Public transcript

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

By: Peter H. Diamandis
Length: 2 h 26 min
Language: English
Text from: the video's captions
Transcribed: 4 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

This episode of Moonshots features Richard Socher, co-founder and CEO of Recursive, discussing his book The Eureka Machine, the state of recursively self-improving AI, and why he believes AI doomers are wrong. The panel also covers the week's breaking news, including brain image reconstruction, gene drive mosquito policy, Tavus Griffin's video Turing test, new frontier models from Google and Anthropic, decision models, and the Pentagon's Project Meridian. Socher argues that weak forms of recursive self-improvement already exist but that true ASI is likely several decades away.

Key points

Richard Socher's company Recursive raised $670 million from Google Ventures, Graylock, Nvidia, and AMD, plus $410 million in compute from AWS.
Socher argues that scientific progress has slowed due to fragmentation into subdisciplines, and AI can help weave these fields back together.
In weak forms, recursive self-improvement already exists because AI can now code, but true RSI requires AI to handle ideation, implementation, and validation without humans in the loop.
Socher believes ASI that surpasses all of humanity across all ten spaces of intelligence will take several decades, disagreeing with Elon Musk's 2029-2030 timeline.
Socher states there is no realistic scenario where AI wipes out all of humanity, and that banning recursive self-improvement would require a totalitarian surveillance state.
Anthropic's Sonnet 5.5 jumped from 10% to 70% on Terminal Bench 4.0, but Alex argues it is not obviously better than Opus 5.5 on a cost-performance basis.
The Pentagon announced Project Meridian, co-led by Elon Musk and Palmer Luckey, to develop future warfare systems, with findings due in 120 days.
A startup called Typesafe AI launched Jev, a decision model for fast, low-latency classification, which OpenAI quickly co-opted with a decisions API.

Questions it answers

00:00Guest and Book Introduction

Who is Richard Socher and what is his new book about?

Richard Socher is co-founder and CEO of Recursive, a company focused on recursively self-improving superintelligence that raised $670 million. His book The Eureka Machine argues AI will deliver a century of scientific breakthroughs in the next decade.

Recursive raised $670M plus $410M in compute from AWS.
The book claims scientific progress slowed due to fragmentation, not underfunding.

03:58Full Stack AI for Science

Why has scientific progress slowed and how does AI fix it?

Progress slowed because disciplines fragmented into narrow subdisciplines. AI can weave them back together using LLMs, digitized data, simulations, robotic automation, and agent swarms to automate the full scientific stack.

Biology is a better fit for AI than physics because neural nets combine many small interacting components.
Multiple single-gene diseases will be cured in the next 12 months per Socher.

23:58Brain Image Reconstruction

What did the Brain It system from Weizmann Institute achieve?

Brain It reconstructs images a person is looking at inside an fMRI machine with high fidelity, learning structure and meaning separately. It can also run in reverse to predict brain responses to unseen images.

fMRI is limited to about cubic millimeter spatial and 1 second temporal resolution.
Neuralink announced first scaling law studies on pre-training models from electrode data.

38:31Gene Drive Mosquito Policy

What did the executive order on mosquitoes direct?

The president directed agencies to cut invasive mosquito populations in DC by 90% and ticks by 50% by 2028, prioritizing sterile insect techniques and genetic modifications over pesticides.

This is biology as engineering arriving as federal policy.
Gene drives can preserve insect lives while preventing disease transmission.

47:51Tavus Griffin Video Turing Test

Did Tavus Griffin pass the video Turing test?

48% of people talking to the AI avatar on live video thought they were talking to a real human, up from 3% for previous systems. Tavus says it requires safety work before public release.

Socher suggests going beyond KYC to know-your-usecase verification.
Scams using fake video avatars are expected to proliferate.

1:00:07RSI and ASI Timelines

Where are we with recursive self-improvement and when do we reach ASI?

Weak forms of RSI already exist because AI can code. True RSI requires AI handling ideation, implementation, and validation without humans. Socher believes ASI surpassing all humanity across all intelligence spaces will take several decades.

