Transformer Weekly: Blackburn’s preemption package, G7 AI talks, and OpenAI’s big new hires
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 #
Sen. Marsha Blackburn is negotiating with theWhite House over a bill package that will tradechild safety anddeepfake protections for some form ofpreemption over state AI laws.AI was a big talking point at the
G7, with calls for** international cooperationand a trusted access scheme**.** OpenAIhired Noam Shazeerto work on new AI architectures, and Dean Ball**as head of strategic futures.
But first…
THE BIG STORY #
It has been a week since Fable was pulled, and there is still much we do not know. We do not know whether the jailbreak Amazon supposedly found is particularly severe — or if it even was a jailbreak. We do not know whether the Trump administration was motivated by national security concerns, disdain for Anthropic, or a combination of the two. We do not know whether the White House’s thinking was shaped by Anthropic reportedly expanding Project Glasswing without the government’s permission.
That uncertainty and chaos may end up being a very good thing. The Fable saga could turn out to be the wake up call governments need — a sign of how bad things can get without clear rules of the road, and an incentive to ensure something like this never happens again.
There are already signs this is happening. Yesterday, Politico reported that Anthropic and the White House are “developing a common set of benchmarks that could be used to assess future jailbreaks, including the extent to which safeguards were bypassed, the capabilities exposed, and the practical consequences of the breach.” A set of defined safety standards, in other words.
Congress, too, appears to be grappling with its lack of action on AI, which has handed all the power to the White House. Members from both parties are now seeking more information, while Rep. Josh Gottheimer — the co-chair of the House Democratic AI commission who has been endorsed by both pro-industry PAC Leading the Future and pro-safety group Public First Action — said he’s preparing legislation requiring mandatory testing of frontier AI models.
And at the G7, foreign leaders reportedly discussed a “trusted partners” scheme to guarantee themselves access to advanced models — having learned in recent weeks the importance of having that access.
The road ahead will not be smooth. It is unclear if the Center for AI Standards and Innovation, the most AI-literate part of government, is involved in drawing up the reported technical standards. It should be. And though Congress may feel more urgency to act, there is still little chance of movement before the midterms.
But the vibe shift, as it were, will hopefully last. Fable has made many realize just how high stakes AI governance is and will be. That can only be a good thing.
— Shakeel Hashim
ALSO NOTABLE #
Washington weighs trading child safety for preemption #
Tennessee Senator Marsha Blackburn’s negotiations with the White House over a bill package that will trade child safety and deepfake protections for some form of preemption over state AI laws appear to have progressed. The the * Daily Signal* reports she will introduce the package “soon.” Blackburn’s office is keeping bill text close to the chest, saying preemption would be narrow — but the range of measures she may be prepared to trade for it is not.
Her landmark child safety bill KOSA would require most online platforms to “exercise reasonable care” to prevent mental health disorders among children, disable addictive features such as infinite scroll for minors, and empower the FTC to enforce such provisions. Her NO FAKES bill, meanwhile, would make an individual’s voice and likeness a federal property right and impose liability for posting unauthorized deepfakes. The App Store Accountability Act, which she reportedly also wants to include, requires stores to verify user ages and impose parental controls for children. Some child safety groups are pushing to add the SCREEN Act or SAFE for Kids Act, which require age verification for porn sites, and the GUARD Act, which restricts access to AI chatbots for minors and imposes liability for mental health-related harms, as well.
Most advocates are hesitating to support or oppose the legislation until they see the full text. Congress is in limbo, too. NO FAKES moved through the Senate Judiciary Committee Thursday, building some momentum, but three Republicans and one Democrat expressed reservations about its impact on freedom of speech. One of those senators is Ted Cruz, who has his own ideas about AI regulation. He reportedly recently asked for Republicans on his committee to submit lists of their AI priorities, and has been involved in the preemption fight before. The House, meanwhile, has its own version of KOSA that removes the standard of care and focuses on parental controls, so it might be hard to get Republicans in the lower chamber on board. And the White House, though engaged in discussions about the package and reportedly meeting with online safety groups yesterday, is yet to indicate its view on KOSA, NO FAKES, or the App Store Accountability Act.
