Coding on Credit

LLMs are shockingly bad at spatial reasoning, color theory, and working out problems in the physical world. They are, however, very good at writing code—as long as you keep in mind what’s actually happening: they’re making mathematical predictions based on a massive collection of publicly available code.

It’s a bit like coding on credit. Because so many developers have published their code on GitHub and similar platforms, LLMs have an enormous pool of examples to draw from. They’re not solving new problems; they’re regurgitating solutions that have already been worked out in computer science.

And at some point, your credit gets maxed out—you end up living in tech debt, which is never a good thing.

I haven’t seen many examples of an LLM coming up with something genuinely new. The examples I have seen are really just novel connections between existing ideas. And that’s where the fundamental limitation of LLMs in software development lies.

So how do you work with them effectively?

Use them for the solved problems. Boilerplate, common patterns, glue code—this is where LLMs shine. The problem has been solved a thousand times before, and they’ve seen all thousand solutions. Save your own thinking for the novel parts.

Verify everything. LLMs are confident even when they’re wrong. Treat their output like Stack Overflow answers: useful starting points, not gospel. Read what they give you. Understand it before you ship it.

Know when to stop asking. If you’re on your third rephrasing of the same prompt and still not getting what you need, you’ve hit the credit limit. Time to think it through yourself.

Credit is a useful tool. Just don’t mistake it for wealth.