Post #1824112

Claude 4.7 is an *idiot savant*: it is very 'clever', as in it can take instructions involving task-specific knowledge (e.g. 'write an example of this construction') and succesfully complete it, but it is very 'dumb' as it's understanding is definitely still a pose. It misses the big picture and constantly overlooks stuff. But what I'm still amazed by is how fast its capabilities degrade with time. These models have O(1M) context windows and their performance perceivably degrades after O(100K), to the point they cannot write well-formatted replies anymore. Yann LeCunn has 'predicted' this failure mode basically immediately after LLM became big. Any improvement on long-horizon tasks is likely to be completely due to fine-tuning to that specific domain, rather than genuine improvement of the underlying AI.