Can Alexandr Wang save Meta?

Transformer Weekly: More Republican moratorium opposition, the RAISE Act passes, and Altman’s ‘gentle singularity’.

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: this week’s edition is slightly abridged as I’m in the middle of a bunch of travel.

Top stories

• Meta bought a 49% stake in Scale AI for $14.3b. As part of the deal, Alexandr Wang is leaving Scale for Meta, where he’ll work on a new “superintelligence” team (presumably in a leadership role).

  • This isn’t a typical acquihire: only “around four Scale employees” are expected to join Meta, Platformer reported. Jason Droege will become interim CEO of Scale.

    • It’s not clear what will happen to Scale: while the company said it will continue serving other customers, Platformer reported that “Scale expects [DeepMind and OpenAI] will stop using Scale's services”.

  • This is all an attempt by Mark Zuckerberg to salvage the company’s AI ambitions, after years of chaos and disappointing products.

    • The deal gets Meta access to Wang’s deep knowledge of how AI companies are training their systems — as well as injecting some much-needed energy into Meta’s flailing efforts.

  • Zuckerberg has reportedly spent recent months recruiting for the new team, with Google DeepMind’s Jack Rae and Sesame AI’s Johan Schalkwyk also reported to be joining. He reportedly tried to hire Google’s Koray Kavukcuoglu and former OpenAI CTO Mira Murati.

  • The NYT reported that Meta has offered “seven- to nine-figure compensation packages to dozens of researchers”, with Zuckerberg personally recruiting for the new team.

    • Platformer’s Casey Newton has the best analysis I’ve seen of what’s going on here: “In some real sense Meta's future depends on hiring and retaining those people, and Meta has already committed $65 billion to AI infrastructure this year. What's a few hundred million more?”

  • One potential wrinkle: “Some officials in Donald Trump’s administration were upset by reports about the deal and have privately called it an issue of national security due to Scale’s contracts with the Department of Defense”, The Information reported.

The discourse

• Sam Altman published his vision for a “gentle singularity”.

  • “We are past the event horizon; the takeoff has started. Humanity is close to building digital superintelligence.”

  • “Advanced AI is interesting for many reasons, but perhaps nothing is quite as significant as the fact that we can use it to do faster AI research.”

  • He also said that an average ChatGPT query uses about "one fifteenth of a teaspoon" of water and 0.34 watt-hours of energy.

• Zach Stein-Perlman thinks “AI companies' eval reports mostly don't support their claims”:

  • “The companies usually don't explain why they think the results, which often seem strong, actually indicate safety, especially for biothreat and cyber capabilities.”

Policy

• Republican opposition to the state AI rules moratorium grew, with Sens. Ron Johnson, Rick Scott and Angela Paxton all now opposing the move.

  • With Sens. Hawley and Blackburn also opposed, the GOP no longer has a Senate majority for the provision.

  • The House Freedom Caucus also came out against the provision, as did Gov. Ron DeSantis. And even JD Vance acknowledged concerns about it.

  • Others, meanwhile, oppose the provision for making it harder for states to decide where data centers should be built.

  • Sen. Ed Markey is introducing an amendment to scrap it, while Senate Commerce Dems. demanded a committee vote on the moratorium.

• The leaders of the House China committee introduced the Advanced AI Security Readiness Act.

  • It aims to “empower”’ the NSA “to detect and neutralize threats targeting advanced AI technology”, among other things.

• Lots of details in the US-UAE AI deal are still yet to be finalized, Reuters reported, including security conditions and how the agreement will be enforced.

• The RAISE Act passed the New York state legislature, and now goes to Gov. Kathy Hochul.

  • Lobbying against the bill is ramping up in an effort to get it vetoed.

• The New York assembly passed the AI Training Data Transparency Act.

• The final draft for the EU GPAI code of practice is reportedly expected by the first week of July.

  • EU digital chief Henna Virkkunen didn't "rule out" postponing parts of the AI Act, though.

