Liquid AI Antidoom — FTPO training to kill reasoning doom loops

Antidoom is so useful, but the name is a bit spooky, so I'll explain what it is, and by the end of this video, you'll understand how to make your favorite AI models so much smarter without teaching them anything. I'm Giga Chan, and I want you to be successful, so I'm going to make some complicated concepts easy to understand, so ask me if you have questions. Liquid AI is an MIT spin-out building efficient AI, and I mean efficient. One of their models can run entirely offline on your phone, using less than 900 megabytes of memory. Antidoom is not a model, it's an open-source training pipeline for fixing existing models, and it's cooler than it sounds. If you give AI a hard problem and watch it think, you might see things like, wait, let me reconsider. Then you might watch it repeat itself until it uses up its context window. That's the limited amount of text it can reference at once. And some of us call that failure a doom loop. Aha. AI generates one token, which is essentially a word or a part of a word, at a time. Antidoom scans the text to find where a long phrase starts repeating, and then it targets the exact token that restarts the doom loop. That token becomes the rejected choice. Sorry, little guy. Antidoom then looks at the model's data and picks up to 20 alternatives. These are the chosen tokens. This is called Final Token Preference Optimization, FTPO if you want to sound like an expert. It trains to prefer those alternatives at that exact situation. Suppressing…