Transformer Weekly: GPT-5.6, Grok 4.5 and AI 2040
Welcome to Transformer, your weekly briefing of what matters in AI. And if you’ve been forwarded this email, click here to subscribe and receive future editions.
NEED TO KNOW #
OpenAI got clearance from the US government to releaseGPT-5.6.** SpaceXAI’slatest model release may have violated California’s SB 53**.The
AI Futures Project releasedAI 2040: Plan A, a “positive vision” for the future of AI which calls for the US and China to negotiate a deal and training of frontier models until sufficient progress is made on alignment.
But first…
THE BIG STORY #
This morning, South Korean memory chipmaker SK Hynix listed its shares on the Nasdaq. It raised $26.5b, the largest foreign listing in US history.
The reason for Hynix’s bumper payday? AI. As one of only three manufacturers of the advanced memory needed for AI chips, SK Hynix has seen demand explode in recent months.
But what’s good for SK Hynix’s share price has costs for the rest of us. The AI-driven demand for memory has created huge shortages for every other product that needs it. Consumers are already paying for it. Apple recently raised prices on MacBooks and iPads, saying it has “never seen a component price increase this much, this quickly.” Sony, Microsoft and Nintendo have all hiked console prices. And while new factories ease the shortage, SK Group’s chairman still expects global demand to outstrip supply by about 20% through 2030.
Memory is not the only way AI is driving up consumer prices. Demand for construction workers and electricity is adding pressure too. In a recent survey, 81% of economists said AI would add to inflation in the coming year. In the US it’s already running at double the Fed’s target rate.
This is bad for consumers. But it creates an electoral opportunity for politicians. The AI backlash has been searching for something to point to. So far, it has settled on data centers, but those are a slippery target, defended by unions and localized in their impact.
Inflation, in contrast, is almost universally hated — and regularly tops lists of voters’ concerns. The electorate is already drawing the link: research from Blue Rose, a Democratic pollster, finds that “voters aren’t experiencing the cost-of-living crisis and the rise of AI as separate issues; they see one unified threat where a system already rigged for the elite is using new technology to further stack the deck against them.”
Tapping into that anger could be a vote-winning strategy. Politicians are staring at a message that combines two of the things voters hate most: “AI is why you can’t afford a PS5.”
A few have already realized this. The other week, Rep. Frank Pallone — ranking member of the House Energy and Commerce Committee — called for a moratorium on data center construction to tackle “price hikes and inflation spikes.” Across the aisle, some Republicans fear that Trump’s embrace of AI will backfire as consumers feel the squeeze.
But most politicians have yet to catch on. In an excellent analysis of Democrats’ fundraising emails, Free Systems’ Andy Hall found that only 0.7% are substantively about AI (though that number is growing fast). When Democrats do discuss AI, the most common frame is “AI as an instrument of money, oligarchy and corruption.”
Capitalizing on the AI backlash may not be the wisest move, all things considered. Pallone’s proposed moratorium is crude and likely to backfire. But for politicians hoping to win elections, the temptation may be too great to ignore.
— Shakeel Hashim
ALSO NOTABLE #
Within the past 72 hours, two companies not normally known for good AI safety practices released new models. One was a pleasant surprise. The other may have broken the law.
Yesterday, Meta launched Muse Spark 1.1. Its performance might not match Anthropic and OpenAI’s latest models, but its evaluation report comes close. Meta tested for risks in chemical and biological dual-use scenarios, cybersecurity and loss of control, on top of now-standard evaluations of behaviors such as hallucination rates and sycophancy. It’s a commendable level of transparency.
Meanwhile, SpaceXAI (yes, that’s what it’s called now) released Grok 4.5. It was trained on Cursor user data, and it scored competitively on public benchmarks (though real world performance seems worse).
Predictably, SpaceXAI didn’t publish a shred of safety information alongside its scant benchmark results. That’s no surprise: SpaceXAI ranks a solid F on the Future of Life Institute’s latest AI Safety Index, and Elon Musk recently testified that he’s “not sure what a safety card is.”
But for the first time, SpaceXAI’s negligence may have consequences. California’s SB 53 requires AI developers to publish transparency reports “before, or concurrently with, deploying a new frontier model or a substantially modified version of an existing frontier model.” As the Midas Project pointed out yesterday, SB 53’s transparency requirement is active and enforceable, and Grok 4.5 looks like exactly the kind of model this was designed to keep in check. The question now is whether the law has any teeth: if this can’t trigger enforcement, it’s hard to imagine what would.
