12 May 2026 · LinkedIn
OpenAI's chief scientist gave an interview to MIT Technology Review the other week describing the firm's new "north star". The goal is a fully automated AI researcher. An autonomous research intern by September of this year. A multi-agent research system, in his words "a whole research lab in a data center", by 2028.
Dario Amodei, in his January essay The Adolescence of Technology, described the same target in different language. A "country of geniuses in a datacenter". 50 million minds, each more capable than any Nobel laureate, running at superhuman speed. He put the timeline at one to three years. He also noted, plainly, that AI is now writing much of the code at Anthropic, and that the point at which the current generation of AI autonomously builds the next may be only one or two years away.
Two of the three frontier labs have now publicly committed, by name and by date, to building a system that does original research without human direction.
The press has framed this as a story about scientific discovery. Proofs, biology, materials science. That framing is too narrow. Research is not a department of the economy. It is the underlying capability that produces every other complex knowledge function. Strategy, law, finance, medicine, policy, drug development, software architecture, consulting, executive judgement - all of it is downstream of the ability to take an unstructured problem, generate hypotheses, test them against evidence, and converge on something usable.
Software engineering compressed first because the labs needed it to build themselves. The same dynamic is already visible at Google, where Pichai disclosed in April that around three-quarters of new code is now AI-generated. The function that accelerates the construction of the next model was always going to be the function that compressed first.
What the labs are now pointing at is a level above that. The capability that sits underneath everything humans get paid the most to do. Not execution. Not throughput. The capacity to work out what to do when nobody yet knows.
If they succeed even partially, the result does not stay inside the laboratory. It is sold, embedded, deployed. Every white-collar function whose feedback loops have been treated as too loose, too contextual, or too judgement-heavy for AI to handle is suddenly within range. The system that does the research is the same system that figures out how to do the rest.
The labour conversation has spent three years arguing about coders, drivers, radiologists, paralegals. The capability now being built is for the part of the economy where the answer was not yet known.
The firms putting the timeline on the table also own the compute it runs on.
Join the conversation on LinkedIn →
← All writing
