Follow questions whose endings you can't see. Zhengdong enjoyed software internships, but after two or three months he could already see where it was going. Research felt like a better long-term game because of its expansive aperture: “[I] really, really want to find the answer.”
Don't leave the weird thing unexplained. One of his favorite researcher archetypes notices an anomaly and keeps asking why: “you just keep going and asking, why is this weird?” Chase enough of them and you see patterns you couldn't have planned to find. It's the spirit of Laura Deming's “The Rage of Research,” though he translates the rage gently: “just really wanting to know something.”
Knowing what to skip is learned, not stated. Zhengdong overcorrects every project: move fast and you skip things that mattered; examine everything and the field moves without you. There's no rule for the balance between covering ground and going deep — “finding this balance is just something that you need to get a lot of reps in to improve at.”
Zoomed out, research is just choosing the eval. AI models are becoming “this super laser that once you focus it on something, it will be solved.” He thinks research resolves to creating a new evaluation, with the rest just optimizing numbers. Even recursive self-improvement is “just some kind of meta-search” that procrastinates the real question: what makes a good eval?
Know the conversation, think for yourself. Entry into a field runs through its discourse: “you just need to be part of the conversation, whatever that conversation is,” even when that means Twitter instead of journals. But knowing it is in service of independence: close enough to not miss the obvious, removed enough to have ideas that aren't in everyone's stream.
Automating the tasks doesn't automate the job. Humans have struggled to assess job performance since the pyramids (“How many rocks did you place?”). Every time a tool swallows the specifiable part, the job reforms around what's left: judgment, purpose, even how much joy you bring your colleagues. Automate the tasks and the work migrates to what the spreadsheet can’t capture.
We'll build AGI before we can define it. From his 2023 letter: “I more and more see the world where we build a machine that people agree is AGI before we write some words that people agree defines AGI.” Defining it may even be “possibly hopeless”: humans have written books about distinctions for thousands of years. The machines will keep marching past goalposts and we’ll keep moving them.
Genius may just be very large search. AlphaGo's Move 37 looked like genius intuition. Zhengdong's suspicion: “Without having exhaustively done that search, it looks like genius to you. But if you had looked at all the millions of moves, then it's sort of obvious.” Brute-forced creativity: first Go, then knowledge work, eventually invention.