• The European Commission launched a stakeholder consultation on classifying high-risk AI systems under the AI Act.

• The UK AI Bill is reportedly delayed by at least a year. The bill will now address copyright issues alongside safety concerns, The Guardian says.

  • The UK Data (Use and Access) Bill passed without the controversial AI copyright amendment.

• Keir Starmer pledged an extra £1b for UK AI infrastructure.

Influence

• More than 20 tech safety advocacy groups urged Congress to pass the bipartisan AI Whistleblower Protection Act.

• The tech industry, led by the Data Center Coalition, is lobbying to preserve clean-energy tax credits, citing concerns about power shortages for AI development.

• Policy Exchange released a report urging the UK government to restructure its institutions for superintelligent AI.

Industry

• OpenAI said it’s reached $10b in annual recurring revenue, up from $5.5b last year.

  • It’s reportedly discussed raising funds from Saudi Arabia's PIF and India's Reliance Industries as part of its $40b funding round.

  • Employees have reportedly cashed out nearly $3b in shares since 2021, with SoftBank purchasing about half of that.

• OpenAI is reportedly buying some cloud capacity from Google.

  • Some of that will reportedly come from CoreWeave, resold by Google.

• OpenAI launched o3-pro, and cut o3 prices by 80%.

  • OpenAI didn’t release an updated system card for o3-pro, prompting criticism from Miles Brundage.

• Huawei CEO Ren Zhengfei said the company’s chips are one generation behind US competitors, arguing that the US has “exaggerated Huawei's achievements.”

  • Commerce Under Secretary Jeffrey Kessler said this week that Huawei will produce only 200,000 AI chips this year.

• Microsoft plans to rank the safety of AI models on its platform.

• Apple announced a way for developers to integrate its on-device AI models into their apps.

• Oracle said the company is seeing "almost insatiable" demand, driven by AI.

• Qualcomm bought Alphawave for $2.4b.

• Crusoe plans to purchase $400m worth of AMD AI chips.

• Mistral said it’s building a 40MW data center near Paris, with the aim of growing to 100MW in the next 18 months.

  • Mistral also released two new reasoning models. Benchmarks show they underperform the competition.

Moves

• Google appointed Koray Kavukcuoglu as its new chief AI architect. Meta had reportedly tried to poach him.

• Richard Fontaine joined Anthropic’s Long-Term Benefit Trust.

• Richard Johnson joined OpenAI as its national security risk mitigation lead.

• Kevin McKinley joined Andreessen Horowitz to lead state-level policy efforts.

• Americans for Responsible Innovation announced a bunch of hires: Brandie Nonnecke as senior policy director, Rachel Hovde as director of policy, Sarah Kessel as manager of government affairs, and Evan Sarnor as a legislative analyst.

• OpenAI chief product officer Kevin Weil is joining Detachment 201, a new Army Reserve program focused on tech. Bob McGrew, formerly of OpenAI, is also joining the program.

Best of the rest

• Disney and NBCUniversal sued Midjourney for copyright infringement.

• The Getty-Stability lawsuit kicked off in the UK.

• Scientific American has a profile of Epoch AI’s FrontierMath event, at which top mathematicians met to test o4-mini. The researchers seem to have been very impressed.

• Wired profiled the "Worthy Successor" symposium, where researchers gathered to discuss creating superintelligent AI that would replace humanity.

• Users are accidentally publishing sensitive information in Meta AI's "discover" feed.

• Meta launched a new AI "world model".

• The Schwartz Reisman Institute published a report on "regulatory markets" for AI governance.

• Alibaba, Tencent, and Moonshot disabled AI photo recognition features during China's annual "gaokao" college entrance exams to prevent cheating.

• Politico's AI tool for subscribers generated fabricated information.

• Mattel has partnered with OpenAI to develop AI-powered toys and games.

Thanks for reading; have a great weekend.