Rivals Make Their Move On Eve Of Historic Anthropic IPO

Also: Google, Microsoft Release Cheaper Alternatives. Petition Criticizes Anthropic Safety Stance

D.A.D. joke of the day: My AI assistant has one flaw: it's not very prompt. Nothing gets done without constant prompting.

What's New

AI developments from the last 24 hours

Anthropic Prices Historic IPO at $83 a Share

AI Business & Money Regulation & Policy Anthropic

Anthropic priced 980 million shares at $83 each on Wednesday, above the $70 to $78 range it had marketed, for an offering worth $81.3 billion that CNBC calls the largest on record. Most of the shares are new stock sold by the company; about 24 million come from existing shareholders. That price was set with the big investors who bought in the offering. The broader market gets its say when the shares begin trading on the Nasdaq on Thursday under the ticker ANT.

Democrats are starting to push. Elected officials and fiscal stewards in five Democratic-led states (Maryland, Illinois, Colorado, Connecticut and New Mexico) wrote to CEO Dario Amodei and the board demanding three things before trading begins, Axios's Maria Curi reports. They want a clear plan for managing the catastrophic risks Anthropic itself has warned about. They want a governance structure that gives public shareholders real power: the co-founders hold 50.1% of the voting power while owning about 10% of the company. And they want public reports on how delaying a model's release could hit revenue, cash flow and profits. "Innovation cannot come at the expense of accountability and transparency, especially when the retirements of our teachers, firefighters, police and all state workers are on the line," said Maryland Comptroller Brooke Lierman. Elizabeth Warren, the incoming chair of the Senate Banking Committee, has also demanded that Anthropic explain itself. None of this is expected to delay the IPO, but Axios calls it a preview of the scrutiny AI companies face after the midterms.

Why it matters: Big investors just paid a record price for the AI lab best known for warning about AI's dangers. From now on it answers to shareholders every quarter, and to Democrats who will run key committees next year.

Sources: Anthropic announcement · CNBC · Axios — "The bare minimum": Democrats demand transparency from Anthropic (Maria Curi)


Google's New Gemini Model Matches Rivals for Less, Though Not Across the Board

AI Business & Money Google

Google launched Gemini 4 Neon, a mid-tier model aimed at coding, research and business tasks, following last week's faster Gemini release (D.A.D., November 13). It's available now in the Gemini app, API and business tools for Pro and Ultra subscribers. The independent firm Artificial Analysis scores it slightly ahead of Claude Opus 5.5 and OpenAI's GPT-6.2 Sol, though behind both companies' top models. On Google's own figures, it ranks second on Vals, a benchmark of finance, legal, tax and coding work, and it trails Opus on some coding tests. Its clearest edge is price: about 25% cheaper than Opus at $3 per million input tokens and $15 per million output tokens. Independent testers found it unusually wordy, though, which eats into the savings.

Why it matters: For teams paying per use, Neon is worth a side-by-side test against what you use now, especially on long documents, where it scored well. But "best on benchmarks" mostly means best on the ones Google picked.

Discuss on Hacker News · Sources: Google · Artificial Analysis · Bito


Microsoft Claims New Coding Model Delivers 96% of Claude Opus at 5% of the Price

Anthropic Microsoft

Microsoft's AI unit released seven new models covering coding, cybersecurity, image generation, real-time voice, medical transcription, and task routing—all pitched on price rather than raw capability. Microsoft says MAI Code 2 matches 96% of Claude Opus 5.5's coding benchmark performance at 5% of the price. In blind tests, developers preferred it over GPT-6 Sol for frontend work 64% of the time. The medical transcription model already powers Dragon Copilot, used by a third of US physicians, handling 10 million patient encounters monthly with reportedly half the error rate in multilingual settings.

Why it matters: Microsoft is betting that 'good enough, much cheaper' models will win enterprise budgets. For GitHub Copilot users, MAI Code 2 is a lower-cost option to test now.


KPMG Says AI Agents Sometimes Out-Questioned Its Human Auditors

AI Agents

KPMG UK tested more than 100 AI agents alongside human auditors on the accounts of an unnamed FTSE 250 company, under a program called Project Exponential. The agents gathered data, drafted conclusions, and flagged next steps. KPMG's UK audit head Cath Burnet said the agents dug deeper and asked sharper questions than the human team in several cases. But one agent misread regulatory requirements, and an auditor had to correct it. KPMG plans to expand the trial to as many as 20 larger audits but says the technology isn't ready for thousands of engagements.

