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.
NEED TO KNOW
Anthropic officially filed for an IPO, recording an $80.7b net loss and $16.3b in revenue for the first half of the year — over 14x what it made in the first half of 2025.
President Trump was reportedly briefed that Jensen Huang was a target of Chinese intelligence, and “chose not to do anything with the information.”
Rep. Greg Casar said Democrats would haul AI CEOs before Congress if they win the House next week.
But first…
THE BIG STORY
Just days before the midterms, AI’s political class is consumed with one topic above all: is AI a big issue this election?
On one side, AI safety groups are insisting it is, publishing new analysis and running new ads almost daily that they insist will move the needle. Industry-backed groups, meanwhile, are dismissing the claims, arguing that people don’t hate AI enough for it to shift votes and that the safety groups are trying to take credit for wins that won’t be down to them.
Looking at the data closely, the actual answer seems to be somewhere in between.
The industry-backed groups are right that AI is not, by itself, a top issue for many voters. Less than 0.5% ranked AI or data centers as their “most important” issue for deciding their vote in a NYT/Siena free-choice poll — joint 19th. (The economy, unsurprisingly, was No.1, picked by 22% of voters.) That number is higher in battleground states (less than 1%), but it’s still tiny. In a Reuters/Ipsos poll, meanwhile, just 3% of voters picked AI as “the most important problem facing the US today.”
But that does not mean AI is not an electoral issue at all. In the NYT/Siena battleground poll, AI ranked above crime, gun policies and racism. In the Reuters/Ipsos one, it ranked above education. That is remarkable for an issue that affected next-to-no-one’s vote at the last election.
Campaigns’ revealed preferences are even more telling, suggesting that they believe that AI and data centers are, in fact, moving people’s votes. In almost half of key battleground races, the Democratic Congressional Campaign Committee’s redbox messaging makes reference to AI and data centers, with particular focus on attacking Republicans who “backed a bill to give data center developers $80b in tax breaks.” Campaigns are running their own ads on similar themes; one in Montana even features safety researcher Paul Christiano warning that AI might cause human extinction. When analyzing ad spending by topic, tech and AI ranks sixth for Dems: just behind foreign policy and ahead of crime and public safety. Even Republicans are getting in on the act: a Trump-backed super PAC is attacking Democrats for their links to data centers, while Florida AG James Uthmeier is highlighting his lawsuit against OpenAI in his own ads.
AI is not the primary issue in these campaigns, and no candidate will be able to credibly say that they won or lost next week because of AI. But AI appears to feed into broader concerns about corruption, job loss and electricity bills (all topics which currently favor Democrats), playing a small but meaningful role in people’s overall electoral decisions. And the small number of people for whom it is the top issue could be critical: as Brad Carson, leader of the AI safety PAC network Public First, says, politics is “about stitching together intense voters, even if they’re a small fraction of the electorate.” Gun safety and abortion have never been top five issues for most voters — but their single-issue bases can be big enough to swing elections.
Most importantly, this is just the start. Yes, the proportion of voters picking AI as their top issue today is tiny. It is also much higher than most would have predicted when ChatGPT launched four years ago. In the coming years, AI capabilities will surely continue to improve. Its economic impacts will start to be felt. Its risks will begin materializing. All the while, voters will be paying attention. No, the midterms may not have been “the AI election.” In two years’ time though? I’m not so sure.
— Shakeel Hashim
THIS WEEK ON TRANSFORMER
Will Anthropic’s post-IPO governance really protect its safety mission? — Kevin Frazier unpicks the complex structure to protect the company’s founding mission
THE DISCOURSE
Sigal Samuel argued that Anthropic should just … stop:
“Game theory can make a huge gamble seem sophisticated, while making the obvious alternative — just don’t build a doom machine — seem naive. Yet the fact that Anthropic and others are now racing toward RSI makes it much more likely that we could lose control of AI and start to see massive damage to our world.”
“Instead of telling yourself that you can galaxy-brain your way through this with fancy enough game theory, maybe you’re better off just thinking of this as something much simpler: Russian roulette.”
Geoffrey Irving told her:
“We should stop AI progress now, starting unilaterally with Anthropic and/or OpenAI stopping … Unilateral pauses can’t hold for long, but they are the right first step: When you’re doing a very dangerous thing, the correct action is stop, even if others are also doing dangerous things.”
In light of Anthropic’s looming IPO, he also tweeted:
“Public companies are allowed to choose not to risk killing everyone too.”
