"how to be good at research" by @itsreallyvivek #70

Activity

1. lintool

   added

   advice-by-others

   on Sep 6, 2026
2. lintool

   changed the title [-]how to be good at research[/-] [+]"how to be good at research" by @itsreallyvivek[/+] on Sep 6, 2026
3. lintool commented on Sep 6, 2026

   

   lintool

   on Sep 6, 2026

   OwnerAuthor

   More actions

   Adding a TL;DR and some reactions to Vivek’s how to be good at research, published June 10, 2026:

   TL;DR: Research ability consists of smaller skills that can be deliberately trained. The post emphasizes everyday habits that improve judgment and accumulate into meaningful work over time:

   • Choose problems deliberately: Start from outcomes you care about, understand why they matter, and work backward to experiments.
   • Train your judgment: Predict results before running experiments, then compare expectations with outcomes.
   • Broaden your reading: Read original papers—including appendices and limitations—alongside older work and neighboring disciplines.
   • Keep an honest research log: Record hypotheses, expectations, results, and changes in belief, especially evidence against your ideas.
   • Shorten the experiment cycle: Build reproducible tooling, test small versions first, tune baselines, and use ablations to identify what actually works.
   • Inspect data and failures directly: Read examples and transcripts; aggregate scores alone can conceal the most useful information.
   • Explore and collaborate: Try different areas, share tools and explanations, and cultivate colleagues who will challenge your ideas early.

   The recurring message is that reading, writing, experimentation, and generous collaboration reinforce each other through sustained practice. Original post

   Reactions:

   • Applicability beyond research: Ananya Gupta’s X reply argues that the advice extends beyond AI research. Evan Payne’s longer response connects it to everyday agent engineering, identifying concrete changes: predicting evaluation outcomes, reading failed transcripts, and returning to primary sources. Gupta’s reply · Payne’s response
   • Resource constraints: Lost Temple’s commentary praises the synthesis of familiar ideas into practical habits, but identifies a missing discussion of structural barriers. In particular, exploring multiple subfields may be difficult for researchers with limited time, compute, or colleagues. Commentary, June 13
   • Questions about AI authorship: Benno Sturgeon’s X reply and Diana Kimball Berlin’s newsletter both praise the advice while suggesting AI involvement in its writing. Berlin particularly highlights training judgment through prediction and correction. These are their assessments, not verified evidence of how the article was written. Sturgeon’s reply · Berlin’s newsletter, June 15
   • Translation and recommendation: MinLi published a Chinese translation with an enthusiastic endorsement, saying the article would have been valuable during undergraduate and graduate study. This is evidence of recommendation and circulation, rather than a detailed critique. Translation and introduction, June 14

   X required login to see all replies, so this roundup reflects accessible examples rather than a comprehensive survey.

   The summary and reaction roundup above were generated by GPT-6 Astra in Codex.