Joaquín Quiñonero Candela spent nearly 20 years building AI at Microsoft, LinkedIn, and Meta.
He founded Facebook’s Applied Machine Learning team and built the AI platform behind its products.
Then he joined OpenAI, where he led Preparedness, the team that tests frontier models for catastrophic risks, and later ran recruiting and people.
I asked him what founders should do differently now that AI is everywhere.
1. Spend on tokens, not headcount
The old advice was not to raise too much before product-market fit, because the extra cash ends up in hires you don’t need.
Joaquín still believes in small teams and focus, but he thinks the share of a startup’s budget that goes to tokens instead of salaries has to grow a lot.
Because burning compute is what lets a small team test its ideas faster than anyone else.
2. Don’t do research just to impress investors
More and more startups are building research teams early, and he thinks most of them are “investor bait.”
Research that moves a field, like what came out of Bell Labs, needs years of funding with no pressure to show results, and no startup can afford that.
What a startup can do is applied research that makes its own product better, like cutting the cost of inference by 10x.
3. Stay close to the work, whatever your title
As a company grows, the founder usually ends up further from customers and from the technology.
At Facebook, he left his job leading applied machine learning to go back to writing code.
At OpenAI, when leadership offered him a much bigger role, he asked to be an intern.
He says careers aren’t ladders, they’re helixes: you keep circling back to hands-on work, and every time you return, the tools have changed.
In the episode, we also talk about why the labs warning loudest about AI keep racing to build it, and why he’d rather speed up safety research rather than slow down the frontier.
[YouTube] → [Spotify] → [Apple Podcast]
Also in this episode:
[12:00] Why careers are helixes, not ladders
[37:25] Why the labs warning loudest about AI keep racing to build it
[39:20] Why no lab can tell its investors it’s slowing down
