@VincentLogic: If Ilya Is Right, the Three Strongest Consensuses in AI Over the Past Few Years Might All Be Wrong: Scaling Is No Longer the Universal Answer. High Benchmark Scores Don't Equal True Intelligence. RL Might Even Be Making Models 'Dumber'. This Interview, Called 'the Last Interview Before Ilya Disappeared'...

If Ilya's judgment is correct, then the three most solid consensuses in the AI industry over the past few years might all be wrong: Scaling is no longer the universal answer. High benchmark scores do not equal true intelligence. RL might even be making models 'dumber'. In this conversation, dubbed 'the last interview before Ilya disappeared', he compared current large models to competitive programming contestants: they can solve very difficult problems, but when faced with real projects, they fix one bug and create two more, in a repeating cycle. The problem may not be that the models aren't large enough, but that the reward mechanism drives them to pursue the correct answer excessively, gradually losing common sense, intuition, and learning ability. His more radical judgment is: The dividend from pre-training has peaked, and the RL dividend is also nearly exhausted. AI has already regressed from the era of Scaling back to the 'era of research'. True superintelligence will not be a finished product that is downloaded and knows everything. Instead, it will be more like a fifteen-year-old genius: possessing extremely strong learning ability and then continuously growing in the real world. If this direction holds, the key to the next round of AI competition will no longer be who piles up more data and GPUs, but who first cracks how humans can learn a new thing just by seeing a few examples. Ilya rarely talks nonsense, but this time he chatted for over forty minutes. It's worth watching the whole thing. Which part most challenges your understanding?