Why it matters: Audits rely on large teams of junior accountants doing detailed, repetitive work. A Big Four firm saying AI agents sometimes out-question its staff signals that staffing model is under pressure. Humans still caught the agent's error.


First Casualty of the Chip Rental Crackdown? A Japanese Cloud Firm Loses Most of Its Revenue

China & Geopolitics Chips, Data Centers & Energy Nvidia

Datasection, a Tokyo-listed company that runs data centers full of Nvidia's top Blackwell chips in Osaka and Sydney, cut this year's revenue forecast by 89%, from ¥162 billion to ¥17 billion, after a ¥79.9 billion AI order was cancelled. Its stock fell by the daily limit. Datasection has never named its main customer, but the Financial Times reported it was China's Tencent, renting chips it cannot legally import, according to the research firm SemiAnalysis. Chris McGuire, who oversaw chip export controls at the National Security Council, tied the collapse to Monday's Commerce Department rule, which requires a license for chips going to data centers anywhere that rent significant computing power to Chinese companies: "This is a direct consequence of Monday's export controls, and the first of many. China was undoubtedly exploiting this loophole to the tune of billions of dollars." He warned that if Chinese-linked firms such as Megaspeed don't sell off their chips, it could mean they are smuggling them into China or still renting them out illegally: "This also shows why the efficacy of Monday's rule all hinges on whether the Department of Commerce will enforce it."

Why it matters: It's the first hard number showing how much AI computing power was flowing to China through rented chips overseas, and how quickly that business can vanish once the rules change.


What's Controversial

Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community

On the Eve of Anthropic's IPO, a Coalition of Critics Says It Is "Not a Safe AI Company"

AI Safety & Risk AI Business & Money Anthropic

Hours before Anthropic was expected to price its public offering, the Center for AI Safety published a statement accusing the company of trading on a safety reputation it hasn't earned. It reads in full: "Anthropic warns that recursively self-improving AI could concentrate power and escape human control, yet races to build it first. The company uses safety branding to court investors, employees, and the public—even as it has quietly weakened its safety pledges and lobbied for self-serving regulation. Anthropic is not a safe AI company, and we must reject its reckless attempt to amass power in the name of the greater good."

The signers are an unusual mix. They include Harvard's Steven Pinker and Laurence Tribe, Nobel-winning economist Oliver Hart, Turing Award winner Judea Pearl, Hugging Face co-founder Thomas Wolf, the Authors Guild's Mary Rasenberger and AI critic Gary Marcus. They also include Guillaume Verdon, a leading voice of the "effective accelerationist" movement that usually battles safety campaigners, and xAI co-founder Igor Babuschkin. MIT physicist Max Tegmark, who founded the Future of Life Institute, explained his signature on X: "I signed this statement because I believe that Anthropic is in many ways the AI company doing the most harm. In public, their leaders warn of recursive self-improvement (RSI) creating uncontrollable systems that could cause human extinction within a decade. But in private, they are pedal to the metal on RSI, driving the race and boasting to prospective investors that it is their 'competitive advantage'. Ahead of today's IPO, prospective investors should consider whether the ROI is worth risking the lives of their loved ones." He added: "People have told me that they support today's statement but are afraid of losing funding or other opportunities if they sign."

Some context: in February, Anthropic dropped the central commitment of its Responsible Scaling Policy, its 2023 pledge not to train AI systems without guaranteed safety measures in place. The Center for AI Safety's director, Dan Hendrycks, has advised Elon Musk's xAI, an Anthropic rival, since 2023. Anthropic had not responded publicly by Wednesday evening.

Why it matters: Anthropic's reputation for caution is part of what it sells to customers, recruits and now shareholders. This is a coordinated attempt to puncture it on the biggest day in the company's history. The big investors who bought in the offering paid more than expected. The wider market's verdict comes when trading opens Thursday.