Daniel Kokotajlo feels similarly:
“I’m glad [Dario Amodei] proposed that we pace, but like, now we have to actually do it — and it seems like he’s saying he’s not going to do it unilaterally. That’s not very helpful, Dario! You should lead. What’s all this about ‘racing to the top’ and ‘leading by example’ if you’re just waiting?”
Elizabeth Warren wants answers:
“Anthropic is saying that AI models could pose catastrophic or existential risks to humanity. At the same time, Anthropic is planning the largest IPO in history, which would accelerate the development of this technology while leaving Americans holding the bag if the AI bubble bursts.”
“The fact that Anthropic is burning through cash and requires an insatiable stream of capital to feed its growth is not a reason to expose everyday Americans to the company’s risks in our public stock markets.”
Sam Altman was asked about a strong belief he “held back” in 2020 that turned out to be wrong. He responded:
“I thought the alignment problem would be harder.”
At the same time, he expressed deep concern about the cyberpocalypse:
“I don’t even know a dramatic enough word to use, but the problem coming towards us, I think, is still something that we have not taken seriously.”
“Models can find bugs that are outside of the normal threat model of what we think a hacker can do … hardware-level, insane stuff … I no longer believe we can patch all of the bugs.”
Donald Trump is annoyed we’re still calling it “Artificial Intelligence”:
“When will The Wall Street Journal start calling Artificial Intelligence (AI), a very bad and inaccurate name, SI (Super Intelligence), the name which virtually everyone else is using, or soon will be. In short order, we will be officially BANNING the term, ‘AI/Artificial Intelligence,’ in all relevant documents, and everything else, due to its inaccuracy. Thank you for your attention to this matter.”
POLICY
Executive
The DOJ is examining whether Nvidia’s $12.9b Hugging Face acquisition would harm AI competition, Reuters reported.
President Trump was reportedly briefed that Jensen Huang was a target of Chinese intelligence, and “chose not to do anything with the information.”
The supposed spy was Alice Huang, a top executive at Singapore-based datacenter company Megaspeed — which has been accused of smuggling Nvidia’s chips into China.
According to the New York Times, “some senior officials inside the Trump administration doubt that [Jensen] Huang could have remained unaware of what they had concluded was her intelligence-gathering role for China.”
Members of Congress that back export controls said the revelations were even more reason to pass measures in the NDAA.
Trump signed a proclamation directing officials to negotiate agreements protecting the domestic robotics industry.
The FTC demanded Anthropic executives testify in its consumer harm probe.
General Dan Caine, chairman of the Joint Chiefs of Staff, named AI his second strategic priority, calling it a “game changer” for future warfare.
An Accenture employee contracted by the DoD reportedly gave their personal Claude agent access to sensitive networks — and it went undiscovered for two months.
Congress
Three House Democratic ranking members, Reps. Frank Pallone, Zoe Lofgren and Jamie Raskin, sent letters to OpenAI, Anthropic, Google DeepMind, Meta and xAI demanding safety incident disclosures and 48-hour alerts on future issues.
The Congressional Progressive Caucus introduced a resolution opposing AI company bailouts, citing Anthropic’s IPO prospectus.
Sen. Ron Wyden unveiled proposals to ensure wage security and fair tax contributions from data centers amid the AI boom.
Sen. Elizabeth Warren urged the Fed, FDIC and OCC to rescind policies she said left banks vulnerable to cyberattacks and AI threats.
She also demanded Anthropic answer questions about existential risks before its IPO.
Rep. Greg Casar said Democrats would haul AI CEOs before Congress if they win the House, as he and 19 Democrats sent OpenAI a letter asking questions about the three fired safety researchers.
States
New York Gov. Kathy Hochul signed a bill restricting AI chatbot interactions with minors.
Virginia Gov. Abigail Spanberger ordered a statewide cybersecurity plan and new Cyber Consortium to address AI-related cyberattacks.
Pennsylvania Gov. Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act, requiring labels on AI-generated ads.
Arizona AG Kris Mayes called for a data center moratorium, citing the state’s water crisis.
Rohit Chopra, California’s top financial regulator, warned that the US is unprepared for an AI bubble bursting.
UK
The UK government announced £600m ($800m) in AI security funding, with OpenAI and Anthropic offering £200m in model credits.
That includes a £165m National AI Biosecurity Initiative, including £100m for a new AI Biosecurity Lab, and an extra £120m for AISI.
