Haoxing Du: Go AI Defeated with Only a Two-Stone Handicap — The 'Orders of Magnitude' Claim Needs Revision
In July 2026, Korean Go player Shin Jin-seo defeated the open-source Go AI KataGo 2-1 with only a two-stone handicap. Haoxing Du, who conducted an interpretability project on a Go-playing model in 2022, wrote an analysis of this result, arguing that while Go AIs are genuinely superhuman, the popular claim that they are 'orders of magnitude stronger than any human alive' does not hold. The deeper problem is that even in Go — a domain with a clear handicap metric — human assessment of AI capability has long been confused, which means our intuitions about superintelligence in other domains are even less reliable.
ImportanceMaterialEvidenceE3 inspectableWrite-upStandard
In July 2026, Korean professional Go player Shin Jin-seo defeated the open-source Go AI KataGo 2-1, accepting only a two-stone handicap advantage. This result surprised Haoxing Du — he had done an interpretability project on a Go-playing model in 2022 and had long treated Go AIs as an intuition pump for superintelligence in other domains. His core question: if even Go AIs are not 'orders of magnitude beyond humanity,' are our imaginings about superintelligence in other fields also built on a faulty analogy?
Why is this result surprising? Go is the canonical domain where AI models are considered to have most thoroughly surpassed humans. Chinese professional player Ke Jie called AlphaGo 'the god of Go' after losing to it in 2017; OpenAI researcher Noam Brown said on the Dwarkesh Podcast in September 2026 that Go models are 'orders of magnitude stronger than any human alive.' KataGo, initiated in 2017 by David Wu, a Go enthusiast at Jane Street, has been in self-play training for nearly a decade, with a self-play Elo exceeding 14,000. Yet Shin Jin-seo won with only a two-stone advantage.
After one day of research, Haoxing Du offers a three-factor explanation: first, Shin Jin-seo is among the greatest Go players in history and was among the first professionals trained with AI tools; second, KataGo's match configuration was likely weakened — it did not use pondering (idle-time thinking), had only about 20 seconds of compute per move, and its handicap parameter PDA was likely set too low; third, Shin spent months studying KataGo's behavioral patterns, while KataGo could not model a human opponent. He specifically notes that Shin did not win by exploiting adversarial examples or known vulnerabilities.