Rodrigo Schmidt grew up in a fishing town in the south of Brazil. No money, no network.
Fifteen years later he was leading the engineering behind Instagram’s feed, the personalization machine billions of people open every day.
His take on AI agents: they’re making the mistake Instagram almost made. They optimize for what you want right now instead of what you actually want.
At Instagram the easy path was showing people whatever kept them scrolling. His name for it: candy for lunch. Amazing the first time, miserable by week two. The hard part was never the ranking model, it was choosing an objective function that paid off long-term, even at the cost of engagement today.
Agents take the candy. They validate whatever you say and optimize for the next reply. Feels great in the moment. It’s not the relationship you wants with agents.
Then add memory. Everyone’s racing to bigger context windows, assuming more context means better answers. Past a point it’s the opposite: the extra data makes the agent worse.
Rodrigo says forgetting isn’t a flaw in memory, it’s what makes memory usable. The win is knowing what to throw away, and that’s the app’s job, not the model’s.
Rodrigo left Meta to do something about it, co-founding Vetto to build the research-grade data frontier models train on, betting the next leap comes from better signal, not more of it.
[YouTube] → [Spotify] → [Apple Podcast]
Also in this episode:
[21:30] Why Instagram’s early feed almost fell into the same trap
[33:20] Why a bigger context window can quietly make your agent worse
[35:30] Why forgetting isn’t a bug, it’s what makes memory work at all
[43:20] His lessons from Facebook’s AI Research on building product & research orgs
