Rumors have emerged about a purported new GPT-4.5 language model for ChatGPT. However, they are very suspicious: many clues point to a set-up.
There rumor has circulated on social networks, like X (formerly Twitter), but also among some communities on Reddit. It is materialized by a screenshot or quotes repeating its content. The general idea ? OpenAI, the American company behind ChatGPT, has reportedly finalized an even more efficient version of its chatbot’s language model.
Her name ? GPT-4.5. It would be, according to allegations circulating onlinea new version of its large language model (LLM, for large language model in English). An LLM is a central element for the functioning of a tool like ChatGPT. It constitutes the engine, which allows the conversational agent to respond to requests from Internet users.
A screenshot purportedly taken from the pricing page at OpenAI suggests that this GPT-4.5 model is “ the most advanced » of OpenAI and that it “ provides multimodal capabilities across language, audio, vision, video and 3D, as well as complex reasoning and multimodal understanding “. In addition, GPT-4.5 would be available in three versions.
In artificial intelligence, multimodality refers to the ability of a system to process and interpret information of various types (audio, video, photo, text, etc.) and to generate a rendering in other data. ChatGPT has this ability. This also means that the model has been trained on a wider variety of content — text, but not also images, for example.
An implausible leak
However, there are signs that suggest the greatest caution regarding information on this GPT-4.5. First, there is no mention of this model on the service pricing page or in any other section of the site. There may be mentions on the forum, but these are questions from Internet users which have nothing to do with an OpenAI announcement.
Is this then a failure on the part of the company, which would have updated its page too soon to mention GPT-4.5? Here too, caution is required: there is no trace of this model either in Google’s cache or in a page captured by the Internet Archive site, which records the entire web. In any case, nothing notable has been noticed in the last few days.
Another point that goes in the direction of a charade: Sam Altman, the founder and CEO of OpenAI, also contested the authenticity of the GPT-4.5 leak. A denial that is worth what it is worth: some could argue that the boss of OpenAI would never confirm such a leak, even if it was true. However, this rebuttal must be linked to other speeches by Sam Altman.
The first thing to note is that OpenAI has already recently presented a new version of its natural language processing system, which is already a significant leap compared to GPT-4, the presentation of which dates back to month of March 2023. It was in November 2023, with the presentation of GPT-4 Turbo.
It is this computer model which is today the most modern in the company’s offering. However, would it have been launched in November by OpenAI if it also had the imminent release of GPT-4.5 a month later in the pipeline? What is the point of presenting GPT-4 Turbo on stage during the first “OpenAI Dev Day” conference, so that GPT-4.5 is launched, visibly without fanfare?
The timing appears doubtful to say the least. Historically, GPT versions are spaced a few months apart (eight between GPT-4 Turbo and GPT-4, one year between GPT-4 and GPT-3.5, almost two years between GPT-3.5 and GPT-3, and and so on). It would be a first for the group to release two major advances for its sculpin in quick succession.
Last argument which weakens the idea of a GPT-4.5: when Sam Altman talks about the future, it is rather a question of… GPT-5. In any case, it is on this ground that we come to request it. This summer, a trademark registration on GPT-5 was spotted in the United States. More recently, the boss of OpenAI confirmed working on GPT-5. There was no mention of an intermediate version.
Would you like to receive the latest news around artificial intelligence, ChatGPT and language models? Then subscribe to our Artificial newsletter! Concocted by an AI, it is then proofread and verified by the editorial staff before arriving in your inbox.
Subscribe to Numerama on Google News so you don’t miss any news!