[SOTGO]

I thought the section on finding bugs was interesting. I’d be curious how many false positives the LLM identified to get the true positive rate that high. My experience with LLMs is that they will find “bugs” if you ask them too, even if there isn’t one.

[dakshgupta]

This specific case each file had a single bug in it, and the bot was instructed to find exactly one bug. The wrong cases were all false positives, in that it made up a bug

[dheera]

I think this is mostly the fault of RLHF over-indexing on pleasing the user rather than being right.

You can system prompt them to mitigate this to some degree. Explicitly tell it that it is the coding expert and to push back if it thinks the user is wrong or the task is flawed, it is better to be unsure than to bullshit, etc.

[dakshgupta]

This is surprisingly hard to mitigate with system prompts because not being opinionated is ingrained so deeply in (presumably) post-training

[SilverBirch]

Absolutely agree with this, I use ChatGPT to ask about how best to do something, if I say "I'm not sure about that" in response to some proposal the tool will basically always back down and change something even if it was totally right the first time. It's a real problem because it makes it very difficult to interrogate the tool when you're unsure if it's answer is correct.

[never_better]

In my experience that's what's separated the different AI code review tools on the market right now.

Some are tuned far better than others on signal/noise ratio.