Temperament Beats Talent: A Zen Field Guide to Doing AI Research

You become an AI researcher by pairing reading with building—neither alone is enough—and then showing up with the discipline to keep working when insight refuses to arrive. The essay borrows from Zen practice to argue that temperament matters more than raw talent: most experiments fail, most days produce nothing, and the people who succeed are the ones who put in the hours. A practical corollary is to resist over-reading papers; attempt a solution first, hit a real bottleneck, and only then turn to the literature.

On what to work on, the advice is to ignore whatever has been trendy for under six months (harnesses, agents, context engineering) and instead master decades-old fundamentals you can reason about by hand—cross-entropy, SVD, policy gradients. Chasing a higher score on an existing benchmark signals shallow ambition; a more valuable and underrated skill is finding or building a dataset that actually exercises a new method. Because the field is so young—ChatGPT is under four years old, and many of OpenAI’s technical decision-makers are under 35—accumulated experience can even hurt, breeding intuitions that fail at scale. A ‘beginner’s mind’ that refuses to let ego anchor old ideas is an asset.

The closing principles are about emotional posture. Treat experimental outcomes with equanimity: failures teach as much as successes, and suspiciously good results are usually measurement bugs, so seasoned researchers stay reflexively skeptical of anything too good to be true. Don’t measure yourself against others’ published wins; ask instead whether you’re working at a depth where you could have made that discovery. And expect the unglamorous grind—Karpathy hand-labeling ImageNet, the SWE-bench team filtering GitHub for hundreds of hours—plus the reality that many promising ideas die not because they’re wrong but because of an undiscovered bug in the code.