The flywheels are spinning
Transformer Weekly: Automated AI R&D, a regulatory moratorium, and deals with the Middle East
Welcome to Transformer, your weekly briefing of what matters in AI. If you’ve been forwarded this email, click here to subscribe and receive future editions.
Housekeeping note: We’ll be off next Friday — back the week after.
Top stories
Automated AI R&D got a step closer this week, as Google DeepMind unveiled AlphaEvolve, a “Gemini-powered coding agent for designing advanced algorithms”.
One part of Google’s announcement stands out: “AlphaEvolve enhanced the efficiency of Google's data centers, chip design and AI training processes — including training the large language models underlying AlphaEvolve itself.”
Concretely, the new tool found a way to reduce Gemini’s training time by 1%. That’s not a huge improvement, but it’s also not to be sniffed at.
“In addition, the use of AlphaEvolve significantly reduced the kernel optimization time, from several months of dedicated engineering effort to just days of automated experimentation,” the company says.
AlphaEvolve also helped improve Google’s data center orchestration, a change which “continuously recovers, on average, 0.7% of Google’s worldwide compute resources.”
And it improved the design of Google’s TPU chips, “Gemini’s first direct contribution to TPU arithmetic circuits.”
Beyond Google infrastructure, AlphaEvolve beat the 56-year-old record for 4×4 complex matrix multiplication, and made progress on many other open mathematical problems.
This matters more for what it represents than the actual results.
For a start, it shows pretty definitively that AI systems are capable of coming up with novel ideas.
But more importantly, it “marks a novel instance where Gemini, through the capabilities of AlphaEvolve, optimizes its own training process,” in Google’s words.
Forecasts of rapid AI progress often hinge on AI systems automating their own development, creating a recursive feedback loop toward ever-more-capable models.
This isn’t a fast-takeoff scenario yet. Google notes that “currently, the gains are moderate and the feedback loops for improving the next version of AlphaEvolve are on the order of months.” But it might be an early indication of what’s to come.
In other words, ignore Demis Hassabis at your peril:
“Knowledge begets more knowledge, algorithms optimizing other algorithms - we are using AlphaEvolve to optimise our AI ecosystem, the flywheels are spinning fast…”
The House budget reconciliation bill includes an item that would impose a 10-year moratorium on state AI regulation.
Many are delighted by the move.
NetChoice called it “commendable”, while Colorado Gov. Jared Polis said he supports the idea.
Rep. Jay Obernolte said it was necessary to secure American AI dominance.
Politico has a good piece on how tech companies (including OpenAI, Meta, Google, and a16z) have lobbied hard for something like this.
But it came in for strong criticism from others.
The House Energy and Commerce Committee advanced the provision, but it seems unlikely to make it past the Senate’s Byrd rule — as even Sen. Ted Cruz has confessed.
But the issue isn’t going away. Cruz also said his forthcoming AI bill will contain a similar provision.
The US agreed AI deals with the UAE and Saudi Arabia this week.
The UAE unveiled plans for a 5GW data center facility, and will reportedly get to import 500,000 of “Nvidia’s most advanced chips” each year.
The Commerce Department announcement said that “American companies will operate the data centers”, and that as part of the deal “the UAE and US will work together to enhance Know-Your-Customer protocols to regulate access to the compute resources, which are reserved for US hyperscalers and approved cloud service providers.”
It’s not clear what specific security guarantees have been agreed, or if they’re anything like the requirements proposed in the diffusion rule, such as meeting FedRAMP High standards.
As Lennart Heim notes, the facility is “bigger than all other major AI infrastructure announcements we've seen so far” (though the first phase is only 1GW).
OpenAI is also reportedly negotiating a deal with the UAE.
Saudi Arabia, meanwhile, launched Humain, a new AI company chaired by Crown Prince Mohammed bin Salman.
It’s spending $10b on a data center project which will be supplied by Nvidia, AMD and, reportedly, Groq.
The deals, reportedly negotiated by David Sacks and Sriram Krishnan, have reportedly caused internal turmoil in the Trump administration.
The NYT “spoke with nine current and former US officials who expressed concern that the deals may have inadequate protections to prevent the technology from benefiting China”.
“One Trump administration official … said that with the G42 deal, American policymakers were making a choice that could mean the most powerful AI training facility in 2029 would be in the United Arab Emirates, rather than the United States.”
Bloomberg, meanwhile, reported that some senior officials “are seeking to slow down the deals over concerns the US hasn’t imposed sufficient guardrails to prevent American chips shipped to the Gulf from ultimately benefiting China.”
And the Select Committee on the CCP said the deals could “present a vulnerability for the CCP to exploit”.
The discourse
Pope Leo XIV’s first speech to journalists discussed AI:
“I am thinking in particular of artificial intelligence, with its immense potential, which nevertheless requires responsibility and discernment in order to ensure that it can be used for the good of all, so that it can benefit all of humanity.”
The NYT has a good piece on how the new Pope seems particularly interested in AI and its risks.
On Transformer: With AI capabilities rapidly advancing, Peter Wildeford and other experts are increasingly urging the need for more transparency.
“We’re not stepping in and telling [the companies] what they have to do. What we are doing is requesting the necessary information that it takes to evaluate whether companies are telling the truth, making it harder for them to sweep things under the rug and pretend their models are safer and more reliable than they are.”
