Google Gemini team printed and put SemiAnalysis’s bearish report on their desks, and three months later the other side publicly apologized
Google DeepMind Gemini reinforcement learning scaling team member Arthur Douillard posted on November 19 that the team printed SemiAnalysis’s August 7 report, “Gemini is Cooked but GCP is Cooking,” and placed it on their office desks, checking it daily as a reminder. At the time, the report asserted that because DeepMind had lost staff from its reinforcement learning team and compute was being prioritized by Google Cloud for external customers, including competitors, it was no longer a frontier lab and the probability of returning to SOTA had fallen to zero.
Over roughly the next three months, Google consecutively released the Gemini 4 series models Argon, Carbon, Hydrogen, and Neon. According to Artificial Analysis data, Carbon once topped the overall intelligence index, while the latest Neon ranks third on the overall intelligence index with 61 points, ranks first on the coding agent index, slightly above Claude Fable 5.2, and has a significantly lower average per-task cost than the latter.
SemiAnalysis then released a “Google Apology Form,” with options such as “Only looked at X,” “Did not read the model card,” and “Mercury retrograde,” and explicitly wrote, “We were wrong about Gemini 4, and we would like to withdraw one word: Gemini is cooked.” Demis Hassabis replied that he accepted the apology and pointed out that they had failed to check “underestimating this Google DeepMind team.”
Original source:
Title: Google Researcher Reveals The Gemini Team Had Put SemiAnalysis' Comments Saying Deepmind Was No Longer A Frontier Lab On Their Desks For Motivation
Publisher: OfficeChai
Publication date: 2026-11-19
URL: officechai.com/ai/google-researcher-reveals-the-gemini-team-had-put-semianalysis-comments-saying-deepmind-was-no-longer-a-frontier-lab-on-their-desks-for-motivation
Source type: technology media report
Verification: full-text reading, citing Douillard’s original post and key passages from SemiAnalysis’s original text
Supporting content: the desk-printing detail, the report title and date, DeepMind’s quoted line that it was “no longer a frontier lab,” the compute allocation dispute, and the team’s response
ABAB AI Insight
The core basis of SemiAnalysis’s August report was the attrition of DeepMind’s reinforcement learning team and internal compute being prioritized by GCP for external TPU sales, including supply to direct competitors such as Anthropic. Douillard acknowledged that these issues did indeed exist at the time, indicating that the organizational tension over talent and resources was not fabricated, but the next three months of model releases show that the reorganized team was still able to keep iterating with limited internal compute.
The rapid release of the Gemini 4 series, from Argon to Neon, changed market expectations for Google’s model cadence. Carbon topping Artificial Analysis, and Neon leading on coding agents with lower cost, directly challenged the judgment that Google could only sell compute and that its models could not outperform OpenAI or Anthropic. The cost advantage is especially sensitive for enterprise procurement, because differences in average spending on coding agent tasks can directly alter purchasing decisions.
SemiAnalysis’s apology form chose a self-deprecating style to retract the extreme statement “Gemini is cooked,” while still preserving its judgment on the success of GCP commercialization. This way of revising the view separates the misjudgment of model capability from the correct prediction on cloud business, avoiding a wholesale否ulation of its framework and also reflecting how analysis firms quickly adjust after making public forecasts.
Hassabis’s playful reply shifted the focus from “the report was wrong” to “the team was underestimated.” After DeepMind experienced leadership changes, with Hassabis becoming chairman, and staff turnover, this statement strengthened the internal-mobilization narrative: external pessimism can be turned into internal momentum, and actual model progress is what matters for regaining a frontier position.
ABAB News · Cognitive Law
A report that publicly declares something finished can be printed as a daily reminder, but the speed of organizational mobilization still depends on internal resource reorganization
When compute is prioritized for external customers, the pace of model iteration may slow, but it does not necessarily go to zero
When an analysis firm retracts an extreme statement, it often retains its original judgment on the commercialization path; model rankings and cloud revenue belong to different ledgers