@regehr I also noticed they tend to suggest overly complex methods sometimes. Had a DAG reachability problem that had to be computed and kept up to date as the DAG was being modified, and both Opus and GPT suggested cool sounding methods (with paper citations that were real), both of which were pretty complex. I was suspicious and one evening after 45 minutes and a beer I came up with a much cleaner method that was just as efficient but a ton simpler. Had I taken their advice I would have gone down a deep unnecessary hole.

@regehr I’ve had a similar experience. LLMs have been helpful with reviewing and doing simple work for a semi-sophisticated rewriting system I’m building, but when experimented with letting them work on the core algorithms they tend to produce code I wouldn’t have that isn’t as clean or robust as I’d like. I’ve found them useful to help sketch out a method, but I usually do the actual coding work and let it find my typos and subtle bugs when I’m done. Took a weekend to write a confluence checker that would have taken me maybe 3x that before. I wrote the code, Opus reviewed it, and I let it write some mundane stuff I didn’t feel like doing like pretty printers to help me debug. I often wonder what the code looks like for people who primarily vibe code and just shovel tests into the model.

@regehr openclaw fanboys give strong NFT-bro vibes. In some sense it’s useful tech: immediately tells me if someone is a moron.