One of the best ways to understand something is to try to break it

One of the best ways to understand something is to try to break it . . . While scrolling on Luma, I came across an event by AI Tinkerers. This was at the time AI was really blowing up. The event? A capture-the-flag challenge. Honestly, I didn’t even know what that meant at the time. The task? Crack an LLM to reveal a secret code. Each stage of the game got harder as the LLM got smarter with every attempt and round. It gets progressively harder until it was impossible to break. I went prepared, of course I tried a similar test at home (hehe) and got through…but surprise surprise during the live event I was unable to crack the LLM. Lesson learnt: - LLMs get smarter over time - yes, were going to need terminator soon :-) -- - yes, LLMs can be hacked (read: prompt injection) Here’s what I realised: - To win, you had to curate your solution exactly to the problem. - If you used an overkill solution too early, it wouldn’t work later. - As the AI learns your tricks incrementally, it will eventually block you completely. Feel free to try it out at https://hackmerlin.io/. Let me know if you get through :-) On my next post I will share about the hackathon I joined a few weeks later.  That ended a bit differently - and the solution we built can actually be useful in a wide range of real projects. Hopefully guiding you to rabbit holes of your own. #2025IsTheBestYearToBuild 🚀