Google co-founder Sergey Brin rarely speaks publicly. He just sat down for an unscripted Q&A on Frontier AI and admitted something most lab leaders won’t:

Even the people building these models do not fully understand what they have created.

In this 29-minute conversation at AGI House, Brin walks through the surprises that actually matter right now:

↳ Specialized models are converging into one general system faster than anyone predicted.

• Train on coding and math reasoning mysteriously improves.

• Feed it images and geometric word problems get better.

• The capabilities bleed into each other in ways nobody fully engineered.

↳ One of the biggest leaps came from the dumbest-sounding trick imaginable:

• Just telling the model to “think step by step.”

• Brin says there was no obvious reason it should work. It did.

↳ He pushes back on hype around superintelligence (it still can’t solve the impossible), notes that AI mastering a domain has never stopped humans from getting better at it (chess after Deep Blue, Go after AlphaGo), and says something close to transformers is probably enough to reach AGI.

↳ Inside Google, they are already using the AI to build the AI.

• That self-improvement loop is where Brin spends most of his time.

• World models and physical interaction are the missing piece for the version of AGI that can do anything a person can.

Candid, technical, and free of the usual marketing. One of the clearest looks at how the people actually shipping frontier models are thinking in real time.

Instead of another Netflix series today, watch this talk.