When AI Improves Itself | Richard Socher (Recursive)

Summary

Richard Socher makes the case that scientific progress has slowed because knowledge is fragmented across too many specialties, and that AI can help recombine it into a new discovery engine. He describes Recursive’s “Eureka machine” as four pillars: large language models, scientific measurements, simulation, and robotic process automation/experimentation. Socher argues next-token prediction can absorb structure across text, proteins, chemistry, images, video, and sound, while hallucinations can be useful when the objective is novelty rather than factual recall. He thinks AI will materially accelerate coding, biology, drug discovery, cancer research, fusion control, and materials discovery, but biology will still require more perturbation data, organoids, clinical validation, and real-world time. The episode closes with a broader view of intelligence as prediction, action, and goals, and with Recursive’s focus on compute-intensive, productized AI research systems rather than vague “neoclabs” ideas.

Notes

• Socher frames AI as the tool that can recombine the pieces science has split into hyper-specialized fields.
• He links slowed progress to fragmentation across many journals and subfields.
• Recursive’s thesis is that anything simulatable becomes a target for AI-driven recursive self-improvement.
• Next-token prediction is presented as a general mechanism for learning structure, not just language.
• He uses geography as a concrete example: repeated prediction teaches relationships like New York, northbound travel, and Boston.
• Tokens can represent proteins, image patches, video patches, or sound, extending the same machinery across modalities.
• Recursive’s architecture is described as four pillars: LLMs, measurements, simulation, and robotics for verification and data collection.
• The system starts from human knowledge plus LLMs as the first pillar.
• Socher says closed-model knowledge has been distilled back into open source by Chinese companies, creating a feedback loop.
• The second pillar is a model of reality built from scientific measurements beyond human perception.
• He argues that a real foundational model of reality will require collaboration across physics, chemistry, biology, and larger systems.
• The simulation pillar should use layers of abstraction instead of trying to simulate everything at full quantum fidelity at once.
• He sees today’s simulation stack as partial and uneven, from chess and Go to a still-distant virtual cell.
• The fourth pillar is real-world experimentation: labs, satellites, and other measurements when simulation is insufficient.
• He says self-driving robotic labs are starting to look practical within about 2 to 3 years.
• Organoids and human-derived stem-cell systems are part of the path to faster experimentation and reduced animal testing.
• The agent-swarm layer is meant to span as much of the scientific process as possible and recombine ideas in open-ended ways.
• Compute is the main bottleneck, and future systems may allocate compute by problem importance.
• Recursive was founded to build recursive self-improving superintelligence for knowledge discovery and science discovery.
• Socher wants AI to reach the equivalent of roughly 50,000 PhDs in knowledge and capability before tackling harder natural science domains.
• Biology is still under-dataed, so Recursive expects to need organoids and perturbation studies before a usable virtual cell exists.
• Hallucinations are framed as a feature for novelty-seeking scientific exploration, not as a universal bug.
• When factual correctness matters, Socher says the system should be grounded with search or other retrieval.
• He says biology is shifting from a descriptive field to a programmable engineering science.
• Protein language models were an early aha moment for the team, especially the 2018 ProGen work.
• He points to Profluent’s protein design work as evidence that programmable proteins can become real therapeutics.
• AI should help design cancer-specific drug cocktails and RNA sequences because cancers are heterogeneous.
• He argues biology will move faster, but not with a sudden hard takeoff because real experiments still take time.
• AI could also help with fusion control, solar cells, batteries, and new materials.
• He even imagines engineered bacteria that consume microplastics, with strict safety caveats.
• Socher sees economics as hard to benchmark and says that makes academic evaluation politicized.
• He describes a 2018 economics simulation with agents, resource collection, house-building, monopoly behavior, and a tax-and-subsidy meta-agent.
• The reward function in that simulation balanced equality and productivity multiplicatively.
• The system recovered the Saez formula baseline from a more realistic temporal setting.
• Policy AI suggestions should stay under human oversight.
• He argues AI systems are only as good as the people, data, infrastructure, and reward design around them.
• Recursive’s narrow system can already beat months or years of human AI research on some problems.
• It can also generate new CUDA kernels to speed up inference.
• Socher thinks many “neoclabs” are ideas rather than real companies if they do not ship concrete products.
• He defines intelligence around prediction, action, and goals.
• He says intelligence spans multiple spaces, including visual, language, physical, and robotics domains.