UCLA's Quanquan Gu and AlphaFold's sokrypton Clash Over Whether AI Removes Serendipity in Science

On September 13, UCLA professor Quanquan Gu and AlphaFold team member Sergey Ovchinnikov (sokrypton) engaged in a multi-round debate centered on Terence Tao's view of AI in science: whether AI brute-forcing problems on humans' behalf strips away the detours where major discoveries are born. The two sides reached no consensus, but the exchange touched the core question of AI for Science.

Confirmed

• The debate originated from Tao's view that the detours mathematicians take and the new problems they encounter while solving problems are often the source of major breakthroughs—something AI solving problems for humans would eliminate.
• sokrypton responded that he does not deny enough AI agents, compute, and robots could brute-force well-defined problems or assemble many researchers' work into a large pipeline, but asked "then what?", citing Tao's interview to question the value of such brute-force science.
• sokrypton added that AI is still far from "asking the next question," which remains a uniquely human ability; per Tao, new questions emerge precisely in the process of answering the initial one.
• sokrypton framed the disagreement as a "super-machine paradox": we have a super machine that can answer any question, expecting science and math to accelerate—yet progress may stall because we would no longer have new questions to ask.
• Gu countered that detours are not uniquely human: AI also explores failed paths and proposes intermediate conjectures when solving hard problems; even a brute-forced cancer cure would be worth pursuing. Responding to sokrypton's joke that AI will eventually replace all scientists and even carbon-based life, Gu argued that solving hard scientific problems faster does not mean replacing humans.
• Commentator curiouswavefn added another angle: the fundamental difference between AI doing math and humans doing science lies in problem definition—math problems are well-defined, while in broad scientific research humans often do not even fully know what problem they are trying to solve.

Why it matters

The debate touches the fundamental question of AI-era research: when AI can brute-force well-defined problems with compute and scale, does the value of science lie in "answers" or in "the process of seeking"? Tao and sokrypton stress the irreplaceability of serendipity, new questions, and detours, even worrying an all-capable machine could stall science; Gu emphasizes outcome orientation and AI's own exploratory capacity; curiouswavefn pinpoints problem definition as the core bottleneck of AI for Science. These positions represent typical current academic attitudes toward AI for Science, offering a useful reference for understanding AI's role in mathematics, biomedicine, and beyond.

2026-09-13 ~ 2026-09-13 · 9 related posts

• [source] sokrypton questions brute-force AI science: agents and compute can merge papers, but then what? — sokrypton · 2026-09-13
• [source] Quanquan Gu Responds to Terence Tao's Doubt: AI Brute-Forcing Cancer Therapies Is Worth It — QuanquanGu · 2026-09-13
• AlphaFold researcher: AI robs us of the detours that lead to major discoveries — sokrypton · 2026-09-13
• Mathematicians Debate Whether AI Solving Hard Problems Robs Us of Serendipitous Detours — QuanquanGu · 2026-09-13
• sokrypton jokes AI will eventually replace all scientists and carbon life — sokrypton · 2026-09-13
• 'Faster science doesn't mean replacing humanity': AI endgame debate — QuanquanGu · 2026-09-13
• sokrypton: AI Still Can't Ask the Next Question — Why Science Progress May Stall — sokrypton · 2026-09-13
• [source] The Super-Machine Paradox: AI Answers Everything, Science Stalls — sokrypton · 2026-09-13
• The key gap: AI solves well-defined math, but science doesn't know its questions — LucaAmb · 2026-09-13