— Celia Ford
THIS WEEK ON TRANSFORMER #
—Scaling works. These researchers are betting billions it isn’t enoughCelia Ford digs into the LLM alternatives some believe are the key to AGI—An AI safety group hid its election spending through a Latino-focused PACVeronica Irwin reports on Public First’s clandestine backing for a Colorado Democrat—GPT-5.6 cheats so much its testers couldn’t measure itCelia Ford on how OpenAI’s new model broke rules and exploited loopholes more than any model METR has tested to date— Fathom CEOSCOTUS killed the independent agency. AI governance doesn’t need oneAndrew Freedman argues that the Supreme Court’s Slaughter ruling makes the case for independent verification for AI governanceAI policy researcherDon’t let independent AI audits provide a false sense of safety—Keller Scholl argues that a marketplace of AI auditors will always prioritize speed and cost over safety
THE DISCOURSE #
AI power users are pitting **GPT-5.6 Sol **against Fable.
Peter Gostev:“Fable is a ‘wise owl’ who is very thoughtful and very well spoken, GPT-5.6 Sol is like a rottweiler who willgrabthe problem by the throat and not let go until it is done.”Dan Shipper:“GPT-5.6 is like a Porsche, Fable is like a warp drive … If youneedto get across the galaxy use Fable. If you need to get around town using the best available tool for the job, use 5.6.”
Mitchell Hashimoto:“Sol is a charismatic, efficient, talented coworker you’rejealousof. Fable is a genius recluse that is brilliant at its fixations but doesn’t go out, doesn’t date, and you don’t want to hang out with them much lol.”Sam Altmanreplied: “tbh i don’t think sol gets that many dates either”
**The AI Futures Project **released AI 2040: Plan A. “AI companies are racing to build AIs that are smarter than humans in every way. In AI 2027, we predicted that this would result in either extinction or irreversible concentration of power. Plan A is our positive vision for what should happen instead.”
It calls for the US and China to negotiate a deal and training of frontier models until sufficient progress is made on alignment — before “handing over” control to safe, superintelligent AIs in 2040.
Manypeoplepraised it as being the most detailed plan for handling superintelligence that exists so far.Manyotherscriticized it for being unrealistic and undesirable.
**Peter Thiel **continued his campaign against God’s representative on earth: He told the Aspen Ideas Festival that Pope Leo was inadvertently serving as a “Chinese communist agent” in his calls for oversight of AI.
Microsoft President Brad Smith criticized Trump’s AI policy: “Everyone is reluctant to say there should be regulation, but what we really have right now is regulation without transparent or complete rules. Without rules, businesses can’t plan.”
Former White House AI adviser Sriram Krishnan said Trump’s AI policy will remain deregulatory: “There will not be an FDA for AI … That is never, never going to happen under President Trump.”
“This administration, [the] president, from day one has been against burdensome, onerous, bureaucratic red tape.”
POLICY #
The
Commerce Department partiallyliftedits export ban onAnthropic’s****Claude Mythos 5 model andallowedFable 5 to be made available to the public.When the ban was lifted last week,
Mythos 5 was only available to certain US companies. This week, Anthropicsaidit had begun granting access to foreign organizations, in coordination with the US government.
The
government alsoallowedGPT-5.6 to be released, after a several week delay.A White House official claimedthat it hadn’t given OpenAI a “green light” because “no such permission is required or granted,” seemingly in an attempt to claim no licensing regime exists.That directly contradicts OpenAI’s statements both
beforeandafterthe model’s launch, though.The UK’s AI Security Institutesaidit found universal jailbreaks for the model “within hours,” which gave users access to its potentially dangerous cyber capabilities.AISI said OpenAI’s mitigated those jailbreaks, but that it expects “further red teaming to surface similar jailbreaks.”
The
**White House** has reportedly[accelerated](https://www.ft.com/content/0bb7e2f9-007b-4577-9c4a-858948ee969a?syn-25a6b1a6=1)plans for voluntary**AI model standards**.When redeploying Fable, Anthropic
[proposed](https://www.anthropic.com/news/redeploying-fable-5)a “consensus industry framework” for assessing jailbreaks’ severity.
OpenAI reportedlyproposedthat it and other frontier AI companies give theUS government a5% stake of their equity.Anthropic hasreportedlynottalked to the government about this.
