Vinyals: no AI intelligence explosion, two bottlenecks stall self-improvement

Vinyals's talk came days after he left Google DeepMind, where he served as VP of Research and worked on AlphaStar, AlphaCode, and Gemini. His argument is less a prediction than a diagnosis of where AI self-improvement stalls. Labs today mostly infer self-improvement indirectly from capability benchmarks like SWE-Bench Pro or ML-Bench, but those tests mainly cover implementation and experimentation — the steps Vinyals says already work. More direct benchmarks are starting to appear, yet they are expensive: each evaluation requires an agent to work for hours on tasks far removed from the real goal, the way optimizing Tetris is not the same as automating a research lab.

Idea generation is equally underdeveloped. Vinyals calls the instinct for which ideas are worth pursuing "research taste," and notes nobody has really studied how to teach it in LLM training. He expects future evaluations to reward not just how much a system improves but how it gets there, using the criteria conference reviewers apply: originality, elegance, efficiency, and durability. Human review is expensive and not especially good at spotting strong ideas either.

Discovery Loop, his new startup, aims to automate the full scientific loop — hypotheses, experiments, evaluation — starting with AI research itself, with the company as its own first customer. Vinyals concedes idea generation remains the hardest part, so early on humans and machines will develop hypotheses together. Whether that hybrid approach actually moves the needle on the two bottlenecks is the open question, and the founders' stated ambition — that a handful of people could outproduce massive teams — rests on it.