incredibly:

to sound like gpt-4-1106-preview @ temperature: 1.0, I have to set gpt-4.1 wayyy down to 0.2

but 4 doesn’t even begin to sound like 4.1 @ temperature: 0.7(!) until you set it way up to 1.5(!)

….which is also where GPT-4 tends to do the honorable thing & start speaking gibberish instead.

The plight of the stultuloj and their stupid fucking word games played against moody LLM sampling parameters. All for for no tangible reason. Out of boredom. Just because.

The funny thing is that I think this is directionally correct, you’d want a newer model to sound more like GPT-4 at a higher temperature.

But in the olden days, imo, setting GPT-4 to temperature: 1.2 was the sweet spot. Which is still like setting GPT-4.1 somewhere low, between 0.2–0.25, in my view.

In other words, it’s cooked. GPT-4.1 is. And it’s hard to say whether you’re better off lowering the temperature (in the API) or doing a system message (more in like with ChatGPT). There’s a weird alchemy to it. I’m tired of this game.