Some ChatGPT users report that ChatGPT with GPT-4 has access to current information until September 2023. Previously, the end of training data was reported as September 2021 or January 2022.
In the past, ChatGPT has hallucinated incorrect end dates for its training data or misquoted the system prompt, so take it all with a grain of salt.
However, with the current data that has now surfaced up to September 2023, ChatGPT was able to partially reference events that took place after the training data ended in early 2022, such as Will Smith slapping Chris Rock at the Oscars.
However, this update does not seem to affect all ChatGPT models. Some users still get an earlier training end date of January 2022 or September 2021, and ChatGPT does not comment on the Oscar slap. The same thing happened in my test of GPT-4 (January 2022, no slap).
One theory is that OpenAI is known to run a mixture-of-experts model for GPT-4 with 16 networked AI models that may have had different training data or are being successively updated.
Depending on the access and query, the activated model could be different, which could explain the different training dates. OpenAI has not yet commented on the different training end dates.
New GPT-4 model leaked on Reddit
A leak on Reddit also shows that OpenAI is presumably training a new "copilot-gpt-4-2" model, which is up-to-date until March 2023. There may be some blending, A-B testing, or a partial rollout. The new model is expected to have a 32K context window.
OpenAI might provide some clarity at its developer conference on November 6. A new model like GPT-4.5 or even GPT-5 will not be announced, according to OpenAI CEO Sam Altman, but there should still be significant new features.
Since DALL-E 3 has already been released and developers are likely to be particularly interested in efficiency and an updated GPT-4, model improvements could be a possible announcement.