— Veronica Irwin
THIS WEEK ON TRANSFORMER #
—Washington made a frontier AI model disappearShakeel Hashim on the Fable furore—Mythos and Fable can make us all safer. Shutting them down is recklessMozilla CTO Raffi Krikorian argues openness favors defenders over attackers—What’s happened to MAGA’s $100m AI push?Veronica Irwin reports on some strange happenings with Innovation Council—How to fill Congress’s AI knowledge gapChris Stokel-Walker on how the 1970s might hold the key to getting Congress the technical expertise it needs
THE DISCOURSE #
As the Fable saga continues, so does the discourse around it.
Joshua Achiam tweeted: “The thing I’m most concerned about in the ongoing Fable dispute is that it could be the loud noise in the canyon that triggers an avalanche whose outcome is normalizing electronic citizenship verification as a step in using software.”
“The digital universe is a wild west from the perspective of states …The urge to build digital firewalls exists and hasn’t manifested fully. But it could.”
Garrison Lovely speculated: “Now that the stakes have ratcheted up, sr govt people are properly engaging for the first time with things that AI people take for granted — the models are blackboxes, there is no universal defense against jailbreaks — and going ‘what the actual fuck?? this is unacceptable.’ The export controls are the closest thing to pulling the plug that the govt has, so they used it.”
An unnamed source familiar with the Trump administration’s thinking told Axios:
“Anthropic has not done a great job at trying to speak to the administration and appreciate the ideological differences … it’s like they just speak different languages.”
Nat Purser:“not sure the adminrealizeshow bad this makes them look”
**Miles Brundage **has bad news for patriots in AI governance:
“Americans no longer have ‘making fun of Europe’s approach to AI regulation’ privileges …”
**Dean Ball **warned that you can’t stay out of politics while building god:
“[M]aking superintelligence is a profoundly political act even in the healthiest of societies, to say nothing of the filthily political world we Americans currently inhabit. A model like Mythos goes beyond being a mere political act and implicates the sovereignty of the state itself. No company gets to shake the foundation of state sovereignty while staying blithely above the raw reality of politics.”
**Richard Ngo **made a diagnosis: “The AI safety community constructed a memeplex in which ‘taking AGI seriously’ was a prerequisite for being a serious and good person. When inside this memeplex (as many at Anthropic, some at OpenAI, and a few at DeepMind are) … the whole future seems to flow through the ‘one ring’ of controlling recursive self-improvement.”
“[T]he one ring memeplex has an escalating life-cycle. It infects people by letting them harness the narrative that they’re good people for taking AGI seriously, and that making other people take AGI seriously is a boon for the world … then it shuts off their imagination.”
Cal Newport begged Anthropic and OpenAI to stop “doom trolling”: “Like a cat leaving a dead bird at your doorstep, Anthropic catalogs the grim future that its products might produce, shrugs its shoulders and then returns to its furious efforts to make these warnings a reality.”
“If these AI companies insist on continuing to pretend that they’re merely stoic observers of an unavoidable dystopian future, then perhaps it’s time to force the issue.”
(Predictably,
**David Sacks**[endorsed](https://x.com/DavidSacks/status/2067314435685765367)the essay on X.)
**Nate Soares **[said](https://x.com/So8res/status/2067418562952937642?utm_source=substack&utm_medium=email) he told you so:
“AI folk kept telling me that they have to softpedal, because otherwise they’ll piss off people in DC. I replied that people can smell bullshit: AI [company] leaders who believe AI is dangerous should visibly act like it. They didn’t. Now folk in DC are pissed.”
“Would the DC people have been pissed off either way? Maybe. But if they’re going to hate you no matter what you do, you might as well be honest.”
