Comment by tonii141

I’m not demanding anything, I’m pointing out that performance tends to degrade as context scales, which follows from current LLM architectures as autoregressive models.

In that sense, Yann was right.