The bet no one will take

On January 29, 2025, Gary Marcus posted on X an offer to Dario Amodei, the chief executive of Anthropic. The question was straightforward. Would a model “smarter than almost all human beings” exist by the end of 2027? Marcus put $100,000 on no, to be judged by the ten tasks he had set out with Miles Brundage a month earlier. On June 30, 2025 he raised the stakes with a separate offer: $1,000,000 that AI in 2027 would not be “smarter than Nobel Prize winners across most fields in science and engineering.” The stakes are the part that has moved. Amodei’s answer is the part that has not. “Amodei never responded,” Marcus wrote of the million-dollar offer in his March 8, 2026 Substack post “There are no heroes in commercial AI,” and on July 23, 2026 he wrote on X that he wished Amodei “had taken this bet.” A web search of news coverage through October 10, 2026 turned up no reply.

Amodei has spent the past two years telling investors and policymakers that powerful AI could arrive as early as 2026, a “country of geniuses in a datacenter.” In August 2026 he called public distrust of AI a crisis and acknowledged that AI companies, Anthropic included, had not yet fulfilled their promises “to benefit the world”; in a September 12 essay he called for slowing the pace of capability gains and embedding independent evaluators in the labs. In the coverage searched through October 10, 2026, he had not put his name to a settlement criterion drafted by Marcus, at any price.

Two bets, one counterparty profile

The two wagers share a yardstick but not a counterparty, and the difference matters because the second bet is the one a counterparty actually took.

The first is the Amodei bet, direct and public, pitched on X to Amodei specifically, not as an open invitation. Marcus posted it in response to Amodei’s essay on DeepSeek and export controls, published the same day, which said AI “smarter than almost all humans at almost all things” was “most likely to happen in 2026-2027.” The offer echoes that phrase and dates it to the end of 2027, but it does not settle on the phrase alone. Marcus wrote that the bet would run “per criteria I laid out previously with @milesbrundage”: the ten tasks of the second bet. Several of those tasks overlap the bar Amodei set in his October 2024 essay “Machines of Loving Grace,” where powerful AI is “smarter than a Nobel Prize winner across most relevant fields,” able to “write extremely good novels” and “write difficult codebases from scratch,” and could arrive “as early as 2026.” Marcus is offering Amodei the opportunity to settle, in writing, the timeline Amodei has been quoting to reporters.

The second is the Marcus-Brundage 10-task bet, posted to Marcus’s Substack on December 30, 2024. Its criteria are the ones the Amodei offer borrows; its counterparty, stakes and odds are not. Miles Brundage, formerly of OpenAI’s policy team, took the bet at 10:1 odds: $20,000 from Brundage to a Marcus-chosen charity if fewer than 8 of 10 specified tasks are completed by end-2027, $2,000 from Marcus to a Brundage-chosen charity if 8 or more are. The tasks include Pulitzer-caliber books written with little or no human involvement, Nobel-caliber scientific discoveries, bug-free code of more than 10,000 lines from a natural-language specification, and the conversion of proofs from 20 published papers across different areas of mathematics into symbolically verifiable form, building on existing formal libraries, with all 20 required to succeed.

Brundage is the counterparty here. He is a former OpenAI policy researcher, not a frontier-lab chief executive. The 10:1 odds run in Marcus’s favor: Brundage risks $20,000 to win $2,000, terms the published agreement says reflect Brundage’s strong confidence that AI will do eight of the ten tasks. None of the lab chiefs had joined as a counterparty on either bet.

Who has and has not engaged

Begin with the recipient. By Marcus’s account, and in the coverage searched through October 10, 2026, Amodei had not addressed the bet, had not declined it and had not modified the timeline that prompted it. He continued to use language during the same window that maps closely to the Marcus criterion. In May 2025 Amodei told Axios’s Jim VandeHei and Mike Allen that AI could eliminate half of all entry-level white-collar jobs in the next one to five years, a range that opens inside the Marcus end-2027 window and runs to 2030.

Sam Altman had not engaged either bet. Marcus has noted, on Substack and on Bluesky, that Amodei “pretty much follows Sam Altman’s playbook,” which is a reading of the silence rather than evidence about it. Altman wrote in September 2024 that superintelligence was possible “in a few thousand days,” a horizon that does not commit him to Marcus’s end-2027 date.

Demis Hassabis had not engaged. Ilya Sutskever had not engaged. Sutskever’s public statements since founding Safe Superintelligence Inc. have been spare to the point of monastic, and the bet was not addressed in either of the two interviews he has given since the launch.

Richard Sutton, the University of Alberta computer scientist whose “Bitter Lesson” essay supplies the field’s foundational scaling argument, distanced himself from pure-scaling LLM accounts in 2025 podcast appearances. In July 2026 Khurram Javed announced Oak Lab, the startup he and Sutton co-founded after the two left Keen Technologies; its mission statement describes continual, real-time learning that uses multiple orders of magnitude less compute and energy than existing methods. He had not joined the bet on either side. The distinction matters: declining to defend a position in long-form is not the same as putting a number on its falsification.