POLICY #
As more details emerged, we got a slightly clearer timeline of the
Fable saga:The
Commerce Department andCAISIreportedlytested Fable before release, andAnthropic was “not told not to deploy.”But soon after it was released,
Amazon researchers got Fable to disclose flaws in some software code. Amazon shared the findings with Anthropic; an independent cybersecurity expertsaidthe findings weren’t a big deal.In a previously-scheduled call with administration officials, Amazon CEO
Andy Jassy shared the concerns, which appear to have freaked the White House out. They asked Amazon to talk to Treasury SecretaryScott Bessent, and the** NSA**wasreportedlyalso read into the discussions.The government then called Anthropic, and reportedly said it had 90 minutes to pull Mythos and Fable for all users.
They couldn’t get hold of
Dario Amodei immediately, andreportedlydidn’t share much information onwhythe administration was so concerned.Anthropic seems to have pushed back, explaining that the findings weren’t a big concern, but
Bessent and national cyber directorSean Cairncross were “unmoved.”
Shortly after,
TrumpapprovedCommerce Secretary** Howard Lutnick**to impose export controls on the model. Amodei reportedly said: “This means we can’t have the model out.” Lutnick replied: “That’s the point.”The whole saga appears to have
elevatedLutnick’s status as an AI policy decision-maker. Since then, Anthropic has sent senior staffers to Washington, who have been engaged in negotiations all week.
On Monday, the group
heldtechnical meetings involvingCommerce,** Cairncrossand CAISI head Chris Fall.**Treasury officials were reportedlynotincluded.As of Thursday, Anthropic and the government were
reportedlyin talks to develop formal technical standards for assessing the severity of jailbreaks.
We also learned that the
White House was already annoyed with Anthropic for expanding access toMythos without telling the government.In particular, the administration was
annoyedthat South Korea’sSK Telecom was given access, claiming the company had ties to China.Anthropic immediately revoked access when the White House asked, but it appears to have left a bad taste in the admin’s mouth.
Notably, some
Project Glasswing membersstillhave access toMythos Preview.
Fable was the
subjectof much conversation at theG7 summit. French President Emmanuel Macron called it “a good thing” the US recognizes AI’s dangers but said the “very strong decision” was “a bad thing … in some regards strictly nationalist.”Indian PM
Narendra Modi expressed concern about the US move and said democratic nations must have access.Dario Amodei,** Sam Altmanand Demis Hassabisallexpressed supportfor some form of international cooperation, withAltman** urging governments not to “cede your responsibilities to AI labs like mine.”G7 leaders reportedly discussed developing a formal
“trusted access” scheme to guarantee they can use Mythos-class models in future.
FY 2027
NDAA drafts are out, and include a few AI-relevant provisions:Both the Senate and House drafts would
make it harderfor the DoD to declare companies a “supply chain risk.”The Senate version
would alsorequire “appropriate levels of human judgment” for lethal force decisions, and more or less ban the use of AI in nuclear weapon launch decisions.
Lutnick reportedlyexpressedconcerns toASML that one of its extreme ultraviolet lithography machines may have illegally reachedChina.ASML reportedly denies the allegations.
The US reportedly
decidednot to addDeepSeek and over 100 other Chinese companies to itsEntity List despite national security concerns, reportedly to avoid escalating tensions with China.The
DODsaidthat over two-thirds of the department has transitioned off** Anthropicmodels to alternative vendors. DOJlawyers stepped into the lawsuit against xAIin Mississippi,arguingthe company is vital for national security and claiming thatGrok** supports classified military operations and was involved in strikes on Iran.The Trump administration is
lettingtheFederal Data Center Enhancement Act expire in September with no replacement, removing energy efficiency and water use requirements for federal data centers.Rep. Josh Gottheimer ispreparinglegislation requiringmandatory testing of frontier AI models for national security risks.A bipartisan group of lawmakers
introducedtheCREATOR Act, which would grant visual artists control over how AI mimics their creative styles.Experts warn the bill’s definition of “distinctive visual characteristics” is ambiguous and potentially unenforceable.
Sen. Bernie Sandersunveiledlegislation proposing a one-time 50% tax on major AI companies’ equity to create a**$7t** sovereign wealth fund.It would apply to
OpenAI andAnthropic, of course, butalsocover things like data centers and AI compute infrastructure, meaning the likes ofAmazon,** Nvidia, Googleand Microsoftcould be included.** It wouldn’t be allowed to sell equity, so it’s unclear how its planned 5% dividends to Americans would be funded.
