No one ever really taught you how to do research. You get a desk, a problem someone else picked out, and a vague instruction to "make something new."
So most people reverse-engineer the job from what they can see—papers, posts, and announcements—and end up learning how to look like a researcher rather than how to be one. Real research ability is a stack of small skills, and almost every one of them can be cultivated through deliberate practice.
Choose Your Own Problems
Richard Hamming had a habit at Bell Labs that made him unwelcome at lunch. He would ask the person next to him what the important problems in their field were, and then ask them why they weren't working on those. People would switch tables.
The question stings because most of us don't have a good answer. We aren't choosing problems; we're absorbing them—from advisors, from last quarter's announcements by a big lab, from papers everyone is citing and sharing this week.
The trouble with absorbed problems is that you hold the conclusion but not the reasoning behind it. You know some famous lab cares about a direction, but you don't know why, what they expect to find, or what would make them abandon it.
You'll notice their pivot a year later. And on a problem that's already trending, you're racing against 1,000 people who started earlier and have more compute than you.
John Schulman's guide to ML research splits the work into two modes. In the first, you read the literature and look for things to improve. In the second, you choose an outcome you genuinely want to achieve and work backwards to design experiments.
He argues for the latter, the subtle reason being that it manufactures originality. A goal you actually care about will drag you into territory no review paper has ever covered.
As for "taste," people often discuss it as a talent. But it behaves more like a muscle.
Before running each experiment, predict its outcome; cover up a paper's results section and guess the data from its methods; note which results announced this month will still matter in two years, and later check your hit rate. One prediction plus one correction, repeated hundreds of times—every good model is trained that way, including the one in your head.
Upgrade Your Inputs
Shared reading lists produce shared ideas. If your information diet is just the arXiv trending list plus whatever filters through group chats, you'll inevitably reach the same conclusions as everyone else at the same time, making those conclusions nearly worthless.
Old material is severely undervalued. The field keeps replaying its own past with a delay: Mixture of Experts (MoE) traces back to 1991, LSTMs to 1997, backpropagation went mainstream in 1986.
Richard Sutton wrote The Bitter Lesson in 2019 in just over a thousand words, and it predicted the field's trajectory more accurately than reviews ten times its length. Claude Shannon gave a talk on creative thinking in 1952; his first move was to shrink the problem until it was almost trivial, solve the small version, then add the difficulty back bit by bit.
That single move will help you break through more walls than any modern productivity advice.
Breadth is as important as depth. Interpretability research unapologetically borrows from neuroscience; evaluation design is mechanism design in a lab coat; a practical awareness of how GPUs actually move memory lets you judge which architecture papers will fail before benchmarks are even run; and honest statistics is arguably the rarest skill in machine learning, where much published "rigor" is just "vibes with error bars."
One more thing. Read the papers themselves, not the posts that summarize them. The appendix is where secrets are buried, and the "Limitations" section is often the most honest part of the entire document.
Write Everything Down
Paul Graham observed that an idea always feels fully formed until you try to write it down. But words on a page expose the varnished-over holes in your brain: the untested assumptions, the steps that don't actually connect, the two claims that quietly contradict each other.
Feynman's rule was that the first person you must avoid fooling is yourself, because you're the easiest person to fool. Writing is the cheapest defense mechanism ever invented.
Darwin took it further and systematized it: any fact contrary to his theory was written down immediately, because he found his memory deleted inconvenient evidence far faster than favorable evidence. Your memory does the same with your failed runs.
Keep a log: hypotheses, setup, expectations, results, updated understanding. Rereading last month's entries will humble you like no reviewer ever could.
