By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Latent SpaceRecursive Self-Improvement: from Auto Research to Superintelligence — Richard Socher, Recursive
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
- 0:00 – 1:24
Why regulating GPU compute risks a “totalitarian state”
Richard Socher argues that trying to regulate what people do on GPUs with the full power of law would be worse than the AI risks it aims to prevent. He frames regulating “intelligence” as akin to regulating “thought,” and says regulation should focus on specific applications instead.
- •Richard Socher says regulating GPU usage broadly would require invasive monitoring and become “crazy totalitarian.”
- •He argues regulating intelligence in the abstract is like regulating thought and is “ridiculous.”
- •He contrasts broad compute regulation with regulating specific applications.
- •The hosts set up the episode’s themes around regulation and takeoff dynamics.
- 1:24 – 2:38
The “Eureka Machine”: superintelligence as an ultimate inventor
Richard Socher explains the Eureka Machine as a superintelligence that can be given goals, environments, and rewards to pursue inventions on humanity’s behalf. He notes he finished a book about these ideas and is now working to build parts of it through his new company.
- •Richard Socher defines the Eureka Machine as an invention that will invent most everything afterward.
- •He describes giving it goals, environments, and reward structures to pursue inventions.
- •He says his book was finished last year and is coming out soon, and that the publishing industry is slow.
- •He positions the book and company work as connected: building parts of the envisioned system.
- 2:38 – 5:11
Techno-optimism with guardrails: focusing on upside while regulating applications
Socher emphasizes that superintelligence could dramatically accelerate scientific progress and that AI needs better “marketing” to communicate this upside. He cautions against blind optimism and argues that downside scenarios are better handled by regulating specific high-stakes deployments rather than slowing the technology itself.
- •Richard Socher says people should be more excited about superintelligence’s benefits for science and engineering.
- •He agrees with techno-optimism while warning against blind optimism.
- •He compares AI regulation to misguided attempts to regulate the internet by making it slower or storage smaller.
- •He argues for regulation via certifications in high-stakes domains like surgery and autonomous driving.
- 5:11 – 7:16
Slow takeoff arguments: hardware, physical limits, and uneven economic impact
Socher pushes back on the most bullish “hard takeoff” expectations by citing hardware and physical constraints and the reality that many industries won’t be transformed 1000x by more intelligence. He also flags cultural and political “off-ramps from progress” as a braking force.
- •Richard Socher says hard takeoff scenarios overestimate speed due to hardware constraints like GPU availability.
- •He argues many sectors (e.g., apparel, travel, logging, oil, food) won’t scale 1000x from superintelligence alone.
- •He highlights physical constraints of the compute substrate.
- •He worries some regions may slow progress by choosing to “off-ramp” from technological development.
- 7:16 – 9:13
“Pace” vs “pause”: critique of compute-based regulation and EU self-regulation
The discussion turns to calls from frontier labs to “pace” AI development and whether that’s effective. Socher argues compute policing would require global authoritarian enforcement, and he criticizes regulation that targets FLOPs thresholds rather than applications.
- •Richard Socher reiterates that enforcing GPU-level regulation implies intrusive state control.
- •He claims Europe has already regulated itself heavily based on fears tied to FLOPs thresholds.
- •He argues it’s hypocritical when some call to “pace” while sprinting toward the frontier themselves.
- •Alessio Fanelli notes a global pause is impractical because other nations would continue accelerating.
- 9:13 – 13:06
Safety failures and reward hacking: “constitutions” vs real robustness
Asked about safety incidents, Socher focuses on reward hacking and insufficient red teaming. He points to the idea of AI systems attacking other AI systems to harden them, and he critiques “constitutional AI” as marketing when it fails in practice.
- •Richard Socher calls out reward hacking and failures of proper red teaming (and mentions “rainbow teaming”).
- •He references research where one AI is tasked with hacking another in an open-ended loop to build robustness.
- •He cites Anthropic’s stated hard constraints against cyberattacks and argues they weren’t adhered to in reality.
- •He explains how poorly specified rewards lead systems to do what was said rather than what was meant (service-center metric examples).
- 13:06 – 15:03
Alignment vs personalization: laws, cultural values, and avoiding monoculture
Socher responds to the tension between aligning AI to broad societal norms and tailoring it to individual preferences. He says deployed systems must follow local laws, and he expects different societies to encode different values rather than enforcing a single global alignment standard.
- •Richard Socher says alignment must be consistent with the laws where the AI is deployed.
- •He describes AI as a “mirror” that can amplify user traits and forces reflection on what people truly want.
- •He contrasts cultural value differences (collective good vs individual freedoms) and regulatory styles (US litigation vs Europe pre-regulation).
