Sergey Brin | All-In Summit 2024

Key ideas

Brin explains why he returned to technical work on AI, arguing that Google had to accept some embarrassment and release imperfect models rather than wait for perfection.

From a machine transcript of the audio. Quotes link to the moment they are said.

1. 1

   Accept embarrassment in order to ship

   Brin says that as Google grew there was more to lose, and that it was too timid to deploy language models for reasons he calls good ones: they make mistakes and say embarrassing things. He argues the capability is worth it, provided you tell people plainly that it will periodically get things very wrong. He thinks Google has improved at this.

2. 2

   A demo is not a product

   Describing a live video and audio assistant, Brin says the first reaction is amazement, followed by the question of whether it is worth using if it fails or lags 10% of the time. He says getting from there to something responsive and available takes long, repeated work, and that the result is something he calls kind of amazing.

3. 3

   Use the tools you are building

   Brin says he had an AI model write the code to generate and score Sudoku puzzles for testing itself, finishing in about half an hour while he debated it with engineers. He adds that those engineers do not use AI for their own coding as much as he thinks they ought to, and that his kids program with complicated APIs by asking the AI.

4. 4

   Robotics arrived too early for the technology

   Brin says Google has had five or six robotics businesses and that he thinks this was just a little too early. Seeing what general language models with vision can do now, he says the earlier work feels a little silly. He still places most robotics at an impressive stage rather than robust enough for daily use, though he sees no particular obstacle.

5. 5

   Don't extrapolate compute blindly

   Brin says Google is adding compute as fast as it can and has had to turn down cloud customers, so he sees good reasons for the build-out. But he disclaims market expertise and doubts straight-line forecasts of orders-of-magnitude larger training runs, noting that algorithmic gains may be outpacing added compute.

In Sergey’s words

What’s available here

• Original source

  Linked. Opens on YouTube.

• Player on this page

  Available. YouTube’s own player, loaded only when you press play.

• Full text or transcript

  Not reviewed. No license to show a text version here.

• Our summary

  Draft. Written by us with AI from a machine transcript of the video’s audio (Whisper) and checked against it. Machine transcripts can mishear; each quote links to the moment so you can hear it. Not yet rights-reviewed.

• Ask questions with AI

  Not reviewed. Off until there is licensed text to work from.