FULL INTERVIEW: Yong Zheng-Xin says a paper argued LLMs can't make Einstein-style leaps, so they'll never reach general relativity. His answer: Feynman showed you can derive it without Einstein, from theories that already existed. @yong_zhengxin works at @OpenAI on RSI preparedness and safety. 2 weeks after publishing "Hot Take: LLMs Can Jump," he joined @theojaffee and @schisofrenia to discuss whether models can create genuinely new knowledge: 01:18 the 3 kinds of reasoning, and which one the paper says LLMs lack 03:07 why cyber is accelerating and algorithmic progress isn't 05:48 the elevator thought experiment, and what Feynman said about it 07:59 why a model that isn't a specialist finds connections specialists miss 08:45 Yudkowsky's Bayesian superintelligence and three frames of a falling apple 10:59 why interconnected knowledge makes continual learning easier 14:08 the wet lab is the bottleneck, not the reasoning 15:31 whether you need embodiment to know when something feels wrong 17:24 the real constraint on abduction is diversity of hypotheses 19:22 we never ask one scientist to make every breakthrough 22:09 whether GPT-4 could do today's cyber work with infinite compute 23:34 what safety even means once a model is improving itself