Can AI reach superhuman research ability before tackling physical science?

Richard Socher, discussing his company Recursive, states its strategy plainly: "we have to start with AI for AI and then make it really good at doing research on creating better AI so that it has the equivalent of 50,000 PhDs in terms of knowledge and its own capabilities, and only then go after the physical natural sciences like physics, chemistry, and biology." He ties this sequencing to a checkability criterion: "I think we can basically predict where AI will certainly have superhuman capabilities. And those are all scenarios and all domains where we can either have a simulation and/or a verification tool," because such a domain lets "the AI...essentially infinitely many times experiment inside that simulation," and "assuming this simulation doesn't take years to run every time," the problems in that domain get solved.

The reasoning draws on what he calls the four pillars of the "Eureka Machine": human knowledge and LLMs, measurements, simulation, and "robotic process automation to collect even more data and verify whether the inventions really made sense." AI research is the pillar that fits the simulation/verification case most cleanly — benchmarks and training runs are fast, repeatable, scorable — which is why it comes first in the sequence. Physical science, especially biology, depends on the fourth pillar, physical data collection, which does not speed up just because the model doing the reasoning gets smarter.

This domain-level checkability criterion is a coarser, applied version of what Can AI systems invent new concepts rather than reuse trained ones? frames as the verifier gap at the level of individual representational primitives: Socher's question is whether a whole domain can be checked fast at all, not whether a single new concept's value can be judged before reuse. The sequencing itself also parallels the split named in Can AIs learn to specify their own research objectives? between "autoresearch," where "the objective is already specified very cleanly," and open-ended science, where neither humans nor AIs can specify the objective. Socher treats "build better AI" as the specified, measurable target — the same kind of case that note's speaker calls "an extremely measurable, verifiable task" — and treats natural-science questions like curing cancer as the open-ended case, bottlenecked by years-long trials rather than by model capability.

The excerpt gives no benchmark, metric, or timeline for what "50,000-PhD-equivalent" research capability would mean or how Recursive would know it had been reached, and it does not establish that AI research is in fact as cleanly specified as autoresearch — discovering a genuinely new architecture may be closer to the open-ended case. Socher also explicitly rejects a "hard takeoff," attributing the limiting factor to "physics and constraints in the real world" such as multi-year drug trials, not to any claim that AI-for-AI research itself is capped; the sequencing argument and the takeoff-speed argument are two separate claims made in the same answer, not one derived from the other. At the strength the evidence allows, Recursive's roadmap is a bet that AI research is the checkable domain it needs to be — the excerpt does not show that bet has been tested.