They're great at helping me figure out how to make something work with a poorly-documented, buggy framework, which is indeed a large fraction of my job, whether I like it or not.

You say more details as if you gave any to begin with...

Here's a hint: What I input for inference is not in the training data. But the model can generalize well enough to handle the task.

> What I input for inference is not in the training data

??? Like everyone else? This is why LLMs have large context windows.

> But the model can generalize well enough to handle the task.

Sounds like what you're doing is normal corporate business logic. If it's able to generalize it it's probably not particularly novel.

I know the novelty of what I'm working on. If it were in the training data, I'd have reached for a library. You can think what you want, but I'm not particularly interested in defending this debate by exposing the specifics of my own work. I know what is run-of-the-mill business logic, and I know what would be novel to the training data specifically because such systems did not exist when these models were trained nor exist now in public repositories that they could be scraped.

Also, you're contradicting yourself: you said LLMs aren't great at 'anything remotely novel,' then dismissed my novel inference as routine because 'everyone' does it with context windows. Which is it? Either novel input is challenging for LLMs or it's routine and expected: contradicting your claim that LLMs can't handle novel scenarios.