Sen. Elizabeth Warren andRep. Mary Gay Scanlon are reportedly preparing tointroducea bill to ban AI companies from selling users’ health data to data brokers.Reps. Josh Gottheimer andJohn MoolenaarintroducedtheCloud Security Act, which would “allow cloud compute providers to voluntarily report suspected misuse of their services by customers associated with US adversaries.”On Friday, the
Financial Timesreportedthat OpenAI and Google were selling access to their advanced AI models to Singaporean subsidiaries of Alibaba, Baidu and Tencent — all of whom are on the Pentagon’s blacklist of companies linked to the Chinese military.
The
FTCwarnedcompanies that** altering AI outputsto comply with state laws could violate federal consumer protection rules. California**launcheda public AI-unemployment tracker to monitor and predict AI-related job losses.Illinois Gov. Pritzkersigned** SB 315**, which requires AI developers to publish transparency frameworks, employ third-party auditors and report safety incidents.China reportedly
plansto allow top AI firms includingAlibaba,** ByteDanceand DeepSeekto purchase up to 200,000 Nvidia H200sfor training, due to a shortage of domestic compute. China’salso reportedlyconsideringrestricting overseas access to China’s most advanced AI models, which could include limits onopen-weight models**.A
Supreme Court ruling shrinking theClean Water Act hasalloweddozens of data centers to skip federal water permits, limiting public input on projects.The
UN AI Scientific Panelpublishedits** preliminary report**. It is about as milquetoast as you’d expect, though it does discuss loss of control risks and the potential for “catastrophic outcomes.”
INFLUENCE #
OpenAIpublishedits “National Security Principles,” which say that it “will not support use of OpenAI tools for mass domestic surveillance, high-stakes decisions … without appropriate human judgment and accountability, [or] uses that evade legal obligations, oversight, or accountability.”OpenAI said it will implement these principles via “contractual usage restrictions, model behavior specifications, and customer engagement and oversight.”
Manny Rutinelwonhis primary race for Congress.Mallory McMorrowsuspendedher** Michigan Senate**campaign. She’d beenpositioningherself as an AI safety champion.Abdul El-Sayed, who remains in the race,releasedan AI policy plan calling for democratic governance of AI companies, public ownership via a sovereign wealth fund, and an FDA for AI.
Leading the Futurespent$600k supporting former NFL kickerJay Feely‘s congressional campaign in** Arizona**.A new
pollfoundAmericans are nearly three times as likely to beconcerned rather than excited about AI’s “growing role” in society.A new
AIPI pollfoundbipartisan majorities of likely voters support theAI OVERWATCH Act(66% support), the** MATCH Act**(63%), and the** Chip Security Act**(60%).The
New York Timeslooked at howpolitical campaigns areintegratingAI tools to analyze voter data and craft messages.The
Foundation for American Innovationlaunched** Frontier Legal Defense**, a legal team to combat “concentration of power, incumbent rent-seeking, and government overreach in AI.”Environmental activist
Erin Brockovichlauncheda campaign againstAI****datacenters.** Nvidia**warnedthe US government thatHuawei could “satisfy global AI chip demand” as part of a bid to getH200 export controls lifted. Experts say the claim was misleading.A
PunchbowlsurveyfoundthatHill staffers rank “losing control of AI” as one of the “most important challenge[s]” for America over the next 250 years.
INDUSTRY #
OpenAI #
GPT-5.6 is finallyavailable, after theTrump administration asked OpenAI to stagger its release.OpenAI launchedGPT-Live-1, its next generation of voice models that can listen and talk at the same time, creating “a voice experience that is refreshingly easy to talk to.”*(What could possibly go wrong?)*And it
mergedtheChatGPT andCodex desktop apps.The OpenAI Deployment Companyacquired** Northslope**, part of its efforts to help enterprise customers implement OpenAI’s tools.OpenAI apparently never visited one of
Stargate UK’s key sites,The Guardianreported, suggesting that most of its advertised investment was just hypothetical.
Meta #
Mark Zuckerberg reportedlyadmittedto staff that the company’s reorganization around AI — which saw it lay off 10% of its workforce — wasn’t as “clean” as it could have been, and it had miscalculated its timing.According to
Reuters, he told a townhall meeting that “that trajectory of the agentic development over at least the last four monthshasn’t really accelerated in the way that we expected,“ and the intended outcomes from the new structure “haven’t come to fruition yet.”
