expectedwrong hindsight

GPT-4.5 Is a Presentation Layer, Not a Model

One expensive, specific thing it's actually good for.

1 min read 211 words #llms #gpt-4.5 #pipelines #writing
hindsight — still happening

GPT-4.5 as a presentation layer — not for reasoning, for prose — is still the one use case where it justifies the cost. The rhetorical choice the machine makes better than you is still annoying.

The price will come down. It always does. And when it does, GPT-4.5 will still not be worth using for most of what you're tempted to use it for — reasoning tasks where a cheaper model works fine, retrieval where it's overkill, classification where it's embarrassing overkill.

But I found the one thing.

If you have a pipeline — something that ingests data, assembles facts, maybe does a few inference steps — and that pipeline terminates in language meant for a human to read, the final draft should go through GPT-4.5. Not for the logic. For the prose.

The email draft I got back was not "good for an AI." It was good. The kind of good where you read it and feel slightly annoyed that a machine wrote it, because it made a rhetorical choice you wouldn't have made and it was the right one.

I think it might be able to tell stories. I'm not ready to say that with confidence, but something is different about the way it reaches for a sentence — it's not predicting the next token toward some platonic correct response, it's doing something that looks, from the outside, suspiciously like craft.

Use it at the end of things. The end is what gets read.