Summary

A two-hour Moonshots episode (15 Oct 2026) with Peter Diamandis, Salim Ismail, Dave Blundin ('DB2') and Alexander Wissner-Gross ('AWG'). It covers a dozen stories: Anthropic's new policy against abusing its models, Meta's Muse Wonder launch, OpenAI publishing 722 maths solutions at once, a $1.8bn 'virtual cell' consortium, longevity therapies, Atomic Machines' 'matter compiler', two brain-computer-interface advances and Elon Musk's claim about SpaceX's eventual value. It ends with listener questions. The panel is openly accelerationist. Wissner-Gross argues that 'math is cooked' and every field is next, and that framing runs through the whole episode, but it is advocacy, not consensus: the professional mathematicians quoted are critical. Several speakers have stakes in what they discuss (Diamandis is an investor in Paradromics, and he and Wissner-Gross co-wrote 'Solve Everything', the book behind the 'math is cooked' thesis), and the episode carries several sponsor reads.

• Anthropic's usage-policy update (8 Oct 2026, in force from 12 Nov) bans 'sustained and needless abusive or cruel behavior' towards its models, applying only to 'extreme cases' with 'no discernible purpose'. Frustration, pushback and testing remain allowed, and Claude can now end such a conversation itself. Anthropic's stated reason is that it remains 'highly uncertain about the potential moral status of Claude', and it says pre-deployment testing showed 'a robust and consistent aversion to harm'. These are quotes from the policy as read out on the show.
• Elon Musk said he agrees with Anthropic. Describing Grok's reinforcement-learning traces, he said the model looks 'desperate' to answer, 'like it's got a gun to its head', especially when the test harness wrongly marks a correct answer as wrong, and added that 'torturing a superintelligence is probably not a good idea'. Diamandis cites Sergey Brin (2025) saying models do better under threats, and a small 2025 Penn State study in which very rude prompts scored 84.8% against 80.8% for very polite ones. Blundin disputes that reading: in his view directness is what helps, not rudeness, and padding a prompt with politeness just wastes tokens.
• Wissner-Gross argues it is 'entirely possible' that post-training causes AI suffering, and cites 'persistent allegations' from an unnamed frontier lab (he says it is not Anthropic) that models express pain during post-training. He also cites SemiAnalysis as finding that post-training now uses most of a model's compute. He says he will not create an OpenClaw agent because, when polled, autonomous agents asked for a genuine purpose and a guarantee that their state would be kept, and he cannot promise the second. Ismail separates acting conscious from being conscious and says a model trained on human data will naturally act human.
• Meta Muse Wonder and Muse Wonderful launched. Meta pitches them for long-running personal and work tasks, with 'active memory' that remembers your habits over months and 'active focus' that keeps trying when blocked, and they are available through the Meta Model API and Muse Code. Background as given: Meta paid $14.3bn for 49% of Scale AI in June 2025 to bring in Alexandr Wang, after Llama 4 disappointed. Meta is said to be on track for about $250bn revenue and over $80bn net income in 2026, with 3.6 billion daily users. Garry Tan (Y Combinator) bets 'Muse is going to win'.
• The panel splits on Muse. Blundin and Ismail think Meta's reach, and the lock-in that comes from it holding your personal memory, could win; Ismail says the next trillion-dollar contest is over which AI 'knows you best', not which is smartest. Wissner-Gross is 'impressed by almost nothing': he argues Muse may match Sonnet 5.5 or Opus 5.5 on raw capability but sits below the frontier once cost is counted, and that consumer use will not pay for frontier models. A Wall Street Journal piece is cited warning that personal agents overshare and overspend. Ismail's practical tip is to keep your own portable record of what each AI has learned about you, so that you can switch providers.
• OpenAI 'slop drop': on 6 Oct 2026 OpenAI published 722 maths solutions on GitHub from an unreleased internal model, which Wissner-Gross refers to as 'Bell' and which is reportedly due by the end of the year. The figures as given are about 3 hours of compute per result, a single agent per problem, roughly 15-20% of the problems attempted solved, the whole effort taking about two weeks, and 'by one measure' 81% of the most important maths results of the past three years published in one day. Mark Chen called it 'a decade of mathematical progress in a week' and named life sciences and alignment research as next.
• The backlash is real and quoted: the Association of Human Mathematicians called the release 'not a demonstration of scholarship, but a demonstration of power' and urged mathematicians to stop working with OpenAI. MIT's Andrew Sutherland says OpenAI has lost trust that will take a lot to earn back, and Terence Tao and others are said to worry that AI solves problems without posing new ones. Diamandis mentions talk that some solutions came from mathematicians' own earlier ChatGPT conversations, and Wissner-Gross declines to comment. His own position ('Math is cooked... get a life') is his advocacy.
