GPT 4.5bro

I remember when OpenAI launched GPT-4o. I remember commenting, "It is so good that it should be called GPT-4whoa."

They dropped GPT-4.5 two days ago, and everyone was unpleasantly surprised. The model performs below other frontier models on all the benchmarks, it has a knowledge cutoff of mid-2023, and it is absurdly expensive. But they say it has a better vibe than other models. They say that the subjective experience of chatting with it is more human-like. Essentially, they are saying it is GPT-4.5bro.

Karpathy tried to test the bro-ness hypothesis by doing a blind-test on twitter, asking people to compare outputs from 4 and 4.5bro. People chose GPT-4, four times out of five. He was surprised, as he personally felt the bro-ness of 4.5, and said "...Either the high-taste testers are noticing the new and unique structure but the low-taste ones are overwhelming the poll. Or we're just hallucinating things..." followed by other alternate conjectures.

I don't feel the bro-ness.

On a technical level, they say it is a bigger model and that they have done something new in pre-training. How big? Or what’s new? They don’t say.

The 4whoa days were the good old days of OpenAI, and these are the "we are focusing on GPUs" days of OpenAI. At least, that’s what Sam says. He says he is focused on the long game. I think he is right, but in the meantime, OpenAI is constantly underwhelming.

The Christmas launches were underwhelming. Sora, that we got, was underwhelming. And now this. I still believe in Sam, but I can no longer justify paying $20 for ChatGPT. He needs to build something better if he wants to charge that kind of money.

I canceled last week. But I am positive I will renew it before the year ends. As I said before, I believe in Sam's ability of doing remarkable things.