The models page is clear about the choices of GPT 3.5 turbo and GPT 4 models, and it appears that the last 4-digits number represent the date in MMDD of the snapshot that was taken for that version. E.g.
gpt-4-0613: Snapshot of gpt-4 from June 13th 2023 [...]; Training Data: Up to Sep 2021
This seems to be congruent with the other release announcement dates:
But for us, API consumers, the date that seems to be more relevant is the snapshot of the training data – e.g. both *-0125 and *-1106 have training data “Up to Apr 2023” and these versioning dates seem like they’re only useful for those who have access to their source code.
The version numbers have a y2k problem except it is a y1 problem. Otherwise they are pretty clear when you know they are based on MMDD.
Unlike what the page professes, the models and their operational quality is not a “snapshot”; they continue to be actively updated past the date shown with reinforcement learning and other undisclosed changes. At best it represents a complete model pretraining from corpus - why it took two months for the fixes in 0125 to emerge.
API developers should not considers themselves “consumers” (although that is seen in practice). The usage is governed by the business agreement terms.
API developers should not considers themselves “consumers"
I understand how the word “consumers" can be seen as “business consumers”, which was not my intention.
By “us API consumers” I was referring to the mental concept that we’re writing code that “consumes” the API, to highlight that (AFAIK) we don’t have access to the original source code that is only available to the API developers (the “producers”). Say that I had access to OpenAI’s repo on Github or such: then, the “code date snapshot” would be a bit more meaningful to me, as looking at the git history tree I could see what commits made into that release version based on the date or so.
Unlike what the page professes, the models and their operational quality is not a “snapshot”; they continue to be actively updated past the date shown with reinforcement learning and other undisclosed changes. At best it represents a complete model pretraining from corpus - why it took two months for the fixes in 0125 to emerge.
I guess you’re starting to answer another question I have and I was not explicit about: what does “TRAINING DATA: Up to Apr 2023” really mean? I had an idea that OpenAI engineers and scientists built a massive model based on a slice of the then public internet content they web-crawled up to around that date (is this what you meant by “corpus”?), so asking questions like “who is the current prime-minister of XYZ” or so could only be accurate as that month. But in contrast with the idea that GPT learns from user feedback such as rating, “rejected answers”, Chat.GPT input, etc, your idea that they are updating it makes sense to me. So at first, we should expect that the answer questions would likely to be accurate as the training data date, with the caveat that some answers could be up-to-date if the updated model got enough input from feedback and/or OpenAI’s explicit update…
Right now, there are only two answers of cutoff date: through 2021, or through early 2023 (only in models called “preview”).
Compilation, scraping, and preparation of 45 terabytes (at last public disclosure) of knowledge data is a major undertaking of its own, the first step of the months and millions of dollars pretraining task of building an AI model, and knowledge preparation and even training order essentially could make a model that behaves quite differently, needing new red-teaming and followup techniques to refine it. This training data is what makes the AI even able to speak English or Farsi.
I expect even more eyes are on OpenAI now about the choices they make of copyrighted works (which is almost everything after 1928) to consciously include in new model training.
That is likely why we don’t see common progressive steps forward in knowledge cutoff, and that which was recently added is likely a well-curated subset and training continuation. It is far easier to tune an AI to emit functions and to deny impossible tasks, or extend its context length, and call that a “version”, than it is to start all over and make GPT-5.
You can propose changes. There was already a big change. We didn’t get a text-davinci-004 model, in the series of model sizes ada-babbage-curie-davinci.