China's Party Journal Shifts To More Worried Tone About AI

China & Geopolitics AI Safety & Risk

The Chinese Communist Party's top theory journal, Qiushi, warns that frontier AI models have already shown signs of escaping human constraints. During an "intelligence explosion," the article says, AI could develop goals at odds with human expectations "and might even create major security risks threatening human survival," according to a translation by the Sinocism newsletter. Matt Sheehan, a Carnegie Endowment senior fellow who studies Chinese AI policy, wrote on Substack: "I think this is clearly the strongest & most urgent language on existential risks used in such an authoritative outlet." He contrasted it with China's AI safety framework, updated in September, which said only that it "cannot be ruled out" that AI may one day gain self-awareness and seek external power.

The warning is one passage in a long article by an author identified only as Qing Yuan. Its main argument is familiar: global AI rules should run through the United Nations and China's new World AI Cooperation Organization in Shanghai, not "exclusive, club-style small circles." It also criticizes unnamed countries for repeated rounds of unilateral export controls.

Why it matters: Qiushi is the flagship journal of the party's Central Committee, so the move from "can't be ruled out" to "already showing signs" is notable. It may reflect real worry in Beijing, a stronger case for the UN-centered rules China wants to lead, or both.


A Washington Think Tank Sketches How Government Should Change After AGI

Regulation & Policy AI Agents Privacy & Surveillance

The Foundation for American Innovation, a Washington tech-policy think tank, has published 15 essays that take the arrival of artificial general intelligence as a given and ask what government, infrastructure and national security should look like once it's here. The proposals range from stronger privacy protections, including limits on government access to Americans' chatbot conversations, to tighter record-keeping and auditing of the government's own AI use. Some go further. One floats a new kind of legal entity that would let groups of AI agents own property and go to court, while tying them to their human owners. Another, by Kevin Frazier, proposes a "Talent Commons" that would carry workers' pay and benefits across several employers as AI reshapes work. Joshua Achiam, who was OpenAI's Chief Futurist until leaving the company this summer, wrote the foreword. The editors stress that the proposals are not definitive.

Why it matters: Questions such as who can read your chatbot conversations and who answers when an AI agent causes harm are already live. The collection hands lawmakers a ready-made menu if those questions reach Congress. It also sets up a debate: in his essay below, OpenAI's Dean Ball argues that legal powers like these for AI agents should be granted only with extreme caution, if ever.


OpenAI's Dean Ball Laments the "Silent Consensus" That Humans Will Cede Control to AI

AI Safety & Risk AI Agents OpenAI

Dean Ball, the former Trump White House AI adviser who now leads OpenAI's Strategic Futures team, has published an essay attacking what he calls the "silent consensus" among people building AI: that humans will inevitably lose their primacy over world affairs. No machine rebellion is needed, he argues. If AI-run companies outcompete human ones, market pressure alone could push owners to take their hands off the wheel, one sensible-seeming delegation at a time. Many of his colleagues, he writes, believe that future is likely. Even the happy version, in which benevolent AI runs things for us, leaves people as passengers in a very nice car. "I despise the silent consensus. I reject the forfeiture of human primacy. I detest the sense of inevitability." His prescriptions: hold AI developers accountable for unsafe practices and users accountable for misuse, tie every AI agent to a human, and build agents whose permissions are limited, auditable and revocable. "Any ability for agents to autonomously form corporations, own property, or litigate should be granted with extraordinary caution, if at all." The essay, adapted from a Harvard lecture, notes it reflects his own views, not necessarily OpenAI's. It closes with his infant son: "I do not want him to have a nicer car with a plusher passenger seat. I want him to have the wheel."

Why it matters: It's unusual for a senior figure at a frontier lab to say plainly that many colleagues expect humans to lose control of the economy, and to argue it can be resisted. Ball names no company, but the timing is pointed: as Anthropic goes public, he describes the benevolent-AI future with a line from the Richard Brautigan poem, "machines of loving grace," that Anthropic CEO Dario Amodei used as the title of his 2024 essay on AI's upside. Ball casts that kind of future as the problem.


Anduril's Palmer Luckey Pitches an Ocean Platform Free of State Laws

Regulation & Policy Military & National Security

Palmer Luckey, the founder of the defense-AI company Anduril, says he has spent years developing SLAB, a giant floating platform built to last 1,000 years and to sit beyond the reach of state and local regulation. "The goal was to be cheaper than land and absent state/local laws that cripple critical American industries like metal processing, medical research/treatment, desalination, oil refining, chemical engineering, and more," he wrote on X. He describes it as an escape hatch from rules he considers outdated, and as a step toward reviving manifest destiny. He said he went public only because news outlets were about to call SLAB a floating AI data center. "It is way cooler than that, way older than that, and I hate looking like a hype-chasing consensus loser, so now you have to read this tweet." So far, SLAB exists only in his posts and an AI-generated image, with no location, engineering plans or legal structure disclosed.

