Post #3236103

@freddy@social.security.plumbing We faced the same struggle when hosting a CtF earlier this year. Two categories of tasks were (according to participants) surprisingly challenging to solve even with help of with LLMs: 1. Challenges containing non-speech audio files 2. Challenges that need niche knowledge or knowledge of proprietary protocols/tools I suspect the reason for this is simply, that data of neither category is likely to be part of the model training set. I would be very interested in your own experience :)