A Samsung executive has sparked rumours that OpenAI is about to double the size of its flagship large language model (LLM), ChatGPT.
At the Semicon Taiwan conference today, Dr Jung Bae Lee reportedly got up on stage and showed the audience graphic revealing key details of ChatGPT 5 - a model that will reportedly be blessed with PhD-level intelligence.
Although it is not yet clear whether the information was an informed guess or based on actual data, therefore constituting a leak, pictures of the Samsung executive's presentation inspired frantic social media discussion of what had just happened.
Many comments on X were posted by writers who believe the former claim: that the Samsung boss just leaked details of OpenAI's upcoming model.
The image appeared to show either that it would be made up of 3.5 trillion parameters - almost twice as many as OpenAI's current GPT-4 model - or between three and five trillion parameters, depending on how you view the blurry image.
The slide also appears to indicate GPT-5 can or will be trained on Nvidia B100 7000 chips and run HBM3E 8TB/s memory.
You can see the image in the tweet below:
Gpt-5 and a Samsung exec.
— Jimmy Apples 🍎/acc (@apples_jimmy) September 4, 2024
The noise is getting louder. pic.twitter.com/LoTag3iWy1
However, not everyone was convinced that they were seeing concrete data about the upcoming model.
"Why would Samsung have any official information about GPT-5?" an X user asked.
The veracity (or not) of the information is now being discussed heartily on Reddit.
Some posters queried the accuracy of the numbers, questioning whether the slide is a case of Samsung asking: "How do I get the biggest compute number to make my argument look better?"
"Samsung is one of the main developers of the type of memory that GPU use," another person wrote. "To clarify. His slide is basically saying: 'These models keep going up and we plan on making our bank account go up along with them'."
