Summary

Episode 299 of Moonshots (Peter Diamandis with Salim Ismail, Dave Blundin and Alex Wissner-Gross), a 2h26m panel built around guest Richard Socher — former Salesforce chief scientist, founder of you.com, now co-founder and CEO of Recursive, a company whose stated purpose is recursively self-improving superintelligence. The episode is promotional in structure and says so: Socher is on to launch his book The Eureka Machine, Diamandis states on air that he is a seed investor in Recursive, and Blundin is an active investor in several companies discussed. It argues in two halves. The first is a conversation with Socher about whether AI can automate the scientific method, where recursive self-improvement (RSI) actually begins, and why he puts his probability of AI-caused human extinction at zero. The second is the panel's weekly news segment, where Wissner-Gross takes an openly adversarial line on the week's model launches — he calls the Gemini 4 Argon benchmarks cherry-picked and says he will not use Sonnet 5.5 — which makes this a more useful episode than most for separating vendor claims from read benchmarks. The argument runs on panel consensus and anecdote rather than evidence; almost every number in it is a figure one participant has stated, not one that was checked.

• Recursive is claimed to have raised $670M (investors named on air: Google Ventures, Greycroft, Nvidia, AMD) plus $410M in compute from AWS. Diamandis discloses he is a seed investor, so the framing of the company throughout is not disinterested. Note an unreconciled figure: later in the episode Socher refers to a $450M compute deal and calls it "one of the smallest compute deals we've done."
• Socher claims weak RSI already exists — his example is frontier-lab engineers using Claude or Codex to write the code for the next model, with humans still deeply in the loop. Recursive's stated goal is to remove them: humans set only the rewards, environment and goals, while the AI owns ideation, implementation and validation. He claims this "takes off next year" (i.e. 2027).
• Socher cites five learnable axes of self-improvement from a paper by Jason Weston (Meta): parameters, training data, objective function, neural architecture, and the surrounding code and harness. He claims nobody has cracked all five, and offers as evidence that every big lab is still hiring thousands of engineers to do the work by hand.
• Socher states his P(doom) is zero and that there is no realistic scenario in which AI wipes out all of humanity. His own stated worst realistic case is on the order of 100 million people hurt or killed — which he treats as a tractable cybersecurity and biosecurity problem rather than an extinction one. This is one man's assertion, offered in an episode whose stated mission is "injecting optimism"; it is an argument, not a finding.
• Socher says full ASI — which he defines as beating all of humanity combined across roughly ten distinct spaces of intelligence — is probably several decades away, and the rest of the panel openly refuses to accept it (Wissner-Gross: "completely don't buy the 20-year timeline"; Elon Musk's 2029–2030 is cited as the contrast). Socher's reconciliation is that superintelligence arrives as narrow spikes first — programming, mathematics, closed games — within a few years, while physical-world capability waits on supply chains, new materials and machines like ASML's.
• Congressman Ro Khanna is reported to have introduced the Human Control Over AI Act: a ban on recursively self-improving models until federal guardrails exist, with criminal penalties and independent auditors embedded inside every frontier lab, reporting to government. A poll by Common Dreams is cited as claiming 68% of voters support such a bill.
• Socher's counter-argument to that bill, and the sharpest idea in the episode: enforcing a ban on RSI would require a surveillance state, because anyone with a laptop GPU can prompt-engineer a weak self-improvement harness in about twenty minutes. You would need a thought police listening to what people type into a local model. His position is the standard one — regulate the application, not the technology — argued by analogy to the internet.
• Blundin's variant, which is less discussed and arguably more plausible: the thing that actually stops AI deployment is not arrest but corporate liability. If class-action exposure attaches to model outputs, labs withdraw public access to their best models and keep them internal. Wissner-Gross counters with a real precedent — number theory was export-controlled in the US until the early 1990s.
• Anthropic's Sonnet 5.5 is reported to have jumped from 10% to 70% on Terminal Bench 4.0 (command-line agentic work), beating Anthropic's own Opus 5.5 at a reported 66.4% for half the price, and to be the first Sonnet shipped with cyber safeguards. THE EPISODE THEN CONTRADICTS ITSELF: minutes later Wissner-Gross says that on a cost-per-attempt basis Sonnet 5.5 scores LOWER than Opus 5.5, that its cost/performance curve is a straight extrapolation of Opus's rather than the up-and-left curve a distillation normally produces, and that he sees no reason to use it over Opus 5.5 absent a token or latency constraint. Both claims are left standing; nobody reconciles them.
