I think a lot of issues with language model coding come down to three issues. The first is the models themselves: 1. A lack of longer context, i.e. white boar

I think a lot of issues with language model coding come down to three issues. The first is the models themselves: 1. A lack of longer context, i.e. white boarding or other means of breaking down problems into components, being able to focus context in and out, etc. This is a direction models are going to go. Just like us, they are going to benefit from organizational tools. The other two are just the need for normal feedback, like we developers get: 2. They need a code, run, evaluate, improve cycle. With a hard feedback cycle today's models do much better. They quickly respond to iterative manual feedback. Which is just an inefficient form of direct feedback. 3. A lack of multiple perspectives. Groups of models competing and critiquing each other on each task should improve results for the more tricky problems. -- I personally think it is astounding that models generate somewhat reasonable, somewhat buggy code on their first pass. Just like I do! I don't know any coder that doesn't repeatedly tweak code they just wrote due to feedback. The fact that models can output first pass code so quickly without feedback today, suggests they are going to be very very good once they can test their own code, and converge on a solution informed by different attempts by different models. A group of models can also critique each others' code organization for simplicity, readability and maintainability.