mayaskOpenAI has proclaimed that it has rolled out a fresh version of its GPT-4o large language model to power its ChatGPT chatbot. However, it has chosen not to disclose precisely how the updated model varies from its predecessor.
“To be explicit, this is an enhancement to GPT-4o and not a novel frontier model,” the company declared on X (formerly Twitter) on Monday.
Recommended Videos
There’s a new GPT-4o model available in ChatGPT since last week. Hope you all are relishing it and give it a try if you haven’t! We reckon you’ll fancy it.
— ChatGPT (@ChatGPTapp) August 12, 2024
“We’ve brought in an update to GPT-4o that we’ve discovered, through experiment outcomes and qualitative feedback, ChatGPT users tend to favor,” the company noted in its Model Release Notes. “It’s not a new frontier-class model. Even though we’d love to tell you exactly how the model responses are distinct, determining how to granularly benchmark and communicate model behavior improvements is an ongoing area of research in itself (which we’re working on!).”
In the absence of specific particulars from the company, unsurprisingly, many users have begun to conjecture about the nature of the changes and whether they represent new features. X user @misaligned_agi ventured onto social media to suggest that the new update had implemented a multistep reasoning approach rather than an entirely new model.
OpenAI promptly quashed that line of thought, with a spokesperson informing VentureBeat that it was not, in fact, a new reasoning process and that the behavior observed by @misaligned_agi could have been triggered by the structure of their prompt.
Other users also expressed their theories on social media, arguing that GPT-4o has recently started behaving in subtly different and superior ways, and that its image-generating quality has improved. “For the first time in a long while, it offered better ‘vibes’ on an output than 3.5 Sonnet,” observed X user @mattshumer_.
After allowing users to make their best guesses at defining the new iteration, which OpenAI is naming ChatGPT-4o-latest, the company added a few meager details on its Models page on Wednesday.
Described as a “dynamic model continuously updated to the current version of GPT-4o in ChatGPT,” GPT-4o-latest has a knowledge cutoff of October 2023, and can handle 128,000 tokens, or 96,000 words, per conversation, just like the previous GPT-40 version. It can output 16,384 tokens, or 12,288 words, matching the newer GPT-40-mini model and approximately quadrupling what the older GPT-4o can manage.
Unfortunately, concrete statistics like these do not offer an insight into what the new model is truly capable of, and, apparently, neither will its developers.
