February 17, 2026
Checkmate—or check the hype?
Contra "Grandmaster-level chess without search" (2024)
Grandmaster bot “without thinking”? Commenters call check on the hype
TLDR: DeepMind’s chess bot trained on Stockfish hit a sky-high online blitz rating and claims “Grandmaster-level” play without heavy calculation. Commenters are split: skeptics say it’s old tricks and clever branding, defenders say it’s a solid proof-of-approach, and everyone’s arguing whether “no search” really means no thinking at all.
DeepMind dropped a flashy claim: a chess bot trained to mimic Stockfish can play at human Grandmaster level—without doing the usual deep “thinking” search. It blitzed to a 2895 rating on Lichess, which is bonkers high. But the internet instantly split like a forked king. Critics argue it’s mostly “copying the teacher” and not new, pointing to the open-source Leela Chess Zero project that already pushed this style to wild heights. The blog post dragging the paper calls it “not serious” and blasts the team for leaning on human masters to judge games—when top engines are stronger than any human.
Defenders clap back: the paper never said it’s the strongest bot, just that their method gets to a legit level with a different setup. Commenter mquander basically says, calm down—this is a proof-of-approach, not a world championship. Meanwhile, meme-lords pile in with jokes about a “no-think chess god” and “50ms Stockfish intern” doing the real work. Others nitpick the “no search” label, noting the bot still does a tiny one-move lookahead, and wonder if its magic fades at longer time controls where humans and engines can think longer.
Bottom line: impressive blitz numbers, big “but actually” energy, and a comment section doing its favorite opening—the Sicilian Drag-on-DeepMind.
Key Points
- •DeepMind’s transformer chess model is trained to imitate Stockfish, outputting state value, action-values, and a policy distribution based on 50ms searches.
- •The approach closely resembles AlphaZero networks (policy and value), with the action-value output being a new addition.
- •Model strength was assessed via policy-only play and depth-one rollout using the value head; the latter is effectively a minimal search.
- •The paper reports a Lichess Blitz rating of 2895, supporting a claim of Grandmaster-level play in blitz, with noted time-control limitations.
- •The article argues the contribution may not be new, citing Lc0’s stronger policy-only networks (BT4 vs T30/T40), and questions parts of the evaluation and analysis approach.