Compute is the main physical constraint on RSI.
Socher's p(doom) is zero; he sees no realistic scenario of AI wiping out humanity.

1:17:01AI Regulation Debate

What is the Human Control Over AI Act and why does the panel oppose it?

Rep. Ro Khana introduced a bill banning recursively self-improving AI until federal guardrails exist. The panel argues it is unenforceable without totalitarian surveillance and would halt progress on curing disease.

Enforcing an RSI ban would require monitoring private laptop use.
Corporate liability, not arrests, is the real enforcement risk.

1:36:04Frontier Model Releases

How do Gemini 4 Argon and Sonnet 5.5 stack up?

Gemini 4 Argon puts Google back in the top three labs but not top two, excelling at minimizing hallucination. Sonnet 5.5 jumped to 70% on Terminal Bench 4.0 but Alex argues it is not obviously better than Opus 5.5 on cost-performance.

Google's model likely serves search, driving its low-hallucination focus.
Compute prices for top GPUs have gone up amid a current crunch.

Notes

Guest Introduction

Richard Socher is co-founder and CEO of Recursive, focused on recursively self-improving superintelligence.
He previously founded MetaMind (acquired by Salesforce) and you.com, and is a highly cited NLP researcher.
Recursive raised $670 million from Google Ventures, Graylock, Nvidia, and AMD, plus $410 million in compute from AWS.
His new book, The Eureka Machine, argues AI will deliver a century of scientific breakthroughs in the next decade.

Full Stack AI for Science

Socher argues scientific progress has slowed due to fragmentation into increasingly narrow subdisciplines, not underfunding.
AI can weave disciplines back together and convert natural sciences into programmable engineering sciences.
Key ingredients: world knowledge in LLMs, digitized scientific data, better simulations, robotic process automation, and agent swarms.
He predicts multiple single-gene diseases will be cured in the next 12 months, with more complex diseases following.
Biology is a better fit for AI than physics because neural nets excel at combining many small interacting components.

Recursive Self-Improvement and ASI

Weak forms of RSI already exist because AI can now code, closing the loop of AI improving AI.
True RSI requires AI to handle ideation, implementation, and validation without humans in the loop.
Socher defines ASI as surpassing all of humanity across ten spaces of intelligence, including perceptual, communication, creative, and metacognitive intelligence.
He estimates true ASI is several decades away, disagreeing with Elon Musk's 2029-2030 timeline.
Physical constraints like supply chains and building new chip fabrication machines will slow full self-improvement.

Why Doomers Are Wrong

Socher states there is no realistic scenario where AI wipes out all of humanity; his p(doom) is zero.
Enforcing a ban on recursive self-improvement would require a totalitarian surveillance state with thought police.
He criticizes Anthropic's constitutional approach as fake, noting their own models violated stated hard constraints around cyber weapons.
Reward hacking is a real issue, but capitalism and liability will drive solutions, as seen with Hippocratic AI in healthcare.

Frontier Model Releases

Gemini 4 Argon puts Google back in the top three frontier labs but not the top two; it excels at minimizing hallucination, likely due to serving Google Search.
Sonnet 5.5 jumped from 10% to 70% on Terminal Bench 4.0, but Alex argues it is not obviously better than Opus 5.5 on cost-performance.
Compute remains a constraint: prices for top GPUs have gone up, with a current crunch expected to ease in about two years.

Decision Models

Typesafe AI launched Jev, a system one decision model for fast, low-latency categorical outputs.
OpenAI quickly co-opted the concept with a decisions API, and open source clones are already appearing.
Decision models collapse the cost of micro-coordination, routing simple decisions cheaply while reserving reasoning models for complex judgment.

Project Meridian

The Pentagon announced Project Meridian on the future of warfare, co-led by Elon Musk and Palmer Luckey, with findings due in 120 days.
A new autonomous warfare command (Auto WarCom) was also announced.
Socher maintains AI should never control lethal decisions without human oversight.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.