Temperament matters more than talent in AI research
via Zen and the Art of Machine Learning Research — Jack Morris
The core thesis: talent is overrated in AI research; temperament is what actually separates great researchers. A few principles worth keeping:
Read and build in combination. You can’t do only one. The loop between them is how you become a researcher.
Chase fundamentals, not trends. Concepts like agents, harnesses, and context engineering will change. Cross-entropy, SVD, and policy gradients won’t. Go deep on the basics.
Beginner’s mind is a real advantage. Experience in AI can actually be counterproductive — many pre-scaling-era researchers still optimize for small-scale performance. No one has been at this long enough to have a huge head start.
Inspiration happens away from the keyboard. Most “aha moments” come on walks, not at the screen. Darwin, Tesla, Feynman, Aristotle — all famous walkers.
Experimental equanimity. Treat bad results as equally informative as good ones. Be more skeptical of suspiciously good results — they’re usually a bug.
Don’t benchmark-chase. If the best possible outcome of your project is a higher score on an existing benchmark, you’re not going deep enough.
Healthy paranoia. If something looks wrong in your metrics, you cannot move on. Understand every anomaly, because something may be wrong.
Fast feedback loops beat parallel slow ones. Context switching is harmful. Design experiments to return results quickly rather than running many slow ones at once.
Gruntwork is the job. The more ambitious and forward-thinking an idea, the more obscure labor behind it. This difficulty is a feature, not a bug.
Don’t outsource understanding. Coding agents can write, run, and interpret experiments for you — but small omissions (shortened prompts, wrong configs) can materially change results. You need to understand the system that produced your findings.