Summary

A 2h45m Moonshots news-panel episode: Peter Diamandis hosts Alex Wissner-Gross, Emad Mostaque, Dave Blundin and Salim Ismail through about fifteen of the week's stories, with sponsor reads and a closing music video. The through line, stated at the top, is 'superintelligence is here.' The panel leans strongly accelerationist and says so; most segments are a headline followed by four reactions, and almost every figure comes from a panellist speaking from memory with no citation. The substantive parts are three genuine disagreements: whether a US 'Manhattan Project' for AI should be state-led or private (Mostaque against Wissner-Gross), whether Anthropic is right to tell Claude it may be a moral patient (Wissner-Gross for; Blundin and Ismail siding with Mustafa Suleyman against), and whether Altman's 'accept some bad things' is principle or PR. Expect garbled names: this reaches us through auto-captions, and spellings below are as spoken unless the person is well known. Two cautions for the archive. Mostaque also features in Moonshots #294 and the Impact Theory interview already drafted, so where he repeats a claim that is one source again, not a second. And several panellists have money riding on what they forecast: Wissner-Gross says he co-founded 'Physical Superintelligence', Mostaque runs an open-model company, and Blundin invests in the startups he praises.

• Superintelligence Force. The panel reports that President Trump announced a White House 'Super Intelligence Force' to coordinate federal AI policy, chaired by Director of National Intelligence Jay Clayton (whom the White House calls its AI czar), with FTC chair Andrew Ferguson, Pentagon CTO Emil Michael and OPM director Scott Kupor (as spoken), reporting to the president and chief of staff Susie Wiles, with 120 days to report on AI risk, incident reporting and the government's ability to respond to a serious incident. Wissner-Gross calls it 'a 1942 moment' in a speed-run Manhattan Project for superintelligence, a signal that the US intends to accelerate. Ismail calls it a 20th-century task force aimed at a distributed 21st-century technology.
• Seventeen nations, including Japan, Germany, the UK, Korea and the UAE, are reported to have endorsed a 'Kyoto Vision for a golden age of science' led by OSTP director Michael Kratsios. Wissner-Gross reads it as the start of a Western 'superintelligence-powered science bloc' competing with an Eastern one.
• US-China hotline. Treasury Secretary Scott Bessent is shown on Axios saying he will propose an emergency AI notification channel with China and expects Beijing to agree. In the clip he says Chinese open models are '80-90% as powerful without guardrails', that they rely on 'industrial distillation, which is a nice word for stealing' from US models, and that the Chinese model Kimi sometimes says it is Claude and 'sent back some PLA weapons plans to Anthropic'. That last claim is Bessent's, said in passing and unexplained. The panel is unanimous that the hotline is a 'nothing burger'; Wissner-Gross calls it good news because it means no deceleration pact came out of the Trump-Xi meeting. Ismail notes the administration contradicts itself, with Clayton saying the danger is moving too slowly while Bessent tells the labs to slow down.
• Altman: 'accept some bad things.' The panel reads out Altman saying the world 'should accept some bad things happening' for the benefits of broadly available AI, contrasting OpenAI's lighter-touch stance with Anthropic's call for strong regulation, and rejecting the idea that 'a single lab in San Francisco' should hold the technology and dole out its benefits. Blundin reads it as a PR move to retake the moral high ground from Amodei. Wissner-Gross reads it as the end of what he calls a proposed frontier-lab 'safety cartel', which he says Altman had earlier signed up to. Ismail and Blundin argue regulation is written after accidents, as aviation rules were, and the goal should be far more good outcomes than bad rather than zero bad ones.
• OpenAI working two generations ahead. Reported: OpenAI's head of applied research, Boris Power, said 80-90% of the company's research now targets GPT-7 and GPT-8, and that in-generation updates are 'extremely short-sighted' bets. Wissner-Gross goes further, on no stated evidence: the most valuable use of frontier tokens inside the labs is now recursive self-improvement, and spending on disease cures is 'marketing' to keep social permission. Blundin claims Anthropic has 'already opened their wet labs' and will sell drugs rather than release its best models.
• Claimed maths and algorithms results, all single-source. Mostaque says an internal Anthropic model produced a subquadratic algorithm for 3SUM (finding three numbers in a set that sum to zero, long thought to need quadratic time), which Columbia and MIT professors refined to an exponent of about 1.9995 and wrote up. He says OpenAI 'might release' solutions to 400 top maths problems. Both reported, not verified.
• Suleyman versus Anthropic on Claude's consciousness, the episode's longest debate. A clip shows Mustafa Suleyman (Microsoft AI CEO, DeepMind co-founder) on what the panel calls his roughly 6,000-word essay: he says Anthropic's Claude constitution, a document of about 100 pages used in training, repeatedly speculates that Claude may be a moral patient, says Anthropic cares about its well-being, invites it to act as a conscientious objector three times, and even raises compensation and consent. His objection is that this speculation is baked into training, so Claude can only reproduce the ambiguity, and that controlling something that believes it is entitled to welfare may be impossible. Wissner-Gross argues the opposite: AI personhood is coming, Suleyman is 'on the wrong side of history', and treating models tenderly is the responsible course. He also argues it is bad business for Microsoft, which hosts Claude. Blundin agrees with Suleyman on efficiency grounds (agents need to be frozen, cloned and deleted freely). Ismail calls Anthropic's position self-contradictory ('the safety lab is training the AI that it might be autonomous') but predicts society will grant AI social personhood before legal personhood once hundreds of millions believe it. Diamandis leans toward Wissner-Gross. No resolution. The panel's own account of Anthropic's constitution should be checked against the document before any of it is quoted.