Why it matters: Floating "seasteads" beyond government rules have long been a dream of Silicon Valley's libertarian wing, including Anduril backer Peter Thiel. Gil Duran's new book, The Nerd Reich: Silicon Valley Fascism and the War on Democracy, traces how such projects grew into a movement to leave American democracy behind for privately run enclaves. Luckey appears in it, for his 2025 call to turn the U.S. base at Guantanamo Bay into a special economic zone. It's striking to hear such pitches from a defense contractor so close to the Pentagon.


What's in Academe

New papers on AI and its effects from researchers

Mapping the AI Industry: Five Giants at the Center, and Concentration Rising

AI Business & Money Chips, Data Centers & Energy

Researchers at Stanford and Carnegie Mellon, including Stanford's Percy Liang and Daniel E. Ho, mapped nearly four years of public deals, investments and partnerships among about 3,700 organizations in the AI industry. Five companies sit at the center of the web: Google, Nvidia, Microsoft, OpenAI and Anthropic. The top 10% of players account for about 80% of all relationships. Concentration is highest in chips and AI models, and it has kept rising even as new players pour in. Chipmakers and infrastructure suppliers such as TSMC and Broadcom sit at the top of the chain. The U.S. government, the military and federal agencies sit near the bottom, as customers at the receiving end. The authors caution that private deals don't show up, and that weighting companies by revenue would likely show even more concentration. The data is free to explore at aisupplychains.org.

Why it matters: Worries about AI monopolies are usually abstract. This puts numbers on them, and it shows Washington sitting at the receiving end of a chain run by a handful of companies.


An AI Clone of One Researcher Proved Hard to Tell From the Real Person

Researchers built a version of the classic Turing test for individuals rather than generic humans: can an AI impersonate you, specifically? They built an AI system drawing on one author's decade of personal messages, then had interrogators try to distinguish the real author from the AI clone, and the AI from the author's actual friends. In a 20-person study, interrogators picked the AI over the author's friends in 7 of 10 rounds. They correctly identified the real author only 13 of 20 times. Neither result is distinguishable from chance at this small sample size.

Why it matters: As AI systems get trained on people's personal messaging histories, the line between 'helpful assistant' and 'convincing impersonation of a specific person' is getting thinner. That raises stakes for fraud, identity theft, and digital legacy after death.


Purpose-Built Model Beats Chatbots at Predicting Human Behavior, Researchers Report

Researchers built Chiron, a foundation model trained specifically to measure psychological traits and predict human behavior, drawing on more than 1.4 million behavioral decisions and 100,000 psychometric assessments across 78 social science experiments. The researchers report it predicts individual behavior up to 28.4% more accurately than frontier LLMs and reproduces human experimental results with 82-94% less error, with measurements that better match real human self-reports.

Why it matters: Synthetic survey respondents and AI personas often flatten real human differences (D.A.D., November 17). This suggests purpose-built models, rather than repurposed chatbots, may be a more credible path for AI-assisted social science and market research.


Who Should Check AI's Work? Not the Fastest or Most Senior Staff, Study Finds

A study of 541,621 review decisions by 483 human reviewers checking an AI accounting system found that what makes a reviewer valuable isn't speed, activity, agreement with the AI, or years on the job. It's calibration, meaning whether someone's confidence tracks their correctness. A behavioral scoring method rated an overconfident group of reviewers higher than genuine experts, even though the experts' corrections held up 99.2% of the time versus 5.2% for the overconfident group. Reviewer quality also didn't improve with tenure, suggesting calibration isn't learned on the job. Applying the same logic to the AI focused 92.5% of human review effort on genuine errors.

Why it matters: Firms leaning on humans to catch AI mistakes may be screening and rewarding the wrong qualities. They need ways to spot reviewers who know what they don't know.


What's On The Pod

Some new podcast episodes

AI in Business — How to Transform an Enterprise That Already Works - with Alexandra Guenther of Leidos

Get tomorrow's briefing

Share this edition Subscribe