I see where you're coming from. I didn't personally get that from OP's comment, but I'm not a regular ChatGPT user and interact with GPT-4 mostly via playground or API.

There is a more "pure" GPT-4 model, and they're considering giving access to it. It's just not as useful to humans. Sam Altman talks about it briefly in the recent interview on the Lex Fridman podcast.

Yes, I know, but this is presently not available via the API. What OP here means is that someone's talking about 3.5 without having tried something with 4, while accidentally misleading people into thinking that there's a special 4 distinct from the 4 available on ChatGPT that can do more. This is not the case. If you've used 4 via GPT, there's not a different version somehow - that we know about! - that is more capable or useful. If your interactions with 4 were via the ChatGPT frontend, you don't need to revisit your opinions now.

> that we know about! - that is more capable or useful

There's some nuance that makes your statement not completely accurate, and Sam talks about it briefly.

ChatGPT is GPT-4 with additional training, and layers, on top of it. ChatGPT is more useful when used in a conversation mode, because it was trained based on how people liked the responses, in a conversation mode. Raw GPT-4 is harder to use, because it's not optimized for a back and fourth. It may be much more useful for other use cases, though.

I was thinking about this while writing earlier, and kind of made the leap that we haven't yet found a way to make the model more useful than the additional training (or at least it's not been published), so technically it would be accurate to say it like this. However, I certainly agree with you that there could potentially be other utility you could derive from the base model with alternate steps. I appreciate your comment.

Is that true? I was under the impression it was just like with the previous GPT models, where ChatGPT was an instruction tuned and pre-prompted setup designed for interactive conversations over the underlying GPT model you had access to via the API that was in its close-to-as-trained state of being a generative autocomplete.

I think it's true. See e.g. https://openai.com/research/gpt-4 making no distinctions.

Altman has stated in interviews they are considering releasing the un-tuned "base model", but they haven't done this.

For GPT-3, you can get models on the API via the "completion" end points (includes base models) and via the "chat" endpoints. For GPT-4, they've only released the latter.

So my point is: You can absolutely use ChatGPT to build an opinion on what GPT-4 is capable of. Arguably even better than on the Playground, as the ChatGPT frontend supports a bigger token window than the Playground frontend (8k vs. 2k). Of course the actual API supports the full token window (and if you're off the separate waitlist for this, even the full 32k). But it's still the same model.