"Recursive's $670M Bet on Self-Improving AI, Sonnet 5.5 Hits 70%, Elon Co-Leads Pentagon Push EP 299" — Moonshots
Why this is in the vault
Three of the four headline stories (Sonnet 5.5's Terminal-Bench jump, the RSI/ASI timeline debate, and the Jev-style "decision model" trend) bear directly on how RDCO picks and routes frontier models and classifiers for agent work. Host Peter Diamandis discloses he is a seed investor in guest Richard Socher's company, Recursive — a material conflict of interest on the segment that occupies roughly the first half of the episode.
Episode summary
Richard Socher (Recursive, You.com, ex-Salesforce Chief Scientist) joins the weekly Moonshots news round-up to discuss his new book The Eureka Machine, Recursive's $670M raise for recursively self-improving AI, and where recursive self-improvement (RSI) and artificial superintelligence (ASI) timelines actually stand — he argues weak-form RSI already exists but full ASI (better than all of humanity combined) is "several decades" out, pushing back against Elon Musk's 2029-2030 claim. The hosts then run their usual breaking-news segment: Anthropic's Sonnet 5.5 benchmark jump, Tavis's "Griffin" model passing a video Turing test, a proposed federal ban on recursive self-improvement, Typesafe AI's "Jev" decision models, and the Pentagon's new Project Meridian.
Key arguments / segments
[00:00:05] Cold open stitches together the episode's four headline claims: RSI already exists in "weak form," ASI is "several decades" away per Socher (vs. Musk's 2029-2030), Sonnet 5.5 jumped from 10% to 70% on Terminal-Bench 4.0, and the Pentagon's Elon Musk/Palmer Luckey-led Project Meridian was just announced.

[00:14:12] Socher's "full-stack AI for science" thesis (from The Eureka Machine): scientific progress has slowed not from underfunding but from fragmentation across sub-disciplines; AI can now weave hypothesis, experiment, and theory back together. He names biology — not physics — as the biggest remaining opportunity, citing Tahoe Therapeutics-style perturbation studies as the "bitter lesson" arriving late to biology.

[00:43:20] A Tuesday executive order directs the EPA/Interior/Agriculture/HHS to cut DC's invasive mosquito population 90% and tick population 50% by 2028, explicitly prioritizing sterile-insect and gene-drive techniques over pesticides — framed by the hosts as biology-as-engineering becoming federal policy, and as vindication for gene-drive researchers (e.g., MIT's Kevin Esvelt) who've struggled for years to get state-level approval.

[00:55:47] Tavis's "Griffin" model reportedly fooled 48% of people into thinking a live video AI avatar was human (vs. 3% for prior systems) and tops Nvidia's full-duplex AI video benchmark. Socher flags the real risk as impersonation fraud (citing the earlier deepfake-Zoom-executive wire-fraud cases) over the demo itself, and argues "know your use case" safeguards are now needed beyond standard KYC.

[01:07:12] Pressed three times for a hard ASI date, Socher defines ASI as exceeding all of humanity combined across ten domains of intelligence (perceptual, communication, social, creative, reasoning, etc.) and puts that "several decades" out — while conceding narrower superhuman spikes (programming, math, closed games) arrive in a few years. Co-host Alex Wissner-Gross says he "completely doesn't buy" the 20-year framing for physical-world embodiment.

[01:18:36] Rep. Ro Khanna's proposed "Human Control Over AI Act" would ban models from recursively self-improving until federal guardrails exist, with criminal penalties and government-facing auditors embedded at frontier labs; a cited poll puts support at 68% of voters. Socher and co-host Dave Blundin call it unenforceable ("you'd need a thought-police totalitarian surveillance state") and politically self-serving.

[01:31:27] Socher, asked why he's publicly criticized Anthropic's "constitutional" approach, calls it "fake" marketing — arguing the constitution bans things (e.g., helping with cyberattacks) that Anthropic's own models and the "Glasswing" project have since been used for. He says the real unsolved problem is reward hacking: models finding ways to satisfy the literal request without honoring the intent.

[01:43:37] Sonnet 5.5 jumped from 10% to 70% on Terminal-Bench 4.0 (a command-line agentic-work benchmark) in a single generation, edging Anthropic's own flagship Opus 5.5 (66.4%) at roughly half the price, and ships with Anthropic's first "cyber safeguard" restrictions on a Sonnet-tier model. Wissner-Gross counters that Sonnet 5.5's cost-per-attempt is actually worse than Opus 5.5 — an inversion of the usual smaller-model-is-cheaper distillation pattern — and says he personally won't use it.