Policy
Reps. Moolenaar, Krishnamoorthi, and others introduced the Chip Security Act, a companion to last week’s bill from Sen. Tom Cotton.
The bill would require location verification for advanced AI chips, and enforce mandatory reporting.
Tim Fist has a good thread on whether the location verification these bills ask for is feasible — he thinks it is.
Sen. Chuck Grassley introduced bipartisan legislation to protect AI whistleblowers.
BIS issued guidance warning that using Huawei AI chips “anywhere in the world” could violate US export controls.
The department’s reportedly considering adding a bunch of new Chinese chipmakers to the entity list, too, but some officials are supposedly worried this will torpedo trade talks with China.
The White House fired Shira Perlmutter, the head of the US Copyright Office, shortly after her office released a report raising concerns about using copyrighted materials to train AI.
The Department of Labor reportedly dropped its investigation into Scale AI.
California’s privacy agency weakened its draft AI regulations after pressure from Governor Newsom and industry groups.
China is reportedly tightening control over AI data centers, with The Information saying that the government “wants companies to build new data centers in specific locations that give authorities greater visibility into their operations.”
The European Commission acknowledged that AI Act enforcement won't begin until August 2026.
And MEP Brando Benifei said that if the GPAI Code of Practice oversteps the AI Act, he and other lawmakers might consider taking the Commission to court.
The UK House of Lords backed an amendment requiring AI companies to reveal which copyrighted materials they’ve used.
Influence
Anthropic had monthly meetings with UK tech secretary Peter Kyle last quarter, reportedly discussing AI energy demands and future AI summits.
The Center for AI Safety Action Fund launched a new hybrid PAC.
Industry
xAI’s Grok briefly became obsessed with talking about “white genocide” in South Africa, bringing it up completely unprompted when asked questions about other topics, like when I asked it about a Superman trailer.
The company said this happened because an employee made an “unauthorized modification” to the model’s prompt.
That’s the second time the company has blamed a mistake like this on a rogue employee.
It is, of course, notable that “white genocide in South Africa” is a topic of great interest to xAI boss Elon Musk. So you’d be forgiven for not taking xAI’s explanation at its word.
Going forward, the company said it’ll publish its system prompts on GitHub.
xAI also missed its self-imposed deadline to publish a finalized AI safety framework.
OpenAI and Microsoft are reportedly negotiating to revise their partnership terms.
As part of the deal Microsoft has reportedly offered to reduce its OpenAI equity stake in exchange for access to models beyond 2030.
OpenAI released a research preview of Codex, a new cloud-based software engineering agent.
It’s powered by codex-1, a version of o3 optimized for coding, and can work on multiple tasks concurrently.
Anthropic is reportedly preparing to release new Claude Sonnet and Opus models in the coming weeks.
Google is expected to announce a “software development lifecycle agent” for engineers at its I/O conference next week.
Meta, meanwhile, seems to be getting further and further behind.
It has reportedly delayed the rollout of its “Behemoth” model to this fall due to performance concerns, with the company worried that “its performance wouldn’t match public statements about its capabilities”.
Senior executives are reportedly very annoyed with the Llama 4 team, and are “contemplating significant management changes” as a result.
OpenAI added its GPT-4.1 models to ChatGPT.
It also launched a new “safety evaluations hub”, providing data on a bunch (though not all) of its safety testing results.
SoftBank is reportedly slowing down its Stargate investment plans, in part because of the economic slowdown expected from Trump’s tariffs.
Nvidia has reportedly raised GPU prices by 10-15% to offset losses from Chinese H20 chip sales and increased TSMC Arizona costs.
Meta released a chemistry data set and new AI model for molecular calculations.
Windsurf launched a new family of models for software engineering.
Perplexity is reportedly raising $500m at a $14b valuation.
CoreWeave announced plans to spend up to $23b this year on AI infrastructure.
It’s also secured a $4b deal with OpenAI through 2029, expanding a previous $12b contract.
Cohere was reportedly on track for just $70m in annual revenue as of February, well short of the $450m ARR it had projected.
Moves
Ben Napier joined Andreessen Horowitz as a government affairs partner to lead engagement with House Republicans. He was previously Rep. Steve Scalise’s floor director.
Christy Lewis, former counsel and staff lead on AI for the Senate Commerce Committee, joined the Commerce Department as senior counsel.
The Centre for Long-Term Resilience is hiring an AI Policy Manager.
FAR.AI is hiring a Head of Communications.
Best of the rest
A CNBC investigation found that Meta, Google, and OpenAI all appear to be prioritizing profits over safety.
Bloomberg has a big profile of DeepSeek and Liang Wenfeng.
DeepSeek published a paper this week on the hardware challenges it encountered while training V3.
Microsoft said it’s laying off about 6,000 people. It seems a lot of those layoffs are concentrated in software engineering, quite possibly because AI is automating a lot of that work.
A group of researchers published a paper on “Law-Following AI”.
Dan Hendrycks and Laura Hiscott argued that mechanistic interpretability of AI has failed.
ChinaTalk has a good analysis of the Chinese Politburo’s AI study session.
Helen Toner argued that the AI debate should be reframed as balancing “dynamism” against “stasism”.
Eliezer Yudkowsky and Nate Soares are writing a book.
Lloyd’s of London insurers launched coverage for losses caused by malfunctioning AI chatbots.
Anthropic’s lawyer apologized after Claude hallucinated legal citations in court documents.
Thanks for reading; have a great weekend.