Reps.Guthrie, Latta, and** Evans**introducedtheRatepayer Protection Act, which would require AI data centers to pay for their own grid upgrades rather than passing costs to local ratepayers.** Sen.Rick ScottintroducedtheGenerative AI Terrorism Risk Assessment Act, requiring federal reports on terrorist use of AI. Sens. Warrenand Blumenthal**introducedtheAI Bubble Transparency Act, which would require the Office of Financial Research to collect data from financial institutions on their AI-related debt and equity exposures.A coalition of
state attorneys general reportedlysubpoenaedOpenAI for documents covering advertising, user data handling, and impact on minors and seniors.Californiaadvancedtwo bills,** SB 813and AB 1405**, to create independent** third-party verification infrastructurefor AI governance. Pennsylvania’s House**passedthe state’s first data center regulations, requiring** AI facilities**to fund grid upgrades, contribute to low-income energy assistance and source 32% of power from in-state renewables.The
Ohio Department of Education and Workforcereleaseda model AI policy addressing ethical use, academic integrity and data privacy.The
UK governmentpublished‘AI Scenarios 2030’, examining five scenarios across three trajectories: a slow down, continued progress, and a take off.
INFLUENCE #
The Guardrails Alliance, a new super PAC that has reportedlyraised$5m to support AI regulation, is positioning itself as a grassroots, tech worker-led counterweight toLeading the Future. Crypto executive
Chris Larsen’s super PAC** You Can Push Backreportedlycontractedwith a company to run social media accounts boosting NY-12 candidateAlex Bores**… without disclosing they’re part of a paid campaign.It is also
supportingstateRep. Manny Rutinel in aColorado Democratic primary, joining manyAI employeesand …Ericand.Wendy Schmidt
As the NY-12 primary campaign enters its final stretch,
Boresmetwith** Hillary Clintonto discuss AI regulation. Micah Lasher**is still the strong favorite to win Tuesday’s primary, having a slight lead in the most recent poll in May, and Bores with only a ~37% chance of winning onpredictionmarkets.
Rep. Sam Liccardosaidhe didn’t want financial support fromLeading the Future, despite accepting its endorsement.Liccardo’s formed his own smaller
Innovation for Good PAC.
Over 200
state lawmakers from 42 statesurgedCongress to opposeReps. Jay Obernolte andLori Trahan’s** Great American AI Act’s**preemption provision.Cybersecurity experts led by
Meta’s former chief security officer** Alex Stamos**urgedthe Trump administration to restore access toFable, arguing the restrictions harm cyber defenders more than attackers.The
AI Policy Networkleda coalition supporting the** MATCH Act**, which would seek to get allied countries to more closely align with US export controls on semiconductor equipment to preventChina from accessing advanced AI chips.Americans for Responsible Innovationreleaseda report calling onNIST andCAISI to develop national safety benchmarks for AI chatbots used by minors.The
American Enterprise Institute and theUrban Institutelauncheda bipartisan commission co-chaired by formerCommerce Secretary Gina Raimondo and formerSpeaker Paul Ryan to assess AI’s impact on jobs and develop policy recommendations.The
Center for Shared AI Prosperity ishiringan executive director to “shape a progressive economic vision for a world with AI-driven job loss.”Trump reportedly
mockedMark Zuckerberg andJeff Bezos for trying to suck up to him after the 2024 election.
INDUSTRY #
OpenAI #
OpenAI
burned$3.7b in Q1 and reported39% gross margins, a slight improvement from last year.Ed Zitron
shared2024 and 2025financial datafrom OpenAI with theFinancial Times, which showed that while itsrevenues overtook the cost of serving customers, it made a huge** loss of $39b in 2025**, largely driven by a**$30b charge** from its conversion to a public benefit corporation late last year.Early employees have reportedlycashed outaround**$14b in shares** ahead of OpenAI’sIPO.** ChatGPTwill reportedlygo liveon the Pentagon’sbespoke AI platform in early July. ChatGPT’s market share**fellbelow 50% as Gemini and (to a lesser extent) Claude gained users.It
announcedtheOpenAI Partner Network, which brings together companies that build and co-sell products using OpenAI’s tech.
