Pichai says AI is accelerating Google model work, with limits to self-improvement

What happened?

On Alphabet's October 28 earnings call, CEO Sundar Pichai said AI was helping increase Google's model-training and launch velocity. He forecast further acceleration over the next 12 months because of recursive self-improvement, and said more models were coming, including more efficient versions. He did not give model names, release dates or evidence that specific rumored models are ready.

Google announced Gemini 4 Carbon the same day, initially rolling it out to a subset of cyber defenders before expanding access. Our existing Carbon coverage details the independent benchmark, pricing and restricted rollout.

Source: Google ↗ · Source: Google ↗ · Source: LessCarbs ↗

What self-improvement means here

Recursive self-improvement broadly refers to a system contributing to the development of improved successors. In an interview published October 20, Pichai discussed current AI-assisted loops involving researchers and described agents helping them build models. The available transcript is automatically captioned; we have not independently checked the audio for the precise wording of his qualification about the end state.

The October 28 earnings remarks are a company assessment and forecast, not an independently measured breakdown of AI's contribution or proof that models autonomously design and deploy successors without people.

Source: Google ↗ · Source: Alex Kantrowitz ↗

Selected visuals are not a ranking

The source video compares a botanical magazine and page-turn effects attributed to Carbon and Opus 5.5. Its presenter acknowledges strengths on both sides. The video does not establish complete prompts, selection procedure, equivalent tool settings or authenticated backends, so the examples do not prove a general design or coding lead.

Artificial Analysis's separate coding-agent comparison, covered in our Carbon report, evaluates specific model-and-harness combinations. That distinction matters when applying any benchmark to a different workflow.

Source: LessCarbs ↗ · Source: Universe of AI ↗

Why does it matter?

Our analysis: for teams buying or evaluating AI, a promise of faster releases is not an access date. Treat AI-assisted research as distinct from fully autonomous improvement, and assess currently available models on representative tasks, error rates and cost rather than extrapolating from attractive single demonstrations.