How to Be Good at Research

A widely shared thread that breaks down the unspoken craft of research — how researchers actually learn the job, how to think about novelty, and how to produce meaningful work when the path is unclear.

• An absorbed problem hands you the conclusion without the reasoning. You know a famous lab cares about a direction but not why, not what they expect to find, not what would make them quit — so when they pivot, you learn about it a year late.
• Schulman splits research into two modes: read the literature and hunt for improvements, or pick an outcome you want to exist and reason backwards to experiments. The second manufactures originality, because a goal you care about drags you where no survey goes.
• Old material is underpriced because the field reruns its own past on a delay: mixture of experts is from 1991, LSTMs 1997, backprop went mainstream in 1986. Sutton's thousand-word bitter lesson predicts the field better than surveys ten times its length.
• Shannon's 1952 method for a stuck problem: shrink it until it is nearly trivial, solve the small version, then add the difficulty back one piece at a time.
• Read the paper, not the thread about it. The appendix is where the bodies are buried and the limitations section is usually the most honest paragraph in the document.

• mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.

• videoThe Loop Is the Product — Roland Gavrilescu, IntrospectionAI Engineer
• videoIf we want them to do Knowledge Work, design them as Knowledge Agents — Benjamin Clavié, MixedbreadAI Engineer
• postBuilding Without Foundations Means Reinventing RAG, BadlyAndrew Ng

Checking sign-in…

Loading comments…