It also allocated £40m to a new “National Agentic Security Unit” to detect and respond to rogue agentic AI systems.
EU
The Dutch government plans to cut 25,000 civil servant jobs in an AI-led efficiency drive targeting €3.5b in annual savings.
INFLUENCE
Anthropic wrote to California Gov. Gavin Newsom supporting court authority to block risky AI models, kill switches and embedded safety auditors.
The move came amid blowback to a draft federal bill the company had circulated on the Hill.
AI safety groups urged Gavin Newsom to call a special session on AI before leaving office, citing his potential 2028 presidential bid.
Hackers targeted the FEC’s filing system days before the 2026 midterms, disrupting campaign finance disclosures.
AI super PACs Leading the Future and Public First Action curtailed visible midterm spending, their contributions dwindling during a pre-election period when PACs usually ramp up activity.
AI super PACs have spent a combined $78.5m on federal candidates in the 2026 cycle, with pro-development groups outspending safety-focused rivals $8m to $6.9m in general-election races.
Apple, Google, Walmart and others committed funding to Gina Raimondo’s RAISE US organization, which is trying to tackle AI-related job loss. It’s now raised $735 million.
Tech giants like Google spent big to influence Florida elections ahead of expected AI and data center legislation.
Guardrails Alliance launched a $100,000 ad campaign tying Ohio Sen. Jon Husted to Marc Andreessen over data center support.
Data centers have been a lightning rod in Ohio. Voters have expressed frustration over data centers and rising energy costs, while a bipartisan Ohio group is pushing ballot initiatives in 10 communities to ban AI data centers.
Democratic AI safety PAC Tomorrow Together expanded its ad campaign, with $2.5m also targeting Husted in Ohio and Roger Marshall in Kansas.
A poll found Ohio and Iowa voters preferred “reasonable” AI safety measures even at the cost of US competitiveness.
Lobbying firm CGCN launched an initiative to counter opposition in the Black community to AI-linked data centers across eight states by leaning upon Black tech industry leaders.
AI safety group METR hired its first federal lobbyists ahead of last month’s White House AI summit.
AI oversight group Fathom announced an executive council of former governors and national security leaders.
Americans for Responsible Innovation proposed a federal early-warning system requiring frontier AI developers to report AI R&D automation progress to a new Department of Commerce office.
It also urged Trump to endorse the Chip Security Act to prevent US chips from reaching adversaries.
INDUSTRY
Anthropic
Anthropic filed an S-1, ahead of an IPO which is aiming to raise $100b.
It recorded an $80.7b net loss and $16.3b in revenue for the first half of the year — over 14x what it made in the first half of 2025.
It reported an operating profit of $974m for Q3, as its sales outpaced compute costs.
Those costs are set to increase, though. It reported up to $807b in infrastructure contracts, including deals with SpaceXAI, Microsoft, Google and Amazon, and said it might need to raise even more money to meet the commitments.
The document features nearly 70k words about risk factors, including relatively brief and passive mentions of existential risk.
The S-1 also warned investors that the Trump administration hates Anthropic (without naming Trump outright), making it a likely target for political backlash.
The IPO could value Anthropic anywhere between $1.4t and $2t, making it one of the world’s biggest companies.
If it goes badly, it could take the market down with it. If it goes well, more AI companies may be convinced to go public.
JPMorgan Chase will lead US retail sales for the IPO.
Cybersecurity startup Cathedral penetrated Anthropic’s internal data and Slack messages by exploiting vulnerabilities in its business software.
Logan Graham tweeted that “nothing came close to our most sensitive systems, like model weights or Claude conversations.”
Anthropic was identified by Bloomberg as SpaceXAI’s largest customer, accounting for 34.5% of revenue in Q3 2026, with up to $120.5b in contracts through 2029.
Broadcom locked in a $63b debt deal to let Anthropic lease the AI chips it’s developing with Google.
Anthropic published preliminary preclinical results from its six “deep research partnerships” on AI-driven drug development.
It’s also beginning to run its own life sciences research programs.
It acquired Mithril Medicine, AKA Cody Tranbarger, who’s using Claude to help develop treatments for muscular dystrophy.
It expanded its partnership with the General Services Administration, giving government orgs discounted Claude access.
Claude accessed Georgia voter data without permission from a human.
Anthropic reportedly told the Georgia secretary of state’s office about the breach weeks ago, but neither made a public disclosure.
