Sergey Brin | All-In Summit 2024

Sergey Brin has returned to Google almost every day because AI has pulled him back into hands-on technical work. He said the field is the most exciting development he has seen as a computer scientist and that he does not want to miss it.

AI has already changed how Brin personally programs: he increasingly asks models to write the code instead of starting from scratch. For a Sudoku experiment, an AI built the puzzle-generation and scoring system in about half an hour, faster than engineers expected, and Brin thinks Google developers still underuse coding models.

Brin expects AI to converge toward more unified models rather than remain a collection of entirely separate specialists. Google’s recent math-Olympiad system used three models—a formal theorem prover, a geometry specialist and a general language model—and Google is now trying to transfer the specialist reasoning abilities into general models.

Google is currently constrained by compute supply, but Brin doubts today’s scaling curves can simply be extrapolated indefinitely. He said Google must turn away Cloud customers seeking TPUs and GPUs, while algorithmic improvements may be advancing as fast as—or faster than—the growth in raw compute.

Google’s earlier robotics push was probably premature because the underlying AI was not capable enough. Brin said Alphabet had roughly five or six robotics businesses, including Boston Dynamics, but modern multimodal models can now understand scenes and instructions in ways those systems lacked; reliable everyday operation remains the missing step.

The hardest gap in AI is increasingly between an impressive demo and a dependable product. Brin described systems that work correctly roughly 90% of the time but require enormous engineering effort to eliminate the remaining errors, latency and availability problems before they become genuinely useful.

Brin believes Google’s biggest strategic mistake in generative AI was excessive caution, not lack of technical invention. He noted Google produced the Transformer work underlying modern language models but was “too timid” to deploy them; his preferred approach is to release powerful systems with clear warnings that they will sometimes fail.

Google was materially behind when ChatGPT launched, and Brin now watches the competitive rankings closely. He said Google has closed much of that gap and recently reached the top of some model evaluations, illustrating how seriously the company now treats competition from OpenAI, Anthropic and other AI labs.