[nomendos]

To clarify, in summary so far LLM's can do a bit more than the inputs used for training. Example https://dynomight.net/chess/ as well as some coding solutions are a bit better than each input alone, although if the solution requires more than "a bit more" then LLMs start to hallucinate (spin the wheels). Time will tell if LLM's can jump this "a bit more" barrier? (I can not tell for sure yet, but the current knowledge and my NL tells me if I'd have to put a bet, it would be that the new approach/design is needed)