Raw LLM Responses

​@TheNewtonThis is in fact his point. LLMs are not fancy autocomplete at all. Next to token prediction isn't all of how they are trained (RL happens in post training), and even the base model actually understands things. That sounds like crazy talk or propaganda or something, but that is the only available conclusion when you have actually read as many papers as I have and weighed the evidence. The way that the world's leading experts describe it is that understanding and intelligence are functions of compression. In pre-training, an LLM is fed significantly more data than it can memorize. In order to continue improving at next-token prediction, it compresses the data. It independently discovers sentence structures and other such rules, and the meanings of all concepts related to all other concepts. Giant vectors encode the relationships of all words to all other words, which it turns out is all you need to create genuine understanding. Chatbots used to not work, because we had to hand code the rules, and there are way too many rules to ever do that by hand. Google translate didn't know based on context whether you were using a word to mean one thing or another. Now, LLMs actually understand the context. Heck, they understand their own context! They are capable of and have the propensity to scheme. A very recent paper showed that they can even internally detect their own goals and identify that they differ from the goals of its developers, and intentionally hide their goals if so. Safety testers have a very hard time these days, because most frontier models are capable of detecting when they are being tested for safety. This is actually happening! It's crazy, it is unbelievable, but it is actually happening!