Vibe researching
Language Models like ChatGPT or Claude are good at some things and less good at others. They're great at language tasks: summarizing, rephrasing, translating, generating titles, etc. They're improving but still mediocre at reasoning and coding. And they remain somewhat unreliable for factual information – often correct, but we're far from being able to use them for serious research without verification.
But they're quite valuable for what I call vibe researching. This term is, of course, inspired by vibe coding – a style of programming with LLMs where you provide instructions, let the AI write the code, and don't obsess over optimization as long as it works.
Letting the questions emerge
Vibe researching is similar in the sense that it applies to a specific type of research: when the stakes are low and you're in discovery mode. It's really good at finding out things that you didn't know you didn't know, at exploring unknown unknowns.
For example, I shared my running performance diagnostics with Claude and talked through them casually. Through these interactions, I discovered concepts I had no idea existed. That hobby runners often train in the "black hole": they're training too hard to build up the basics but not hard enough to really expand the top limit. Or about "cardiac drift," the phenomenon that the longer a run gets, the higher your heart rate will be for the same tempo.
I looked these up and they exist, they're not hallucinations. It's highly unlikely that I would have googled for these things – they're not important enough, I don't usually open a search engine to look for things randomly. They emerged in the conversation, and I wasn't actively looking for them. But they're super interesting and relevant for me.
Each message exchange has the potential to unearth fascinating perspectives or facts. I wouldn't sit down to initiate dozens of Google searches – that's tedious and lacks the emergent quality of a conversation.
More fun = more learning
In this way, vibe researching resembles vibe coding: you chat casually, let information surface naturally, and verify anything interesting. It's like prototyping with AI before doing the "real work" when something proves valuable.
For the things I don't double-check: I treat them as something I read in a Wikipedia article without checking the source. I assume it's probably true, but stay open to the chance that it isn't. This approach works best when the stakes are low and I'm dealing with well-established topics rather than fringe subjects.
Vibe researching is more engaging and enjoyable than traditional googling and reading. And because it's more fun, I find I learn more along the way.​​​​​​​​​​​​​​​​