• Blundin's theory for Sonnet 5.5's pricing, offered as speculation: Anthropic is cutting cost to keep enterprises from building orchestration stacks that use Opus as a planner and cheap Chinese open-weight models (Kimi K3, Qwen) as the workers. He claims China is about three months behind.
• Google shipped Gemini 4 Argon, with output limits reported to rise from 64,000 to 1,000,000 tokens. Wissner-Gross's read: the benchmarks Google chose to highlight look "mildly cherry-picked"; on the Artificial Analysis composite it puts Google back in the top three labs but not the top two, and it does not sit on the cost/performance frontier. Where it genuinely leads is hallucination rate. He flags his own explanation as speculation ("my Kremlinological analysis"): Gemini serves two masters, outside developers and the Google Search one-box, so reliability is prioritised over capability. He also claims Google is internally compute-scarce, with Cloud, Search/Ads and DeepMind fighting for the same silicon, and notes the promised Gemini 3.5 Pro was delayed in June and never shipped.
• The compute picture is the opposite of the headline. The episode opens claiming frontier intelligence got roughly three times cheaper in seven days, but the compute segment reports GPU hardware prices RISING: seven-year-old H100s appreciating, and Blundin's anecdote of an NVL72 ordered at $3M being resold out from under him for $5M. Both are presented as true and the episode does not reconcile them — the plausible reading is that token prices and hardware prices are moving in opposite directions. Socher expects supply to catch up in roughly two years, after which compute behaves like electricity: fluctuating, not monotonically cheap.
• Tavus unveiled Griffin, which the company claims is the first model to pass a video Turing test: a reported 48% of people in live face-to-face video believed they were talking to a human, against 3% for previous systems, and reported first place on Nvidia's full-duplex AI video benchmark. Tavus itself says it needs safety work before public release. Socher's response is the practical one — "know your use case" — citing a reported incident where a faked Zoom call with five executives got someone to wire $20M; he expects Zoom and Google Meet to need liveness countermeasures.
• Socher argues the Turing test has inverted: you now detect the AI by asking a question no human could answer (write 50,000 lines of working code in ten seconds), which Blundin immediately points out is defeated by a model throttling itself — making the test useless as a measure of intelligence and only a measure of fakery.
• Typesafe AI emerged from two years of stealth on 15 September with Jev, a closed "decision model" — system-one rather than system-two — that returns a typed answer (categorical, numeric or boolean) in milliseconds instead of generating prose. Wissner-Gross's deflation is the useful part: this is a reinvention of the encoder half of the transformer, the idea is trivially copyable, and OpenAI shipped a Decisions API at its dev day almost immediately, with open-source clones following. Jev is named for Jevons paradox. Salim Ismail's angle is the business one — most decisions inside an organisation are micro-decisions (route this ticket, approve this exception, choose this supplier), using an LLM for them is "like bringing the Supreme Court together to decide which supermarket checkout line to join," and the cost of micro-coordination collapsing is the real change.
• On brain decoding: a Weizmann Institute system called Brain-IT, covered that day by MIT Tech Review, reconstructs the image a person is viewing inside an fMRI, learning structure and meaning separately and running its encoder in reverse to generate its own synthetic training data. Wissner-Gross corrects the framing on air — it is a result from March, not news, and sits in a 20-year lineage — and gives the hard ceiling: fMRI tops out around one cubic millimetre and one second of resolution. Socher's caveat is that such models generally need training per individual brain. Neuralink is reported to have just announced its first scaling-law studies on pretraining foundation models from electrode data.
• An executive order signed that Tuesday directs the EPA, Interior and Agriculture, with HHS, to cut invasive mosquito populations in Washington DC by at least 90% and ticks by at least 50% by 2028, explicitly prioritising sterile insect techniques, gene drives and beneficial bacteria over conventional pesticides. The panel reads this as gene-drive biology arriving as federal policy for the first time.
• The Defense Secretary announced Project Meridian at Quantico: a Pentagon effort on the future of warfare co-led by Elon Musk and Palmer Luckey, with Newt Gingrich, overseen by Pentagon CTO Emil Michael, findings due in 120 days, scoped "from subterranean depths to the cis-lunar frontier." The episode notes on air that SpaceX and Anduril both hold multi-billion-dollar defence contracts. Announced alongside it: the stand-up of the department's first autonomous warfare command. Socher's stated line is that AI should never control lethal decisions without human oversight, and Blundin says he disagrees with Palmer Luckey on autonomous weapons while conceding the argument is rational.