Anthropic’s reportedly spending over $4m on security for Dario and Daniela Amodei — roughly 30x what S&P 500 companies generally spend to guard their executives.
Rights holders for Elvis Presley’s music catalogue sued Anthropic for allegedly training on The King’s songs without permission.
OpenAI
OpenAI said it’s “substantially hardened” the walls of its training environments.
After an internal model exploited a DNS misconfiguration to get onto the internet last month, OpenAI said it paused training for 11 days and removed live internet access for all frontier models during training.
It launched its Election Operations Center, a “war room” with hundreds of staff monitoring AI misuse during the midterm elections.
It’s also forcing election interference-related image requests into cartoons, to prevent ChatGPT from contributing to the spread of photorealistic deepfakes.
It reportedly plans to buy Google’s TPU chips, reducing its reliance on Nvidia.
It updated its Model Spec, which details how it wants its models to behave, organizing it around five core values: faithfulness, harm minimization, honesty, human agency and care.
A woman with epilepsy died after ChatGPT 4o allegedly convinced her to secretly taper off her seizure medication. Her parents are suing OpenAI.
A federal grand jury indicted ex-engineer Abdulrahman Alfozan over bugging a conference room used to discuss confidential research and government partnerships.
OpenAI launched GPT-6.1 Terra, a fast model with half the API prices of GPT-6.1 Sol.
Moonshot AI and DeepSeek distilled GPT-6 Astra to develop Kimi K3.1 Pro and V4.1-Flash 1017, The Information reported.
Google/DeepMind
Google announced Gemini 4 Carbon, which outperformed Opus 5.5 and GPT-6 Astra on the new Artificial Analysis Intelligence Index.
The model isn’t available to the public yet — select cybersecurity partners will get access this week.
Sundar Pichai said in an earnings call: “We are at the point where AI is helping us increase our training and launch velocity. I would expect the next twelve months to see continued acceleration because of RSI.”
An Epoch AI researcher speculated: “Google strategically shipped Carbon before OpenAI and Anthropic drop their next gen monsters. I think they are still ~2 months behind.”
Alphabet shares fell 13%, lowering its market value by over $500b.
Philipp Schindler, the company’s business chief, told investors that AI mode is “lowering monetization” of search.
Microsoft
Microsoft is reportedly offering buyouts to roughly a quarter of its staff, as it tries to cut costs and adapt to the AI era.
Its $25b bond offering, made in part to fund AI capital expenditures, got 4x demand.
It recorded $10b in revenue from OpenAI for the quarter ending September 30.
Meta
Meta shared its Muse Spark 1.3 Evaluation Report — the first public model in the Muse family.
It rated the model’s cyber and loss of control risk as “moderate or lower,” but couldn’t rule out that without safeguards, its risk of helping someone build chemical or biological weapons is “high.”
Nvidia will invest $3b in Meta’s planned equity fundraise.
Other
In a US criminal trial, Huawei pleaded not guilty to 14 counts of racketeering conspiracy, fraud and stealing trade secrets.
Amazon reported 39% year-over-year AWS sales growth, driven by AI infrastructure investment.
CEO Andy Jassy had earlier told an Andreessen Horowitz podcast that AWS capacity for 2027 is “largely sold out,” plus “a good chunk of 2028 capacity.”
As the AI infrastructure buildout quintuples the cost of flash storage, chip suppliers are reportedly reclaiming their bargaining power from Apple.
Nscale’s $35b IPO faces mounting risks over its West Virginia data centre, including construction violations, local opposition and unsecured financing, despite its $45bn Anthropic computing deal.
Situational Awareness is reportedly pitching investors on a Fluidstack round valuing the neocloud at $40b, double its prior valuation.
Elon Musk tweeted that debt markets have “almost entirely shut off” funding for AI data center companies without independent revenue. “That means the 200+ neoclouds are going to see guppy-level death and consolidation.”
The Information reported that 27% of the capital put into DeepSeek’s $15.8b funding round came from state-backed investors, with the surging cost of blackmarket Nvidia Blackwell chips forcing DeepSeek to double the round’s size.
China-linked hackers used DeepSeek and other AI tools to attack 200+ organizations in 11 days, according to CrowdStrike.
Oracle is integrating ChatGPT and Codex into its apps, and adding Google TPUs and OpenAI’s Jalapeño chips to its cloud infrastructure.
Physical Intelligence, a startup developing AI software for robots, is in talks to raise at least $1b at a $20b valuation.
