generates and scores algorithmic candidates, has helped design Google's next-generation TPU chips, aided Terence Tao on open Erdős problems, and improved genomics analysis. At February's inaugural First Proof challenge, ten research problems were kept unpublished so they could not appear in any training data; DeepMind's Aletheia, which pairs a proof generator with a natural-language verifier, solved six, the best result.

Against those results the authors set the limits. Vivek Natarajan, a Co-Scientist lead, warns that "a single hallucinated claim on page 10 of an output can invalidate the whole thing"; the authors add that calibrating confidence for open-ended scientific reasoning remains unsolved. Thang Luong, who led Aletheia, sees mathematics drifting toward "proof indigestion" as machines produce results faster than people can check them. Where an answer can be verified in silico, as with a proof written in Lean, validation keeps pace; elsewhere it lags. An agent can propose a genetic lead to reverse cellular ageing, but cannot say whether it works.

The essay closes with four demands on funders and policymakers: widespread agent access, agent-ready public data, more validation capacity, and reformed peer review. On validation, DeepMind reports building a wet lab inside the UK's Francis Crick Institute and funding independent scientists to run experiments that test agent-generated hypotheses; it points to the US Genesis Mission linking Energy Department labs with academia and industry, the National Science Foundation's $100 million for a network of distributed facilities, and the UK's £81 million Materials Innovation Factory. On peer review, with AI now drafting grants and papers, the authors propose "Human-AI Interaction Cards" that record the inputs and outputs behind key findings, and note the UK Medical Research Council has reinstated interviews for shortlisted applicants.

Enterprise agent engineering, tracked here since Sierra's July 2 account of deploying agents inside large organizations, drew a field-wide reading this week. In a July 14 Latent Space post, "5 Trends That Defined AI Engineering at World's Fair 2026," Richard MacManus argues the field has shifted from building autonomous agents to engineering the systems around them. He frames it through two essays by Lilian Weng, co-founder of Thinking Machines Lab: her 2023 "LLM Powered Autonomous Agents," which cast an agent as planning, memory, and tools, and her July 4 "Harness Engineering for Self-Improvement." Weng defines the harness as the system deciding how a model plans, calls tools, manages context, and judges its output, adding "evaluation, permission controls, and persistent state management" to the older recipe.

MacManus's second trend, "loop engineering," splits the work in two: agents run an inner execution loop while engineers hold an outer loop of direction, feedback, and evaluation. Roland Gavrilescu, co-founder of the self-improvement startup Introspection, described an outer system that "studies and maintains the primary system"; former Google engineering leader Addy Osmani said "that outer loop is still engineering." A closing-day debate aired the strain. HumanLayer's Dex Horthy warned that "the hype is outrunning the discipline," noting Kubernetes control loops are deterministic and agent loops are not. Geoffrey Huntley, who built the Ralph Loop, likened engineers to locomotive operators whose job is "to keep the locomotive on the rails."

Enterprise adoption, MacManus's third trend, runs through "forward deployed engineers" who fit agents to a company's systems and data. Cursor's Pauline Brunet said her team leaves behind long-running agents on the Cursor SDK so that "when we walk away, it is a strict ROI for them," though adoption "is still concentrated among early adopters." Warp CEO Zach Lloyd described Oz, his "software factory" platform, where organizations pick which repositories and lifecycle stages to automate and where humans review high-risk changes.

Fourth on MacManus's list are the coding agents themselves. Claude Code, Codex, Gemini CLI, Cursor, and Warp now explore repositories, edit files, and debug; Vercel's Andrew Qu, presenting the new "eve" agent framework, said agents "are not as predictable as web applications." Conductor's Charlie Holtz resisted the software-factory image, preferring to stand "in front of an orchestra, waving my baton."

MacManus's fifth trend gathers every platform around "skills," the Markdown procedures Anthropic added to Claude last October. Google DeepMind's Philipp Schmid argued that "agents are just files," cutting orchestration code once written in Python; Y Combinator president Garry Tan said AI-native startups encode sales, support, and finance as "written procedures that their agents execute." Paul Bakaus, whose open-source Impeccable gives coding agents a design vocabulary, warned that models "converge in one direction," so if everyone leans on the same skill, "everything ends up looking the same." Weng's essay lists weak evaluation among the bottlenecks: "many research claims do not have a fast and precise verifier."

In a July 14 conversation on his newsletter, the writer Derek Thompson asked the University of Chicago philosopher Agnes Callard to explain her theory of the "uni-context." A context, she says, is a set of circumstances telling you how to act; for most of history, contexts were local and plural, read off whether you stood in a field, church, or bar. The uni-context replaces them with one set of norms holding everywhere. Thompson gets there through "context collapse," where one post reaches boss, parents, and strangers at once; Callard goes past who sees you to how you should act. She rejects pure technological determinism: radio, television, and smartphones caught on because people already wanted to live in one shared context.