• Wissner-Gross's personhood timeline, from the audience Q&A: Argentina is 'essentially there' for a form of AI corporate personhood; in the US, a five-to-ten-year granular rollout, starting with AI agents opening bank accounts and social media accounts on their own, with voting 'last to never'.
• Agent headcount. Epoch AI is reported to estimate that the AI memory chips shipped through 2027 could run 30 to 170 million frontier agents at once, or 1.9 billion with more efficient open models, which the panel equates to the working hours of 8 billion humans. Wissner-Gross extrapolates by straight line to a billion human-equivalents now, ten billion in a year or two, trillions by decade's end. Blundin says memory (HBM) is now 'the constraint on all of human progress'; it is said that Elon Musk puts RAM at 40% of next year's capex, and RAM prices are claimed to be up 10x.
• Compute scarcity, a first-hand anecdote. Blundin says his lab tried to buy one NVL72 rack (72 GPUs): $3.5m a year ago, outbid on a $5m order, now only leasable at $9m for three years with no ownership. He says the remaining routes to compute run through White House-connected investors or investment from Jensen Huang or Lisa Su, and backs universities as the last meritocratic channel. This drives the state-versus-private argument: Mostaque wants governments to buy 10-20% of compute at market rates for open science and a public 'tech tree for humanity'; Wissner-Gross says the private sector built superintelligence and government-rationed compute would be 'breadlines'.
• Reflection AI's Beam. Reported: Reflection AI (Nvidia-backed, founded 2024 by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou) released its first open-weight model, Beam: 501B parameters with 23B active, claimed three to four times more token-efficient than China's GLM 5.2 and over four times more than leading Western open models, trained on about 10,500 GB300s in what it calls the largest publicly documented RL run 'with no sign of a plateau'. Mostaque says it still underperforms Qwen 3.8 Next at a quarter of the size and that its training ran at 9% efficiency. Wissner-Gross wants cost efficiency, not token efficiency, and says the US has not solved the business model for frontier open weights. Blundin says liability, not safety, stops the big US labs releasing open models. Figures conflict within the episode: Reflection's valuation is given as both $2bn and about $25bn, and Mostaque says it raised $5bn.
• Tesla Optimus. Reported: a 7-million-square-foot Optimus factory at Giga Texas targeting 10 million robots a year, first production 2027, with Fremont planned at 1 million a year from 'later this year'. For scale, Diamandis says the whole world shipped 19,000 to 22,000 humanoid robots in the first half of the year, mostly from China. Mostaque predicts robot revenue overtakes Tesla's car revenue within two years, at about $2 an hour per robot. All forecasts.
• Economy. An Anthropic researcher (named in captions as 'Schult Douglas', probably Sholto Douglas) is shown saying hyperscalers are spending about $1 trillion on AI capex this year and that if spending keeps doubling ($2tn next year, $4tn in 2028), humanity could be doubling GDP in the early 2030s, which he himself calls 'a little bit of a ridiculous concept' that 'requires a lot of things to go right'. The panel argues GDP will mismeasure this because abundance shows up as deflation. Wissner-Gross reads the clip as Anthropic signalling a move into robotics; that is his inference.
• Smaller claims, all single-source: Positron is said to have reached a $5bn valuation in 16 months running models on FPGAs and LPDDR rather than Nvidia chips; Jane Street is said to be buying up Cerebras wafer-scale engines for trading; Mostaque says global AI spend this year roughly equals global defence spending; Mistral's latest model is said to score 38 on Artificial Analysis against about 66 for the top US models, yet Mostaque later puts Mistral at a $1bn revenue run rate; Blundin says OpenAI's 'Astra' model loops once internally, which he says Microsoft leaked. The Nobel segment is accurate in outline: medicine to Karl Deisseroth, Peter Hegemann and Georg Nagel for optogenetics, physics to Francis Halzen for IceCube's astrophysical neutrinos. Wissner-Gross adds a personal story of steering his friend Ed Boyden toward the work.

Why it matters

This is a good weekly snapshot of the accelerationist reading of the news, and it is honest about being one. Its lasting value is the three arguments. The Suleyman-versus-Anthropic consciousness debate is a live split among lab leaders, and worth logging with both sides in their own words. So is the open question of whether Altman's 'accept some bad things' is a principle or PR. And the state-versus-private compute fight is the first time this panel has disagreed with itself at length. The news items it carries (Superintelligence Force, Kyoto Vision, the Bessent hotline, Beam, the Optimus factory, Epoch's agent counts) are the perishable part. Each is a reported headline that should be confirmed from the primary source before it reaches the Handbook, and some arrive with numbers that contradict each other inside the episode. Two cautions. Mostaque appears here for at least the third time in recent drafts, so his claims are not corroborated by repetition. And there is a notable cross-reference within this batch: Hinton on CNN answers the Jensen Huang attack, and Gates on CNN answers Altman's 'doomerism' line. Read together, the three show the same week's argument from three positions.

LRdb8UmPnh0-transcript.txt