Recursive's $670M Bet on Self-Improving AI, Sonnet 5.5 Hits 70%, Elon Co-Leads Pentagon Push | EP #299

Recursive, co-founded and led by Richard Socher, raised $670 million from Google Ventures to pursue recursively self-improving superintelligence. Socher said weak forms already exist—AI helps engineers improve AI, but people still choose what to change and validate the results. Fully automated improvement across models, data, objectives, architectures, code, and testing has not been achieved. Recursive wants AI to handle ideation, implementation, and validation while people set goals, rewards, and environments. Socher expects strong recursive self-improvement to take off next year, but said it will require substantial compute.

Socher argued that artificial superintelligence in its broadest sense—capability beyond humanity across many areas, with some ability to choose what to work on—probably remains decades away. AI could exceed humanity in narrower domains, such as mathematics, within a few years; he suggested Elon Musk’s 2029–2030 timeline may refer to such capabilities. He sees AI, digitized scientific data, simulations, and robotics as ways to reconnect knowledge fragmented among specialized fields and speed up discovery. The Eureka Machine’s thesis is that AI could enable a century of scientific breakthroughs within the next decade, but progress will accumulate problem by problem and depend partly on compute.

In biology, Socher predicted that multiple diseases, potentially including simpler single-gene diseases, could be cured within 12 months. That does not mean treatments will quickly reach patients: human studies and FDA approvals could keep some therapies five to 10 years from widespread use. He contrasted biotech companies that once took roughly a decade to move one candidate into late-stage trials with companies now advancing five to 10 compounds into stage three after two or three years. Organoids, virtual cells, simulations, and large biological datasets could accelerate research, but speakers disagreed on whether cell models might eventually replace human trials. Modeling cells remains difficult, and measuring every protein in a single cell currently destroys it.

The episode also examined reported gains in Anthropic’s Sonnet 5.5: it scored 70% on Terminal-Bench 4.0, up from 10%, versus 66.4% for Opus 5.5. The introduction said Sonnet costs half as much, but Alex said the cost-per-attempt comparison did not establish a performance or value advantage; token use, latency, and other factors would matter. One speaker speculated that the release was intended to close a perceived three-month gap with China before enterprises move to open-source models. Another said there was no clear reason to choose Sonnet over Opus without a demonstrated advantage. More broadly, speakers said compute remains physically constrained: GPU prices have risen, buyers are competing for large clusters, and Richard expects added capacity in about two years, not an imminent supply glut.

The Defense Secretary announced Project Meridian, a Pentagon effort on the future of warfare co-led by Elon Musk and Palmer Luckey, with findings due in 120 days. The effort spans domains from subterranean depths to cislunar space. Alex also described Project Agincourt as the Department of War’s first autonomous warfare command, a joint force dedicated to autonomous weapons, including drones. Richard argued that lethal decisions should remain under human oversight and singled out AI connected to nuclear weapons or codes as especially concerning. Palmer Luckey said undersea systems cannot continuously communicate with central command; once in “hunt mode,” they operate without reporting back. Speakers said Russia and Ukraine each plan to produce 10 million drones this year, compared with about 500,000 used between them two years into the conflict.

Socher said he saw no realistic scenario in which AI wipes out all humanity, while acknowledging the possibility of large-scale harm, potentially affecting or killing 100 million people. He argued that specific risks could be reduced through stronger cybersecurity, enforcement of existing bans on illegal gain-of-function research, and scrutiny of online content. A proposed Human Control Over AI Act would prohibit recursively self-improving models until federal guardrails exist and require shutdown controls and independent reporting; a panelist called prevention unworkable and warned of an “AI Stasi.” Speakers also noted that physical-world innovation depends on materials, supply chains, and hardware, and that even a perfect blueprint would not make building an ASML-like machine a two-year task.