Callard uses the theory to explain why the internet skews negative. Goodness is context-dependent, she argues, while a few evils read as bad nearly everywhere: death, pain, illness, violence. Two strangers online, hunting a topic both can care about, coordinate on something bad. Thompson supplies the example she endorses: praising chicken enchiladas bores a global feed; calling their consumption racist travels, because racism reads as bad everywhere. The same asymmetry favors identity over character. Character, a disposition like courage or irascibility, shows only across many situations; identity categories such as woman, gay, or Jewish stay legible in all of them, "a hat you never take off." An ethics of inclusion follows, since the uni-context must hold everybody, and the positivity bias of Aristotle's virtue ethics fades.

Comparison, in her telling, breeds sameness: let families choose between two school districts and they weigh graduation rates and AP counts, and the schools homogenize to compete. Thompson runs it through baseball, where a shared statistic like WAR ranked every player, and team strategies converged. Callard dates the inflection to around 1910, when theorists like Georg Simmel, Max Weber, and Martin Heidegger described the fragmentation of everything: every value, in her reading, had crowded into one context before anyone could reconcile them. Thompson quotes Simmel's "The Metropolis and Mental Life," where money "becomes the frightful leveler"; Callard points him to "The Philosophy of Money," which defines money as abstract value, abstraction being what makes unlike things comparable.

Callard gives attention the same treatment. It runs bottom-up by default, pulled to whatever is salient, useful for a creature that may face a fire or snarling animal. With a screen in nearly every environment, no local cue says what deserves attention, so attending turns into a management problem, distraction into personal failure, and people buy products to fight the phones they bought. Asked whether the uni-context is good or bad, she echoes the close of Simmel's essay: the people it produced cannot judge it, and she does not feel ready to. Asked for a prescription beyond putting the phone away at dinner, she offers none: people keep choosing the uni-context, she says, out of a deep aversion to "world closure" and a hunger for "world openness," a world wider than the one you were born into.

In an essay for The Argument, “Obsession” and the horror of losing yourself, Jeremiah Johnson argues that the horror films winning prestige and cultural attention have moved their fear inside the self. Johnson, whom editor-in-chief Jerusalem Demsas introduced on July 14 as a staff writer, reads the genre through critic Robin Wood’s idea of the return of the repressed, a society’s denied fears coming back as its monsters. Cold War films brought infiltrators and body snatchers; the 1970s and 1980s slashers punished teenagers for sex and gave Carol Clover’s 1987 article the Final Girl. Johnson dates the current mode to Jordan Peele’s Get Out, which grossed nearly $260 million on a $4.5 million budget, earned four Oscar nominations including Best Picture, and won Best Original Screenplay. Its hero, Chris, is threatened with a surviving body and a suppressed inner self.

The films Johnson lists share that predicament. In Us, the protagonists risk replacement by their doppelgangers, the Tethered; in The Substance, a beauty drug splits Elisabeth Sparkle in two and she ruins herself chasing fame; in Companion, a woman learns she is a rented robot built to believe she is human; in Ryan Coogler’s Sinners, the heroes face becoming vampiric versions of themselves that keep the traits of who they were. Johnson stresses that the harm tends to begin with a character’s own choice or with an intimate: Sparkle swallows the drug freely, and Get Out’s Chris is entrapped by his girlfriend. Peele compressed the idea into a title, Johnson notes: “the real villain is Us.”

Two breakout 2026 horror hits prompted the essay. Johnson describes Obsession, the Focus Features film starring Inde Navarrette, as the story of a young woman whose mind is commandeered by a coworker while a facsimile wears her body; in A24’s Backrooms, the monsters are distorted versions of real people. Their success arrives as horror climbs at the Academy Awards. At the 2025 ceremony The Substance and Nosferatu drew nine nominations between them. A year later Sinners collected 16, the most in Academy history; Amy Madigan won Best Supporting Actress for Weapons; and Guillermo del Toro’s Frankenstein won three awards from nine nominations. Johnson takes Peele’s record as corroboration: Nope, his film least concerned with a stolen self, drew his lowest gross, weakest reviews, and least discussion.

Johnson closes by tying the pattern to the digital age’s machinery. Influencers, algorithms, social platforms, and the growing habit of outsourcing thought to AI, he writes, leave a person unsure which tastes and beliefs are genuinely her own. He sets the era’s compulsions on one continuum: gambling, pornography, obesity, and screen addiction, outcomes of both choice and a designed environment. Thousands of engineers, he observes, are paid to build systems that nudge users toward some altered version of themselves. Beneath the genre, on his telling, sits the fear that audiences have already become strangers to themselves, unable to tell a chosen identity from an absorbed one.