SpaceX #
On its first full day of trading, SpaceX
**shares**. It’s now worth around[rose](https://cnbc.com/2026/06/15/spacex-stock-record-ipo-debut.html)20%**$2.5t**.It
[plans](https://ft.com/content/8b73b4ce-3855-4b89-9110-3d892567f28a?syn-25a6b1a6=1)to raise $20b in bonds to repay a bridge loan resulting from merging with**xAI** and X.It
[locked in](https://www.reuters.com/legal/transactional/spacex-buy-anysphere-60-billion-2026-06-16/)its**$60b Cursor acquisition**.It’s
renting outits entireColossus 1 data center toAnthropic. A federal judge
dismissedthe latest version ofxAI’s lawsuit against OpenAI, which accused Sam Altman of stealing trade secrets about Grok from an ex-engineer.
Meta #
Meta’s
Applied AI team is apparently a trainwreck and everyone feels bad.“It’s literally the gulag,” one employee
toldWired.“You have zero purpose in life all of a sudden, you barely interact with anyone, you just have these tasks every week.”CTO Andrew Bosworth thenadmittedMeta’s AI reorganization was “atrocious” in an internal memo, promising stability, communication, and perks.
It
introduced“AI mode” on Facebook, kind of like Google’s AI Overviews … but if the internet was just Meta platforms?** Manus**’ original Chinese investorsplanto buy it back from Meta for $2b after the Chinese government barred the acquisition.Meta president
Dina Powell McCormick is reportedlyexploringraising tens of billions in equity and off-balance-sheet financing from Wall Street and sovereign wealth funds to fund the company’s $600b AI infrastructure push.After famously supporting tokenmaxxing a couple weeks ago, Meta is
now(tokenminning?), imposingtokenminimizingtoken limits on employees.
Microsoft #
Microsoft reportedly might
switchtoDeepSeek V4 as a cheaper alternative to runCopilot Cowork, which is currently powered by pricey Anthropic and OpenAI models.Microsoft
turnedtoAWS for extra compute forGitHub— vibecoders have driven a huge spike in code commits.** ByteDanceis reportedlyspendingover $1b per year onOpenAI models** via Microsoft.
Nvidia #
Kazakhstansigneda**$10b AI infrastructure deal** with Nvidia and startupFirebird to build a computing cluster with 100,000 GPUs.Nvidia’s new
agent harnessenabledcoding agents to** autonomously train physical robots**.It sold$25b in bonds to help fund the AI infrastructure buildout.
Others #
Google is reportedlyencouragingdata centers to adopt its TPU chips by providing financial guarantees, mimickingNvidia‘s strategy, including with a**$3.2b** guarantee for the Lake Mariner cluster in New York state.Apple price increases are “unavoidable” due to AI demand for memory chips quadrupling,CEO Tim CooktoldThe**WSJ.Export controls on Mythos and Fable have reportedly
createdan opportunity forMistral, which has positioned itself around AI sovereignty and open-weight models that customers can deploy independently.** DeepSeek**closeda**$7.4b** funding round at a**$50b+** valuation, with an unusual structure requiring investors to put capital into a limited partnership managed by CEOLiang Wenfeng to ensure he retains complete control — and only China’s national AI fund having voting rights.Midjourneyannouncedplans to build ultrasound-based full-body scanners, aiming for 50,000 worldwide by 2031, as part of a pivot into healthcare.Jeff Bezos joined a**$400m**funding roundfor UK AI materials science startupCuspAI that quadrupled its valuation to $2.6b.World model startup Odyssey
raised$310m at a**$1.45b** valuation, led byNatural Capital with participation fromAmazon,** GVand AMD Ventures**.** Z.ai**released** GLM-5.2**, an MIT-licensed open-weights model that it claims is at the frontier.
