It outputs compilable/runable code. And if it doesn't it can process error messages and retry until it outputs compilable/runable code.

Sucks that this is my life for now or to have to look for something else eventually but even worse is to consider how this rollout might effect other industries where just trying shit until it sort of works is actually extremely harmful.

In my experience, for the very generic software engineering tasks I have given to LLMs, they are decent at writing code and refining it until it works. Sometimes the code is even pretty good. They have gotten much, much better as of late (last 6 months), and now even DeepSeek's cheapest model is solid for the narrow purpose of writing code.

Now, this is not to say anything about the long term effects of this, where engineers have no idea how the codebase they are working on is structured, and all interactions with it have to go through the LLM, nor the broad negative societal consequences of using LLMs at scale, which would be far too long to enumerate in this comment.

For writing the glorified CRUD app for my B2B SaaS employer, the code it generates works well enough. Keep in mind that management never really cared about software quality even before the LLMs. It's hard to keep caring about the quality of the software as a code monkey because who the hell cares about making a good business software for corpos? Even if solving problems in the abstract can be fun for awhile, eventually it gets old.

It's the software equivalent of adulterating food with sawdust to maximize profits.