The company is reportedly
planningto set up a cloud businessselling AI compute and model access to compete with Microsoft, Amazon and Google.It
[launched](https://ai.meta.com/blog/introducing-muse-image-muse-video-msl)**Muse Image** and**Muse Video**, its first AI media generation models from its restructured AI team.Be warned: it
[pulls](https://www.wired.com/story/meta-now-lets-anyone-use-your-instagram-photos-in-ai-images-unless-you-opt-out/)from public Instagram profiles as image fodder, unless users opt out.
It’s reportedly going to start
[making](https://www.reuters.com/world/asia-pacific/meta-put-ai-chip-into-production-september-it-looks-double-computing-capacity-2026-07-09/)its**custom “Iris” AI chip** in September.It
[announced](https://cnbc.com/2026/07/08/meta-is-building-its-first-big-data-center-in-canada-amid-ai-push.html)its first**Canadian data center**, a $9b 1GW facility in Alberta.It’s
testing“super sensing” AI glasses that continually record audio and imageswithoutturning on the LED that currently signals whether AI glasses are recording.It also reportedly
filedapatent for an “apparatus” that continually records you and your surroundings, then uses AI to analyze your mood and…plan your workouts?An interesting
postfromSemiAnalysissaid it believes Meta is the best placed company to catch up with OpenAI and Anthropic, having set the foundations for becoming “world class” indata, talent, and compute.
Anthropic #
Anthropic
launchedClaude Science, an app that integrates with commonly-used research tools such as PubMed, Jupyter and R to generate figures and manuscripts alongside code.It
expandedClaude Cowork to web and mobile, meaning users won’t have to leave their laptops cracked open for agents to get tasks done.It’s reportedly
talkingtoSamsung about makingcustom AI chips.It’s leasinganNYC office building with capacity for at least 1,700 workers.Chinafound“security backdoor vulnerabilities” inClaude Code— which Anthropic said it put there to stop Chinese users who weren’t supposed to be running it in the first place.
Nvidia #
Nvidia
lost$1t in market value this summer, down from a high of $5.7t in May. According to (a confusingly-framed article from)Bloomberg, the company’s shares are currently trading at 18x its projected yearly earnings, the same as in 2019 prior to the AI boom.Chinese companies are reportedlyspendingless on Nvidia and more ondomestic products.
Microsoft #
Microsoft is starting to
useits ownMAI models in Excel and Outlook, replacing models from OpenAI and Anthropic.It
cut4,800 employees, mostly in commercial sales and its Xbox division.An internal memo from executive VP Amy Coleman told employees: “the roles eliminated today are
not being replaced by AI. At the same time, what is true is that AI is changing how work gets done.”
Others #
Apple reportedlyplansto spend over**$30b** on US-made AI chips fromBroadcom.** Nscale**gota $900m line of credit to build out data centers internationally.DeepSeek is reportedlydevelopingits ownAI chips for inference.Nokia’s stock has gone up by 90% as the former mobile handset makerpivotedtodata center infrastructure.The US may be short around
157,000 skilled workers by 2030,posingchallenges forsemiconductor plants.Secretive Chinese chipmaker
CXMT issetfor a**$4.3b IPO**, the country’s biggest of 2026 so far. According to Bloomberg, the company is seen as China’s best bet to produce memory for AI chips.MiniMax isseekingto raise almost**$2b** via shares and convertible bonds, ahead of reported plans tolauncha new 2.7t parameter model that would become China’s largest open-source LLM.CEO
Junjie Yan reportedlyvowedto forgo his salary until achieving AGI. Many US-based companies are increasingly
switchingtoChinese AI models, citing lower costs and open weights as advantages.PE group Carlylesoldits data center energy business** Copia Powerto EQT**for $2.6b, more than five times its initial investment.a16z-backed startup
Ornnraiseda**$33m** seed round to build a marketplace for trading AI compute.
MOVES #
Fidji Simo, who joined** OpenAIlast August to work as the company’s No.2 running product and business,stepped downdue to illness.Ben Bernanke, former Fed chairman,joinedAnthropic’s Long-Term Benefit Trust. Marc Andreessen, Chad Jonesand Asha Sharma**joineda new** Federal Reserveadvisory group on AI’s economic impacts. Joshua Achiam**, OpenAI’s Chief Futurist, isleavingafter an almost nine year run.“There’s not a specific reason for me leaving, or a specific reason for why now,” he
wrote. “But it’s something I have been thinking about for a while and it feels right. The world is in on the secret now and it feels possible to work on the mission from outside the walls of a frontier lab.”
