Sergey Brin said he retired from day-to-day work at Google about a month before the COVID-19 pandemic began.
“And I retired like a month before COVID hit, in theory.”
Person
This page collects what Sergey Brin, Google co-founder, said on the All-In podcast, in a live episode recorded in Miami. He covered Google's work on AI and its Gemini models, his return to the company, its robotics efforts, and the early days of the web.
Brin said he retired from day-to-day work at Google about a month before the COVID-19 pandemic began. He said he later returned to work on AI, and that he became closely involved in pre-training. He also said Google trains Gemini mostly on its own TPU chips while buying large numbers of NVIDIA chips. Each fact below includes his exact quote and a link to the moment.
1 episodes · 22 quoted claims
Sergey Brin said he retired from day-to-day work at Google about a month before the COVID-19 pandemic began.
“And I retired like a month before COVID hit, in theory.”
Brin said an OpenAI employee named Dan, whom he met at a party, helped persuade him to return to working at Google on AI.
“there was a guy from OpenAI this guy named Dan and I ran into him at a little party”
Brin said he has not finished his computer science PhD and is technically on leave of absence.
“I haven't finished my PhD yet, but working on it.”
Brin recalled that the early web browser Mosaic had a What's New page that listed the few new websites each week.
“And you remember there was a What's New page?”
Brin said the code he submitted at Google was minor, such as adding himself to get access to systems and making small changes called CLs.
“I think I needed to, like, add myself to get access to some things and, you know, a minor CL here or there.”
Brin said he became closely involved in pre-training, the large-scale first stage of training AI models, before moving on to post-training for thinking models.
“I was really very close with what we call pre-training.”
Brin said an AI that reads 1,000 search results and follows up on each does work that would take him a week.
“But if it sucks down the top 1 ,000 results and then does follow-on searches for each of those and reads them deeply, that's a week of work for me.”
Brin said in a joking exchange that AI models, not just Google's, tend to perform better when threatened.
“But not just our models, but all models tend to do better if you threaten them.”
Brin said Google acquired and later sold about five robotics companies, including Boston Dynamics.
“I think we've acquired and later sold five or so robotics companies and Boston being one of them.”
Brin said Google built an internal robotics effort called Everyday Robots and later had to transition it away.
“We built out everyday robotics internally and then later had to transition that.”
Brin said Google acquired at least two humanoid robot startups and later sold them, which colours his scepticism about humanoid robots.
“we've at least acquired at least two humanoid robotic startups and later sold them”
Brin said Google had an internal page that listed Gemini among tools engineers were not allowed to use for coding, until he got it removed.
“we had this list of what you're allowed to use to code and what you're not allowed to use to code, and Gemini was on the no list.”
Brin said machine learning moved from different model types for vision and text to the transformer architecture.
“And all of this has shifted to transformers, basically.”
Brin credited DeepSeek with releasing a surprisingly powerful open model in January, which narrowed the gap with proprietary models.
“DeepSeek released a really surprisingly powerful model when it was January or so.”
Brin said Google released Gemma, a family of small open-weight models that fit on a single computer.
“So we released Gemma, which are our open source or open to wait models.”
Brin said Google released its smart glasses too early because the technology was not ready.
“Yeah, I kind of messed that up, I'll be honest. Got the timing totally wrong on that.”
Brin noted that many people say the technological singularity is about five years away, which makes predicting 10 years ahead difficult.
“A lot of people are saying, hey, the singularity is like five years away.”
Brin said an internal AI chat tool at Google could read a whole chat space and answer complicated questions about it.
“But it could suck down a whole chat space and then answer pretty complicated questions.”
Brin said an AI summary of a work chat led him to suggest promoting a quiet young engineer, and her manager agreed she deserved it.
“I actually picked out this woman, this young woman engineer who, like, you know, I didn't even notice she wasn't very vocal, particularly in that group.”
Brin said Google trains Gemini mostly on its own TPU chips while also buying large numbers of NVIDIA chips.
“we mostly use our own TPUs. But we also do support NVIDIA and we're one of the big purchasers of NVIDIA chips”
Brin said users get a few free prompts with Gemini 2.5 Pro but must pay if they use it heavily.
“Yeah, you got a few prompts for free, but if you do it a bunch, you need to pay.”
Brin said each new Gemini generation tends to make the free tier about as good as the previous paid tier.
“every time we've gone forward a generation, then the sort of the new free tier is usually as good as the previous pro tier and sometimes better.”