Richard Socher says recursive self-improvement might arrive in one or two years, but he rejects overnight-takeoff fears. Bigger ideas require bigger models and more compute, and even pure-software experiments cannot train a five-trillion-parameter model in five minutes; iteration cycles remain, and physical constraints and supply chains also matter.
Socher says AI research is stuck at a meta level where humans still generate ideas, use AI to implement them, and validate results. Recursive aims to automate that full loop—idea generation, implementation, and validation—inside human-set goals, rewards, and environments, reaching what he calls strong recursive self-improvement.
Socher argues AI's offbeat hypotheses can be useful because they explore a vast combinatorial space of ideas. But hallucinations become a bug for factual questions; the Eureka machine still needs grounded search with citations, like You.com, to check facts against existing human knowledge. He calls hallucinations sometimes feature, sometimes bug.
Socher says the Eureka machine rests on four pillars: existing world knowledge, scientific measurements, simulations, and real-life experiments, all orchestrated by an agent swarm. He expects this automated system to run through those steps more efficiently than human scientific communities and to verify discoveries before trusting them.
Socher compares biology to linguistics before large neural nets: once enough data exists, handcrafted rules matter less. He says understanding proteins and biological systems will let researchers engineer them, turning biology into an engineering science—a transition that may feel sad but should excite anyone who cares about impact.
Socher says economics is still ideological and built on one-step models with beautiful math but no adaptation. His AI Economist used two-level reinforcement learning, with tax-and-subsidy proposals tested against self-interested agents, to simulate policy outcomes. He wants policy recommendations made more objectively in the future.
After a professor used Claude Code to compress years of planning work into an hour, Socher says researchers will shift toward setting goals and managing agent collaborators. He compares the future role to an NSF program director and says universities must teach students to generate good ideas and use these tools well.
Socher predicted AI would become superhuman at coding because code is simulatable and verifiable. He notes legendary programmers now adopt coding agents, and argues other fields need to become more simulatable or verifiable to get similar AI gains. Programming also remains a useful way of thinking.
Socher says apocalyptic AI fears repeat earlier scares around overpopulation, nuclear energy, and GMOs, which he says held humanity back. He still favors targeted controls—chatbot ratings or FDA-style drug oversight—but warns vivid imagination about the apocalypse can obscure AI's potential benefits and dismiss progress.
Socher cites Perilbio, which uses tiny organoids to test drug interactions with human immune and neural systems. He says these tests are more predictive than mouse studies, and Perilbio received FDA approval to skip animal trials for immune-therapy toxicity testing, saving animal lives and speeding drugs to market. Organoids are smaller than a bee.
Socher says anyone can vibe-code an Instagram clone, but nobody is on your little app, so the network effect is missing. Software built on unique data, marketplaces, network effects, and strong brands will remain relevant, along with jobs in entertainment, health, and research, despite better coding agents.
Asked whether AI can achieve zero-shot innovation, Socher says most breakthroughs combine smaller past innovations. Einstein's relativity sat atop calculus and other prior work. He argues vast underexplored combinations remain, so AI can generate novel-seeming ideas, though people underestimate how much builds on existing knowledge, like new proteins from known amino acids.