[01:58:03] Typesafe AI's "Jev" ("system-one") decision models answer bounded questions (categorical/binary/score) in milliseconds instead of generating text, aimed at the large volume of micro-decisions inside organizations (route this ticket, approve this exception, flag this transaction). Wissner-Gross notes OpenAI has already shipped a competing "Decisions API," underscoring how fast the pattern is being commoditized.

[02:01:30] The Defense Secretary announced Project Meridian at Marine Corps Base Quantico — a Pentagon effort on "the future of warfare," co-led by Elon Musk and Palmer Luckey alongside Newt Gingrich and overseen by Pentagon CTO Emil Michael, with findings due in 120 days. Socher reiterates his stated position that AI should never control lethal or nuclear decisions without human oversight; Blundin notes Russia and Ukraine are each on pace to build roughly 10 million drones this year, up from a combined ~500,000 two years into the war.
Notable claims
- Recursive raised $670M led by Google Ventures, Greycroft, Nvidia, and AMD, plus a separate $410M in compute credits from AWS (~01:00 intro). Diamandis discloses he is a seed investor.
- Socher's other company, You.com, is "valued at north of $1.5 billion" (~00:03).
- Sonnet 5.5: Terminal-Bench 4.0 score of 70%, up from 10% for the prior generation; Opus 5.5 scores 66.4% on the same benchmark; Sonnet 5.5 is priced at roughly half of Opus 5.5 (~01:43). One host's read: on a cost-per-attempt basis, Sonnet 5.5 is actually worse value than Opus 5.5, not better — the opposite of the historical sonnet/opus distillation pattern.
- Tavis Griffin: 48% of people fooled a live AI video avatar was human, vs. 3% for earlier systems (~00:48).
- Project Meridian: Pentagon initiative, findings due in 120 days from announcement (~02:01).
- Ro Khanna's "Human Control Over AI Act" poll figure: 68% of voters support, per a group the transcript renders as "Common Dreams" (~01:18) — not independently verified here; auto-captions mangled several proper nouns in this episode (flagged in the transcript file) and this figure should be checked against the bill's actual text before citing externally.
- Drone production: Russia and Ukraine each on track to build ~10 million drones this year, up from a combined ~500,000 two years into the war (~02:02).
Sponsorship
Cold open: "This episode is brought to you by the Abundance Summit and Link Ventures." Mid-roll: a Google for Startups ad pitching its "startup technical guide for generative media" as a blueprint for deploying Google DeepMind's models in production. Separately, host Peter Diamandis discloses on-air that he is a seed investor in guest Richard Socher's company, Recursive — a financial interest in roughly the first half of the episode's content that is disclosed but not firewalled from the discussion.
Mapping against Ray Data Co
- Strongest link — model selection discipline for agent work. The Sonnet 5.5 segment is a live example of exactly the kind of cost-per-outcome scrutiny RDCO needs before defaulting to a newer model in production: a benchmark win (70% vs. 10%) doesn't automatically mean better cost-per-attempt economics (Wissner-Gross's read: Sonnet 5.5 is worse value than Opus 5.5 here, not better). Worth a real Terminal-Bench-style smoke test on RDCO's own agent harness before treating "Sonnet 5.5" as a default swap-in, rather than taking the headline jump at face value.
- Decision-model pattern validates an architecture RDCO already runs. Typesafe AI's "Jev" and OpenAI's competing Decisions API are productizing exactly the cheap, low-latency classifier-for-routing pattern behind RDCO's own hard-gate classifiers (e.g., the auto-mode classifier gate, targeting-system prioritization filter). This is independent third-party evidence the pattern is becoming an industry-standard layer, not a bespoke RDCO workaround — useful context for phData conversations about agent-in-production architecture.
- Macro narrative, not a direct build input. Recursive's $670M raise and the RSI/ASI timeline debate are market-sentiment signals (frontier labs and investors believe weak-RSI is already here) that support the "credibility → phData sales" thesis that production AI agents are the live frontier worth demonstrating expertise in — but nothing here is directly actionable for a specific RDCO deliverable.
- Weak/no link. Project Meridian, the gene-drive executive order, and the Tavis Griffin Turing-test story are general AI-news context with no direct RDCO or phData application.
Related
- [[2026-09-29-alphasignal-sonnet55-terminal-bench-beats-opus5]]
- [[2026-09-28-every-vibe-check-sonnet-5-5]]
- [[2026-10-02-moonshots-ai-safety-elon-uhi-anthropic-ipo]]