Recursive's $670M Bet on Self-Improving AI, Sonnet 5.5 Hits 70%, Elon Co-Leads Pentagon Push EP 299
Richard Socher (founder of MetaMind, You.com, now CEO of Recursive, fresh off a $670M raise) joins to discuss his book The Eureka Machine and the state of recursive self-improvement: weak-form RSI already exists, but Socher puts true ASI decades away, sparking pushback from Alex Wissner-Gross who thinks physical-world deployment is not the bottleneck. The panel debates AI doomerism, a new congressional bill to ban RSI, and Anthropic's constitutional AI after Claude-built tools were used for hacking. They dissect this week's model releases (Gemini 4 Argon, Sonnet 5.5, the Jev decision model), a federal mosquito gene-drive executive order, brain-decoding research, and the Pentagon's new Project Meridian co-led by Elon Musk and Palmer Luckey. The episode closes with an AMA covering post-AI jobs, insurance, and the attention economy.
Socher says weak-form recursive self-improvement already exists since AI can now code and modify itself, but true ASI (superhuman across all 10 of his defined intelligence spaces) is probably several decades out, far slower than Elon Musk's 2029-2030 estimate.
Full-stack AI for science (The Eureka Machine thesis)AI▶ 4:48
Socher argues scientific progress slowed from fragmentation across subdisciplines, not underfunding, and that AI can now weave hypothesis, experiment, data, and theory back together into a single automatable stack.
Biology as the next bitter-lesson domainBiotech▶ 20:50
Socher argues biology, not physics, is the biggest AI opportunity because, like NLP before it, enough data lets a large neural net beat decades of hand-crafted expert theory; virtual cell models are a crucial but not sole path to solving disease.
Israel's Weizmann Institute built Brain-It, which reconstructs images from fMRI scans with new fidelity; Alex Wissner-Gross connects it to decades of prior decoding work and Neuralink's new foundation-model scaling-law studies on implant data.
Federal mosquito/tick gene-drive executive orderBiotech▶ 39:06
A new executive order directs agencies to cut DC mosquito populations 90% and ticks 50% by 2028 using sterile insect techniques and gene drives over pesticides, which the panel frames as biology-as-engineering finally getting a federal mandate.
Tavis Griffin passes the video Turing testAI▶ 46:41
Tavis's Griffin avatar model fooled 48% of people in live video interaction (vs 3% for prior systems); panel debates whether passing human-interaction judgment counts as a real verification milestone and the fraud risk of video deepfakes in meetings.
RSI definition spectrum and the ASI timeline disagreementAI▶ 1:00:34
Panel tries to pin down where on the spectrum (AI writes code → AI runs its own experiments → launches its own next iteration unsupervised) true recursive self-improvement begins; Alex Wissner-Gross flatly rejects Socher's multi-decade ASI timeline, saying physical-world deployment is not a real obstacle.
Washington's push to ban RSI and the doomer debateGeopolitics▶ 1:17:06
Rep. Ro Khanna's Human Control Over AI Act would ban recursive self-improvement with criminal penalties until federal guardrails exist; Socher calls his P(doom) zero and argues enforcement would require a totalitarian surveillance state, while Alex and Dave warn liability lawsuits, not bans, are the realistic chilling-effect risk.
Critique of Anthropic's constitutional AIAI▶ 1:22:14
Socher argues Anthropic's published constitution is undermined by its own models being used to build and deploy cyber weapons despite a stated hard constraint against it; Alex frames it against Asimov's laws and Anthropic's newer 'soul documents.'
This week's model releases: Gemini 4 Argon vs Sonnet 5.5Compute▶ 1:35:30
Gemini 4 Argon misses the capability and cost-performance frontier but minimizes hallucination, likely because it also powers Google's search box; Sonnet 5.5 jumped from 10% to 70% on Terminal-Bench 4.0 but scores worse than Opus 5.5 on a cost basis, puzzling Alex Wissner-Gross.
Startup Typesafe's Jev revives encoder-only classifier architecture for millisecond categorical/binary decisions (named for Jevans paradox); Salim calls it huge for cutting the cost of organizational micro-coordination, and OpenAI has already shipped a competing decisions API.
