- Idea and implementation sit on two orthogonal axes. Never let anyone stop you from trying an idea just because they tried it before. Ideas now take a week with Claude.
- Talking creates mode collapse in ideas. New ideas are sacred. Sharing creates homogeneity of ideas, especially when you just start at a new place. You can only be a blank slate once.
- Iterate at the speed of light. Waiting a day to get back your experiment’s result will cause you to forget all the context of what you were trying.
- Create robust infra. Infra is a core component to iteration speed and makes research more fun for everyone.
- Few meetings. Great researchers know what the bottleneck is and have the agency to chase them down.
- Do things with substance. Do things well one time and leave it be after. Don’t commit to something you will leave half-baked.
- Write thoughtfully. There are too many low-conviction ideas when using Claude. Writing forces you to stand behind your past work.
- Make the org as flat as possible. It’s valuable for researchers to think full-stack. And hierarchy and teams create infighting for limited resources (GPUs), and thus politics.
- Good ideas take time. The top 1% of ideas contribute 95% of the value (aside: you should have no shame reading papers during work).
- Align incentives. In big AI, bonuses and promotions often optimize for steady and solid incremental work, while moonshots that move the needle are riskier.
- Make both the short-term vision and north star clear. Without one, teams hedge with little bunts that don’t compound. Big things get done by people focused and pushing in the same direction for a long time.
But most of all, a great research org requires great people. Work for teams with the highest of bars. Jump through hoops for them.
All opinions were formed with much appreciation from the great people and teams I have worked with.