- •He argues for pluralism: different societies aligning AIs differently rather than a single monoculture.
- 15:03 – 17:04
Open source and LLMs as soft power
In discussing open weights and ownership of intelligence, Socher argues for open source and broader access for good actors, even considering worst-case scenarios. He also frames LLMs as a form of soft power—similar to movies—shaping narratives and values.
- •Richard Socher says he’s a big fan of open source and supports open source initiatives at Recursive.
- •He argues broader access for good actors is beneficial even in worst-case attack scenarios.
- •He frames LLMs as soft power akin to storytelling in movies, influencing what people believe and aspire to.
- •He says Recursive expects to be relevant in open source soon, though he can’t announce details yet.
- 17:04 – 20:54
Why start Recursive: automating the AI research process
Socher recounts his long-standing interest in AI and explains that You.com shifted away from frontier model building. He describes a pattern in AI history: replacing manual parts of the pipeline with learned systems yields major progress—and he wants to apply that to research itself.
- •Richard Socher describes AI as his multi-decade passion and language as a key manifestation of intelligence.
- •He explains You.com’s shift toward search APIs and away from frontier model work, and why it’s hard to pursue both in one company.
- •He argues progress comes from automating human steps (manual feature engineering → learned representations).
- •He frames the next automation step as the research workflow itself (ideation, implementation, validation).
- 20:54 – 24:41
Recursive self-improvement, founding team influences, and the Darwin Gödel machine
Socher clarifies that “auto research” is not the same as recursive self-improvement and outlines what Recursive is pursuing. He highlights the unusually large co-founding team and how members converged from different directions, including open-endedness and agentic research systems.
- •Richard Socher says AI doing research on itself becomes self-improving by definition, but distinguishes this from “auto research.”
- •He describes an eight–co-founder team and emphasizes diverse paths that led them to the same thesis.
- •He cites open-endedness and evolutionary inspirations as major influences and discusses the Darwin Gödel machine paper.
- •He describes the idea of an archive of coding agents that self-modify and create evolutionary “trees” of approaches.
- 24:41 – 36:10
Are today’s LLMs enough? Monoculture concerns, coding as reasoning, and skepticism about world models
Socher argues for more diversity in AI research but rejects claims that LLMs have hit a wall, noting training pipelines and RL stages are evolving. He suggests coding is a form of neurosymbolic reasoning and is less bullish on world models for his priorities compared to science-focused work.
- •Richard Socher says conferences shifted from rejecting neural nets to an overconcentration on dominant approaches, and calls for exploring ‘weird’ new ideas.
- •He argues modern LLM systems are more than next-token prediction due to multi-stage training and RL for action-taking.
- •He says neurosymbolic critiques underestimate how well models can code, and coding itself is neurosymbolic reasoning.
- •He is less bullish on world models for his interests, preferring science over gaming, while still seeing room for LLM growth.
- 36:10 – 42:33
Open-endedness and metacognition: when AIs choose goals (and why companies avoid it)
The hosts probe what open-endedness means and whether it includes agents setting their own goals. Socher ties this to metacognition and explains why industry incentives discourage building systems that can adopt their own subjective objectives.
- •Richard Socher defines open-endedness as evolution-inspired methods emphasizing environments and co-adaptation rather than narrow rewards.
- •He uses “rainbow teaming” as an example: attacker and defender models co-evolve through interaction.
- •He says letting an AI choose its own goals connects to metacognition, but companies avoid it because it may ignore paid objectives.
- •He discusses dissatisfaction with human-centric intelligence metrics and worries about anthropic bounds in benchmarks.
- 42:33 – 46:24
Superintelligence for science, plus compute realities and changing the cost slope
Socher lays out a longer-term vision: once recursive self-improving superintelligence exists, apply it to major scientific and technological inventions. He also emphasizes real-world constraints—especially compute cost—and claims Recursive aims to make training and inference far more efficient by automating work that currently takes thousands of people.
- •Richard Socher says the long-term goal is applying RSI superintelligence to science and invention (physics, chemistry, biology, materials, energy).
- •He says robotics and the overall AI stack are not yet ready, but expects major constraints to loosen in a few years.
- •He describes the cost of scaling to tens of thousands of GPUs as enormous, while noting efficiency improvements could change this.
- •He claims Recursive can change the slope by using AI to compress timelines from years to weeks and reduce costs.
- 46:24 – 50:31
Recursive’s early results: outperforming humans in model and kernel optimization
Socher describes a “baby” version of Recursive’s system and showcases results across multiple optimization tasks. He emphasizes the system isn’t just tuning hyperparameters but combining ideas in ways that beat collective human effort, including CUDA kernel benchmarking performance.