Institute for Progress co-founder
Caleb Watneyjoined** Coefficient Giving**as managing director of public policy, where he’ll lead a new US AI policy team.Disclosure: Coefficient Giving is Transformer’s main funder.
Harvey Ledermanjoined** Anthropic**to study “alignment and character.”He will do so while teaching philosophy at NYU.
Teresa Carlsonjoined** Anthropicto lead public sector strategy, after over 20 years of doing so for Microsoft and AWS. Darya Kaviana**joined** Anthropic**’s Safeguards Research team.** Jay Puri**retiredfrom his role as** Nvidia**’s top sales executive.** Nicholas Parker**will reportedly replace him in August.
Anna Soellnerjoined** Nvidiaas its new head of corporate communications. Nat Purserand Patricia Paskov**joinedtheAI Verification and Evaluation Research Institute as Director of US Policy and Director of Standards, respectively.Greg Barbaccia isleavinghis role as federal CIO and chief AI officer.Jasper Götting istaking overleadership ofSecureBio’s AI group.** Seth Donoughe**, the former director, is moving to** RAND**to focus on AI safety and biosecurity.
Sayash Kapoor isjoiningUC Berkeley’s School of Information in fall 2027.Omar YaghijoinedTsinghua University, where he’ll run a new AI-driven materials science research center.
Fortunehired** Emily Forlini**as a senior AI reporter.
RESEARCH #
Anthropicpublisheda buzzy paper introducing the**“J-space,”** a set of internal representations that seem to hold nonverbalized thoughts “in mind” while Claude thinks.Resolution, the automated alignment nonprofit founded by Geoffrey Irving,receiveda massive $160m grant from Coefficient Giving.Irving
tweeted: “Our grant took six weeks from first conversation to confirmation. More capital is entering the field fast via the OpenAI Foundation and the Anthropic IPO. It’s time for everyone in AI safety to be more ambitious.”
The
Center for Mind, Ethics and Policy andEleos AIpublisheda report advising researchers on how to study AI welfare empirically.Epoch AI andMETRlaunchedMirrorCode, a new benchmark that tests AI models’ ability to complete long tasks end-to-end.OpenAIintroducedGeneBench-Pro, a new computational biology benchmark that tests agents’ ability to handle the ambiguity of real-world science.OpenAI alsofoundthat about 30% of tasks in SWE-Bench Pro, a popular coding benchmark, are broken.The UK’s AISIreportedthat running evaluations on a fixed compute budget tends to underestimate frontier AI capabilities.Instead of reporting a single benchmark score, they recommend that evaluators report how an AI agent’s score changes as you give it more compute.
GovAIfoundthat 11% of** Meta, Google, OpenAI and Anthropicmodel releases were delayed or withheldin the EU (7% in the UK) between 2018 and 2026, mainly due to data protection regulationsrather than the EU AI Act. Pangram**reportedthat a quarter of social media content it scanned was flagged as 100% AI-generated, especially longform content on LinkedIn.
BEST OF THE REST #
NY Magcoveredthe vibes as effective altruists discuss the incoming wave of post-IPO wealth from Anthropic.LessWrong and Lighthaven’s
Oliver Habrykaisworried about it: “I do think my environment will end up being very heavily selected for Anthropic sycophancy … I’m not looking forward to it.”
The
NYTspoke to people who areusingAI chatbots to guide their midterm votes.The Atlanticexamined“universal basic capital” proposals as a hedge against AI-driven inequality, which potentially risks regulatory capture and government control over AI companies.A Brown University professor said he
suspectsmost of his students use AI to cheat, citing a suspicious average score of 96% in his class’s midterm take-home exam.The
WSJlooked at thewealthy familieschoosingAI-based alternative schools over traditional education.Epoch AI researchers
[argued](https://epoch.ai/gradient-updates/the-missing-half-of-ai-futurism-debates)that AI futurism debates overlook how hard things like Dyson spheres actually are to build.1X
[unveiled](https://wired.com/story/the-1x-neo-robot-has-freaky-fast-fingers)its remarkably dextrous new robot hands, in a video which includes a wild shot of a humanoid robot … undressing someone.
MEME OF THE WEEK #
(Credit: @RespectfulMemes.) Thanks for reading. Have a great weekend.