MOVES #
Noam Shazeer, a Google DeepMind researcher who co-authored the original transformers paper,joinedOpenAI.He will reportedly focus on finding new AI model architectures.
Dean Ball isjoiningOpenAI as head of strategic futures, where he’ll work on frontier AI policy.He’ll still be a non-resident senior fellow at the Foundation for American Innovation, and his Substack
Hyperdimensionalwillremain“active and independent.” Barret Zophleft** OpenAIagain five months after rejoining in January, having previously departed in 2024 to co-found Thinking Machines Lab. Arun Rao**joined** Nvidia,where he’ll work on Nemotron, its line of open LLMs. Bruce Andrews**, a seasoned lobbyist,joinedNvidia to lead government affairs in DC.Julia Kempe isjoiningAMI Labs as director of research, where she’ll focus on world modeling alongsideYann LeCun.Sol Messingjoined Google DeepMind’sSAMBA (Sociotechnical Analysis of Model Behavior and Alignment) team.** Kevin Rooseisleavingtheto start a new media company alongside his New York TimesHard Forkco-hostCasey Newton.Sharon Goldman leftFortunetolaunchher own outlet,Ground Level AI, on Substack.
RESEARCH #
Researchers at
Oxford and the UK’s AISIfoundthat frontier AI models are** more persuasive**than even elite, coached human debaters with time to prep on the topic of their choosing.The score was roughly AI: 275, humans: 0,
supportingSam Altman’s 2023 prediction:“i expect ai to be capable of superhuman persuasion well before it is superhuman at general intelligence, which may lead to some very strange outcomes.”
Tacit Labs andOpenAIannouncedLifeSciBench, a multimodal benchmark to evaluate how well AI systems can dolife sciences****research, from data analysis to drawing conclusions to science communication.** The Center for AI Safety**foundthat different frontier models show** favoritism**towards different people, places and corporations.Perhaps unsurprisingly, only Grok seems to
favorElon Musk. RAND researchersandGeoffrey Irvingsurveyed23 experts about whether** formal methods**can meaningfully reduce AI cybersecurity risks.One takeaway: Verification methods work for infrastructure, but can’t do much about model-related problems such as jailbreaking.
A team of German researchers
demonstratedthat, at least in a sandboxed electronic health record, an autonomous AI agent matched or outperformed humandiagnostic accuracy across eight conditions, including appendicitis and pneumonia.Loft Orbital’s satelliteranGoogle DeepMind’s Gemma 3 on-device to respond to researchers’ commands to identify areas of interest on its own — an AI-in-space first.Pew Research Centerpublishednew survey results revealing that while about half of US adults use AI chatbots, 63% think AI is advancing too quickly.More people think AI will be bad rather than good for society (and themselves), especially adults under 30.
BEST OF THE REST #
Anton Leicht
arguedthat countries who believe frontier AI is important and don’t trust the US need to stop yapping and start building frontier AI models of their own — but that it’ll cost alotof money.Tech Buzz Chinalaunchedthe China AI Atlas, a database tracking the flow of talent (down to their PhD cohorts) and money across 10 major AI developers.The New YorkercomparedSpaceX’s IPO to the 1873 railroad bubble, warning that “given the tendency of bubbles to last for longer than any detached observer might expect, it is the fate of Cassandras to be largely ignored — until it is too late.”Even Hany Farid, renowned UC Berkeley deepfake expert,
strugglesto distinguish between real and AI-generated videos. He told theNew York Times,“I feel like I’m going blind.”Sundar Pichai didn’t mention AI in his commencement speech, but a group of Stanford grads
walked outanyway (primarily, it seems, over Gaza).Amazon ditchedits nearly finished film about Sam Altman, “Artificial”, shortly after partnering with OpenAI. The movie is now being shopped around to other studios.Allbirds’ AI pivot continued with the company
rebrandingas Smartbird and hiring former Amazon Web Services executive Nadia Carlsten as CEO. Its shares jumped 30%.
MEME OF THE WEEK #
Thanks for reading. Have a great weekend.