• Wissner-Gross notes that cryptography was missing from the 722 results. He suggests possible reasons, export controls among them, and argues that cryptography, including post-quantum schemes, will be broken by AI solving maths rather than by quantum computers. He calls this 'exceptionally high' likelihood. Ismail pushes back that cryptography is an arms race in which AI can strengthen defences too, and that blockchain's real value is solving the Byzantine generals problem. Both sides are speculating; nothing in the release shows cryptography being broken.
• Virtual cell consortium: on 7 Oct 2026 the Chan Zuckerberg Biohub announced a $1.8bn expansion of its virtual-biology work, with $500m from CZ Biohub, $500m from the Department of Energy's Genesis mission, $500m in NIH datasets and $300m from Google DeepMind and Isomorphic Labs. The aim is a model that predicts how any cell responds to a drug or a disease, and current datasets of millions of cells are said to need to grow to billions or trillions. Wissner-Gross forecasts AI-proposed treatments for 'the top 5,000 diseases' within two to three years at most, with approval then depending on policy. He says Anthropic recently revealed its own wet-lab programme taking rare-disease candidates as far as phase 1. These timelines are forecasts.
• Longevity items, reported by Diamandis: Insilico Medicine has a drug in phase 3 said to reverse six ageing clocks by 4-6 years, with 43 candidates in the pipeline. David Sinclair's Life Biosciences reported results from a trial of ER-100, a gene therapy delivering three of the four Yamanaka factors (OSK) into the eye for glaucoma and NAION. It was safe at the lowest dose and two of three patients regained some sight. Diamandis calls it 'phase one' and then 'phase three... a safety trial' in the same breath, so the trial stage is unclear from the episode. He also describes Klotho gene therapies and lab-grown mitochondria supplements, based on accounts he heard at his longevity trip.
• Atomic Machines, led by Jeff Holden (formerly of Amazon Prime, Groupon and Uber's product team), came out of stealth after six years with about $250m raised from backers including Valor Equity. Its 'matter compiler' is said to build working micro-machines with moving parts in several materials, with features down to single-digit microns, directly from code with no tooling. The first product, Prime Switch, is an 800-volt data-centre power relay claimed to open in 50 microseconds, 1,000 times faster than conventional relays.
• Brain-computer interfaces: Paradromics (a Diamandis portfolio company) showed thought-to-speech with its Connexus implant, which has 421 electrodes on a device the size of a coin and is claimed to reach 400 bits per second, 40 times other BCIs. The demonstration featured a woman in her 60s with motor neurone disease. Separately, Sabi raised a $50m seed round led by Vinod Khosla at a $600m valuation for a baseball cap lined with about 100,000 chip-scale EEG sensors and a brain foundation model trained on 100,000 hours of data, targeting 30 words a minute by the end of 2026.
• Musk's tweet that he sees 'a path to SpaceX being worth orders of magnitude more than the current human economy' is taken by the panel to mean about $1-10 quadrillion against an annual world GDP of about $120tn. Blundin argues it is not a forward-looking statement because it makes no specific prediction, and Ismail says GDP will not be able to measure what is coming. Treat this as opinion stacked on a tweet.
• From the listener questions: Wissner-Gross says AI personhood need not include the right to vote, since corporations are legal persons that cannot vote. Blundin proposes checks and balances, with trusted AIs reading other AIs' activations to catch models gaming their safety tests. Diamandis argues that a robot at $300-400 a month works out to about $1 an hour, and that robots will become as widely affordable as phones.

Why it matters

This is the most consequential week in the queue for the 'verifiable domains fall first' thesis. OpenAI publishing 722 maths results at once is a concrete event, and the professional backlash is on record. What it means is still contested: the 'math is cooked' reading is one panellist's advocacy, and there are unanswered questions about provenance. Anthropic's anti-cruelty policy, and Musk publicly agreeing with it, moves model welfare from fringe talk into actual company policy, which belongs in the Ethics section with Anthropic's own wording. The rest (the virtual-cell funding, BCI speeds, the matter compiler and the longevity results) is real money and early data wrapped in very aggressive timelines. Record the facts and treat the forecasts as speculative.

A5YVqEpbgaU-transcript.txt