OpenAI staffer says it's a hallucination

I'm getting the same as the shared chat, ChatGPT-4 consistently (on 10 out of 10 rerolls) claiming it is gpt-4.5-turbo, with no custom instructions.

I'm not the kind of guy to take what an LLM says about itself on the face value, but this does seem like a strange kind of thing to be hallucinating consistently, unless there is some clear reason for it. I'm curious what the NO buyers think is the reason.

@Lovre turbo is low temperature so it's easy to reproduce hallucinations

@Lovre It knows about "gpt-3.5-turbo" since the knowledge cutoff was advanced; and it knows it's GPT-4. So it's not totally crazy to hallucinate "gpt-4.5-turbo".

Especially a term like "ChatGPT With Browsing" might indicate to it that it's a more advanced model. Hard to say.

@Lovre I think GPT-4 turbo was once internally named GPT-4.5 and somebody forgot to update something, or the data they used to train it to say that is outdated. Also, don't forget that in all these transcripts it first calls itself GPT-4.

@Lovre Web or mobile?

If on web, you can open dev tools and inspect the network tab, check the api call being made.

That of course says nothing about what is happening no the backend, but for me the "conversation" call has "model: 'gpt-4'".

@DavidBolin I also replicated it and got "gpt-4" as model id. This isn't conclusive because "gpt-4" is an alias that is expected to update to the production variant of "gpt-4-turbo-preview" after they shake out the bugs.

It's worth checking if the backend fingerprint changed. I don't know how often they update it or have a historical list, but if it's ~every couple days, then an update on Thursday correlating with an updated fingerprint would be positive evidence.

@Mira I agree that isn't conclusive (and in fact I use "gpt-4" in production precisely to ensure that it will not be deprecated, for that reason.)

@Lovre It could think it's 4.5-turbo purely due to the date in the prompt changing. ChatGPT is usually updated before the API is.

Thanks everyone, I updated on your thoughts and immediately sold my position with only relatively slight losses, whereas if I had waited they obviously would have been much more substantial.

The link is not evidence until they do it without custom instructions.

I feel it got smarter 2-3 days ago but maybe I am imagining things because of all the rumors flying around

@MartyE is this meant as a joke?

@Soli No I got a pre release access to gpt 4.5 and asked if to Play tic tac toe. It wasn’t very good I pay for ChatGPT plus

@MartyE I am still not sure if you are joking or confused

You most certainly did not get early access to anything. The image shows a custom gpt named GPT-4.5Unofficial and not really gpt-4.5

hahaha

@TheBayesian Haha I see now lol