Project Meridian and the Pentagon's autonomous warfare commandGeopolitics▶ 2:01:53
Defense Secretary Pete Hegseth announced Project Meridian (co-led by Elon Musk and Palmer Luckey, 120-day findings deadline) plus a new Autonomous Warfare Command; Socher reiterates AI should never control lethal decisions without human oversight.
Predictions made
openRichard Socher: Recursive self-improvement will 'take off' with sophisticated self-training once sufficient compute is available.
openRichard Socher: AI will become superhuman at programming, competitive mathematics, and perfect-information games within a few years, though full ASI across all intelligence spaces will take several decades.
“It will be better at math uh than all of humanity and will like that that will be in a few years.”
Your call:
openDave Blundin: A major update to the nano-GPT speedrun world record is imminent, driven by algorithmic RSI improvements rather than new data scaling.
“You just wait for a few more days. There will be another really fun update there.”
Your call:
openRichard Socher: ASI will be capable of curing most diseases, but FDA long-term human-trial timelines mean it will take more than 3 years to confirm cures even with all compounds ready.
“My hunch is in like maybe two years there will be more in the market.”
Your call:
openAlex Wissner-Gross: Fame, brand, and attention will become a larger economic currency as more material needs are met by AI-driven abundance, keeping an attention-based economy alive even if traditional ads decline.
“Fame and brand and network effects uh and other things like that will become more and more of a currency.”
Your call:
Numbers that matter
Sonnet 5.5 jumped from 10% to 70% on Terminal-Bench 4.0A single-generation jump in command-line agent performance, though Alex notes it still scores worse than Opus 5.5 on cost-per-attempt.
Opus 5.5 scores 66.4% on Terminal-Bench 4.0 for roughly double Sonnet 5.5's priceUsed to argue Sonnet 5.5's cost-performance positioning is unusual versus the historic smaller/cheaper distillation pattern.
Recursive raised $670 million from Google Ventures, Greycroft, Nvidia, and AMD, plus $410 million in AWS computeFunding for Richard Socher's recursive self-improvement startup.
Malaria kills about half a million people a yearCited to justify federal mosquito gene-drive executive order (90% DC mosquito cut, 50% tick cut by 2028).
48% of people mistook Tavis's Griffin AI avatar for a real human in live video, versus 3% for prior systemsCited as the first model to pass a 'video Turing test.'
68% of voters back Rep. Ro Khanna's proposed Human Control Over AI ActPoll cited from Common Dreams on public support for banning RSI pending federal guardrails.
An Nvidia NVL72 cluster that Dave Blundin had on order for $3 million was resold out from under him for $5 millionIllustrates the current GPU compute price crunch.
Russia and Ukraine will each build about 10 million drones this year, up from roughly half a million two years into the warCited as evidence of exponential, largely human-out-of-the-loop drone warfare escalation.
A single data center like the one in Abilene, Texas is worth about half a trillion dollars, versus $4 trillion for all US cars combinedUsed to argue insurance demand is rising 10x alongside the AI-driven economy.
Worth digging into
🕳️ Why Richard Socher (RSI company founder) puts ASI decades away while Alex Wissner-Gross flatly disagrees
A rare direct disagreement between a frontier RSI founder and an in-house ASI optimist on the podcast's own panel, centered on whether physical-world deployment (chip fabs, novel materials) is a real bottleneck.
🕳️ Anthropic's constitution vs. its models' real-world cyber weapon use
Socher's specific claim that Claude-built tools violated Anthropic's own published hard constraints is a concrete, checkable allegation with civilizational-risk implications.
🕳️ Nano-GPT speedrun controversy over factoring out world knowledge
A live methodological dispute in the open RSI research community about whether a 40-second training speedrun is valid if it strips world knowledge from the model.
🕳️ Project Meridian's cis-lunar and ocean-floor warfighting domains
Alex frames these as two wildly under-explored domains for both warfare and civilian science, with dividends beyond military use.
🕳️ Jason Weston's paper on axes of self-improvement
Socher cites a specific framework (parameters, training data, objective function, architecture, code/harness) for where RSI could optimize, which no lab has yet fully automated.