sadrasabouri

It argues that training for "math, code, and reasoning" caused this. Do LLMs only trained for natural language (probably older models) have the same characteristics? Does training for these additional domains help in some ways, and this specific style of writing is just a side-effect of that?

When the project scales, the transformation of intent into such pseudocode becomes a big deal. Having the pseudocode as an intermediate level to check agents' artifacts is interesting media, but I couldn't see how this would actually work with Huzzah tbh.

Today I got access to the new GitHubNext product, Copilot Workspace. I decided to test it on an old tiny Python package I developed three years ago for fun. The library is meant to simulate random walks. At that time I opened an issue for handling dead-end cases by randomly jumping to new states. Today I used the same issue for the starter. You can see the results of Copilot Workspace.

Caveats: 1. It looks very neat 2. It took nearly 5 minutes to come up with the solution after specification, planning, and then opening a pull request 3. the proposed solution is not actually true. Although it catches a valid case by checking if there is a row in the transaction matrix summed up to 1, this case is already prevented in `__init__`. A more plausible solution would be to check if the current state has a diagonal 1 in the transition matrix, which results in the random walker going back to the same state after that.

Reading the documentation, I was expecting more tbh. As I found out there is no way to customize the API. The response is always a JSON dump of that google Sheet.