So many of the bugs and poor results that it can introduce are simply due to improper context. When forcibly giving it the necessary context you can clearly see it’s not a model problem but it’s a problem with the approach of gathering disparate 100 line snippets at a time.
Also, it struggles with files over 800ish lines which is extremely annoying
We need some smart deepseek-like innovation in context gathering since the hardware and cost of tokens is the real bottleneck here.