Skip to content

Google employees kept a report predicting DeepMind's decline on their desks, reading it daily. Three months later, SemiAnalysis apologized, admitting they had misjudged the situation.

Nov 20, 14:30

Beating AI News Flash: Arthur Douillard, a member of Google Gemini's reinforcement learning team, revealed that they printed out SemiAnalysis's report that disparaged DeepMind and kept it on their desks, looking at it every day. The outside world all thought they had fallen behind and could never catch up, but they were determined to prove them wrong.

The report was published in August this year, titled "Gemini is Cooked but GCP is Cooking." At the time, SemiAnalysis believed that DeepMind's top talent had left in large numbers, Google Cloud had taken away a large amount of computing power, and model development had fallen behind OpenAI and Anthropic. They even directly asserted that DeepMind was no longer a top AI lab and that it was basically impossible for it to catch up with the first tier again.

As a result, just over three months later, Google continuously released four models: Gemini 4 Argon, Carbon, Hydrogen, and Neon. Carbon once topped the Artificial Analysis comprehensive intelligence leaderboard, and the latest Neon ranked third in the world overall and even took first place in the coding agent evaluation, slightly surpassing Claude Fable 5.2, while the average cost of completing tasks was less than half of the latter's.

Arthur specifically posted to thank SemiAnalysis: "Thank you for the motivation. The problems mentioned in the article did exist at the time, but times have changed."

SemiAnalysis also softened this time, directly posting a "Google Apology Form" and checking several reasons for its mistakes, including "only looked at X," "didn't read the model card," and even "Mercury retrograde." In the end, it admitted that it had misjudged Gemini 4.

DeepMind founder Hassabis also personally joined in on the joke: "Apology accepted! But you missed one item: underestimating this super strong Google DeepMind team."

Source