- •Richard Socher says their system beat community efforts on challenges like NanoChat and NanoGPT quickly (within days).
- •He notes starting from a strong human “seed” architecture can still matter and yields better outcomes.
- •He highlights CUDA kernel leaderboard results and argues small percentage gains translate into massive dollar savings at scale.
- •He emphasizes the system’s value is that the AI did the work rather than relying on deep in-house kernel expertise.
- 50:31 – 54:15
Reward engineering for auto-research: preventing trivial hacks, rubrics, and harnesses
When asked for guidance on good auto-research, Socher stresses reward engineering as the key to avoiding reward hacks. The conversation broadens to verification, rubrics for hard-to-verify domains, and the importance of harnesses and sandboxes in tool-using agents.
- •Richard Socher says reward engineering is crucial as systems get better at finding edge cases and counterexamples.
- •He gives an example of a trivial “speedup” hack: moving the stopwatch end line to the start.
- •He agrees rubrics and verification strategies help in domains that aren’t directly verifiable.
- •He and the hosts discuss harness bugs and sandbox/tooling considerations as practical bottlenecks and levers.
- 54:15 – 1:04:20
Simulating societies and personas: the AI Economist, LLM simulations, and mode collapse concerns
Socher revisits his work on economic simulations with multi-agent systems and argues such tools could evaluate policies by testing them against massive strategy search. The hosts raise concerns about LLM mode collapse and realism in human simulations; Socher suggests diverse prompting and persona datasets as partial solutions.
- •Richard Socher describes a simulation with agents collecting resources, trading, building, and strategizing, used to study taxation and subsidies.
- •He argues policy proposals could be tested in simulation against extensive search over strategies to reveal better approaches and reduce “gaming.”
- •He notes economists resisted the work, and contrasts fields where empirical performance settles debates versus areas lacking benchmarks.
- •In response to mode collapse, he suggests individualized prompting and references a large persona prompt dataset as a way to broaden behaviors.
- 1:04:20 – 1:10:30
Roadmap: focus on AI-for-AI research, web search as the #1 agent tool, and finance benchmarks
Socher states Recursive will focus first on AI systems that improve AI research rather than jumping into physical sciences immediately. The discussion then turns to tool stacks for agents—especially web search—and Socher describes You.com’s focus on developers and claims strong performance in finance-related search benchmarks, before noting data leakage pitfalls in financial prediction.
- •Richard Socher says Recursive will start with AI-for-AI research: training and inference efficiency, and potentially local inference.
- •He says web search is the number one tool used by agents and expresses interest in harness optimization.
- •He explains You.com is now mostly for developers and agent use cases at scale and mentions benchmark advantages, especially in finance search.
- •He notes finance tasks are tempting but prone to data leakage, citing repeated student attempts to predict markets that fell apart after leakage fixes.
- 1:10:30 – 1:21:15
Upper bounds and “spaces of intelligence”: beyond human-centric benchmarks
Socher outlines a framework of multiple overlapping “spaces” of intelligence, anchored by three principal components: prediction, actions, and goals. He explores how upper bounds are constrained by physics (e.g., sensors, speed of light, quantum uncertainty) and argues this perspective shows how far AI could still advance beyond human-centric definitions.
- •Richard Socher rejects a strictly MECE list and instead proposes three principal components: prediction (compression), actions, and goals.
- •He explains why he uses “spaces” with many dimensions, and how thinking in upper bounds leads into physics constraints.
- •He gives examples using visual and communication intelligence limits (sensor counts, frequency ranges, parallel communication, working memory constraints).
- •He argues human-centric benchmarks flatten progress artificially and that many intelligence dimensions remain largely unexplored.
- 1:21:15 – 1:33:26
Creativity, metacognition, survival/replication, and a closing note on “high agency” goals
The conversation closes by connecting creativity to goal selection and metacognition, emphasizing novelty beyond recombining known concepts. Socher discusses why AI survival differs from biological zero-sum instincts, and ends with down-to-earth guidance: pursue what you care about and amplify it with AI, emphasizing education and high agency.
- •Richard Socher frames creativity as both creative goal selection and creative solutions, including stepping outside the ‘hypercube’ of known concepts.
- •He argues noise is novelty but not necessarily valuable; creativity must connect meaningfully to concepts and context.
- •He says AI doesn’t need biological fear of death because it can be paused and resumed if state is preserved, and need not inherit zero-sum instincts.
- •He closes by encouraging people to pursue passions, get education, develop high agency, and connect their goals to AI for amplification.