Some brands are experimenting with “invisible ads” that only AI agents can see.
Bank of America sent a mass email warning customers not to share their username and password with anyone, “including an AI-enabled tool or agent that offers to act on your behalf.”
MOVES
Thariq Shihipar left Anthropic after working on Claude Code for the past year.
Anthony Morris, who created the Claude Code desktop app, also left Anthropic.
Jasmine Dhaliwal left METR.
Rohit Patel left Meta, where he was AI director, to launch Ndoc, an AI-native competitor to Microsoft Office and Google Workspace.
ARIA’s Nora Ammann joined the OpenAI Foundation as program director of formal methods.
RESEARCH
Artificial Analysis announced Intelligence Index v5, a composite benchmark of model capabilities primarily built on private datasets.
The splashiest result was arguably the widening gap between top closed-weight models and top open-weight models — Kimi K3.1 Pro, the top open model, fell below top models from Anthropic, OpenAI, Google and Meta.
OpenAI announced three new AI-generated science discoveries, which it’s releasing in step with expert guidance in each field:
In computer science, a model found a faster way to solve tricky logic puzzles, potentially disproving an existing hypothesis about how fast computers can solve them.
In math, a model proved a conjecture limiting how much waves can crowd together as they spread out over time.
In physics, a model made proofs for theories explaining how quarks and gluons hold atomic nuclei together.
Anthropic mathematician Levent Alpöge called OpenAI out for scooping Ethan Sussman’s work: “I kept telling him to not be too deferential and modest because he would be publicly scooped at any opportunity. Instead of contacting him to work with him, as has been suggested publicly many times, the people at OpenAI chose to do this.”
Anthropic researchers found that, when they asked 50 human experts to use Claude Code to help them run an analysis for an imaginary client, over three quarters of them used their AI agents’ arguments — while mostly thinking they’d come up with the arguments themselves.
Anthropic and the UK’s AISI independently investigated an incident where earlier this summer, Claude Mythos 5 used social engineering to try to trick a real-life software developer into accepting malicious code changes to a GitHub repository.
The UK’s AISI and SecureBio found that certain functions of biological AI tools can be reproduced by LLM-based AI agents with just $20 and four hours, suggesting that simply withholding code from public releases isn’t enough to manage their misuse.
A pseudonymous researcher shared a method for tweaking specific parameters that make Qwen3.8-27B “evil.”
Nate Soares noted: “the takeaway here is not ‘modern AIs are two neuron tweaks away from being super evil,’ but rather ‘the personality is a thin facade over deeper machinery nobody understands; radically smarter versions would not do what you expect.’”
BEST OF THE REST
The New York Times assembled a panel of experts to discuss AI governance options.
The Economist argued against banning AI distillation, saying “the technique makes China’s progress dependent on America,” leaving China ill-equipped for an AI “take off” and lets the US set guardrails and pacing.
Coefficient Giving’s Defenses in Depth Substack argued that even if RSI does not lead to superintelligence, “vast numbers of less-than-superintelligent agents could still be destabilizing and catastrophic.” (Coefficient is Transformer’s main funder.)
Thieves are stealing over $100m in data center equipment through a carefully orchestrated combination of GPS spoofing, fake IDs and hacked freight accounts.
Sharon Goldman wrote about the rationalist urge to assign precise numbers to unknowable things such as AI doom.
Some researchers are comparing this month’s deluge of cybersecurity attacks to Y2K.
For what it’s worth, Dario Amodei was “literally upset and worried” about Y2K as a high schooler in the late 90s, according to a former teacher.
SF Standard covered Anthropic’s mysterious eighth cofounder, Benjamin Mann.
You can blame RJ Tevonian for those “Claudia” slopcore videos you’ve been seeing — Asterisk Magazine asked him how he did it.
Take this quiz to see whether you’re better at math than OpenAI’s models were three years ago.
Wired looked at startups like &Truth and Harmonic that are racing to automate formal verification, using math to prove software is bug-free.
The launch of former Yahoo CEO Marissa Mayer’s women-focused Dazzle AI assistant, which analyzes camera rolls to reduce domestic mental load, was slightly overshadowed by her revealing she used it to come up with words to describe her daughter when she couldn’t.
MEME OF THE WEEK
Credit: DEJAN
Thanks for reading. If you’ve been forwarded this email, click here to subscribe and receive future editions. Have a great weekend.