Elon Musk previously declined a Marcus bet on AGI by end-2025 — the offer escalated from $1 million to $10 million. That declination is a useful baseline. Musk did respond, even if the response was no. The Amodei silence is a different category.

Brundage took his version and laid the 10:1 odds himself, and the structure of that acceptance is the affirmative case. Eight named figures have public positions on the AGI timeline. One became a counterparty, staking ten times what Marcus staked. The Marcus-Brundage 10-task list, drafted with the specificity of a contract — bug-free 10,000-line codebases from natural-language specifications, Pulitzer-caliber books, Nobel-caliber discoveries, symbolic-verification conversions of published proofs — is the rare object in the AGI discourse that survives the question “what would it take to be wrong about this by end-2027?” The lab chiefs whose forecasts prompted the bet have not constructed an equivalent for their own. Their public predictions are dated and dollar-figured, but the settlement criterion has been left unwritten. Marcus’s has been written and signed.

Why a CEO might rationally refuse

The Amodei bet, on its face, is unfavorable to anyone running a frontier lab — regardless of whether that person privately believes the bet is winnable. A chief executive whose product roadmap, fundraising deck, and recruiting pitch all depend on the assumption of undiminished returns to compute has a professional incentive not to put his name on a 2027 settlement criterion drafted by his loudest critic. The downside is asymmetric. If Amodei takes the bet and wins, the headline is that the Anthropic CEO defeated a contrarian academic at his own game, a moderately interesting but commercially marginal story. If Amodei takes the bet and loses, the headline is that the Anthropic CEO lost a public AGI wager to Gary Marcus, a story that becomes the lead paragraph of every Anthropic-related piece for the next decade. Risk-asymmetric refusal is normal executive behavior.

A second version of the counter: the bet’s settlement criterion is itself demanding. Marcus tied “smarter than almost all human beings” to the Marcus-Brundage 10-task list, which narrows the definitional play but still turns on verdicts such as “Pulitzer-caliber” books and “Nobel-caliber” discoveries, and the more specific the criterion, the more obviously favorable the odds become to the side that thinks the tasks are hard. Bug-free 10,000-line codebases from natural-language specs, Pulitzer-caliber books, Nobel-caliber discoveries — none of these were tasks any frontier model could perform when the bet was struck, which is what makes Brundage’s ten-to-one stake a strong statement of confidence. A scaling believer with less conviction would reasonably want different odds.

These are good objections. They establish that an executive could rationally refuse the bet without being a hypocrite, and that the unfavorable structure makes refusal weak evidence against scaling beliefs.

What they do not establish is the absence of any settlement criterion at all. Amodei has been forecasting with public dollar figures attached — the half of entry-level white-collar jobs eliminated, the “country of geniuses in a datacenter,” the doubling of the GDP. He has not been forced to those numbers; he has chosen them. Marcus’s wager asks for one number — a yes or no on his own forecast at his own date — and Amodei had not answered it: “Amodei never responded,” Marcus wrote on March 8, 2026, and no reply turned up in news coverage searched through October 10, 2026.

The right way to read this is not that Amodei’s silence proves the scaling argument is wrong. The bet did not create the asymmetry. It made the asymmetry visible.

A specificity that picks at one claim

Among the 10 Marcus-Brundage tasks, the bug-free 10,000-line code task is worth pausing on. Amodei said at a Council on Foreign Relations event in March 2025 that within three to six months AI would be writing “90% of the code,” and within a year “essentially all” of it. That forecast lands inside the Marcus-Brundage settlement window. It also implies that the bug-free 10,000-line task should resolve in Brundage’s favor by end-2027 — possibly by end-2026.

Brundage took the bet and laid the 10:1 odds himself. He was, until late 2024, inside OpenAI’s policy team, which puts him closer to the production realities of code generation than most outside observers. A former lab insider was willing to risk $20,000 to win $2,000 that AI would do eight of the ten tasks by end-2027.

This is the kind of specificity the bet was designed to elicit. The Pulitzer and Nobel tasks make the unfavorable-odds objection vivid — those criteria are stretched far enough that even a scaling believer might reasonably ask for asymmetric odds. The bug-free code task is calibrated to the executive’s own quoted forecast. It is the task on which the silence costs the most.

What an unanswered bet shows

The argument is not that any of these executives is dishonest. The argument is structural. A debate in which one side has made specific, dated, dollar-figure forecasts and the other side has offered specific, dated, dollar-figure settlement criteria is, until the second side responds, a debate in which only one party is on the hook for being wrong. The forecast survives by being repeated. The bet survives by being ignored.

The Marcus-Brundage structure imports a discipline the field has otherwise gone without. The lab chiefs have not yet written a settlement criterion for their own forecasts.

The point of this bet is not the dollar figure; the point is the date. The scaling argument may yet land within Marcus’s window, and Marcus’s record contains misses as well as hits. The 2027 date will resolve the empirical question. The bet will sit there until it does.