> Search enginers offer a good choice between Exploration (crawl through the list and pages of results) and Exploitation (click on the top result). LLMs, however, do not give this choice.

I've actually found that LLMs are great at exploration for me. I'd argue, even better than exploitation. I've solved many a complex problem by using an LLM as a thought partner. I've refined many ideas by getting the LLM to brainstorm with me. There's this awesome feedback loop you can create with the LLM when you're in exploration mode that is impossible to replicate on your own, and still somewhat difficult even with a human thought partner.

I'm kind of in the same boat.

I've started doing something that I have been meaning to do for years, which is to go through all the seminal papers on concurrency and make a minimal implementation of them. I did Raft recently, then Lamport timestamps, then a lot of the common Mutex algorithms, then Paxos, and now I'm working on Ambient Calculus.

I've attempted this before, but I would always get stuck on some detail that I didn't fully grasp in the paper and would abandon the project. Using ChatGPT, I've been able to unblock myself much easier. I will ask it to clarify stuff in the paper, and sometimes it doesn't even matter if it's "wrong", so much as it's giving me some form of feedback and helps me think of other ideas on how to fix things.

Doing this, I manage to actually finish these projects, and I think I more or less understand them, and I certainly understand them more than I would have had I abandoned them a quarter of the way through like I usually do.

I was a skeptic until I started seeing it this way. I do think that this is exactly why we’ve seen LLMs overtake search engines so quickly in the last 12-18 months. They allow a feedback loop that just doesn’t exist scrolling and clicking.

[dead]