麻豆APP

鈥淲e Are Not (Training GPT-5) And Won鈥檛 For Some Time,鈥� Says OpenAI CEO Sam Altman

While speaking at an event at the Massachusetts Institute of Technology (MIT), OpenAI鈥檚 CEO and co-founder Sam Altman is currently not training the next version of their popular platform ChatGPT, GPT-5. GPT-5 is the presumed successor to the language model’s current GPT-4 which was released in mid-March 2023.

In late March, an open letter was published by the nonprofit Future of Life Institute and co-signed by top tech executives including Elon Musk and some well-known AI researchers. The letter called on AI labs around the globe to pause the development and enrolment of large-scale AI systems more so those more powerful than GPT-4.

The letter highlighted fears of the probable risks these systems pose to society and humanity. 鈥淲e call on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4,鈥� the letter read.

While speaking at the MIT event last week, Altman was asked about this letter. In response, he said that the letter was 鈥渕issing most technical nuance about where we need the pause.鈥� He went on to say openly that the letter, which claimed and implied that OpenAI is currently training GPT-5 is wrong.

See also

Tech & Science

Elon Musk Very Impressed By China鈥檚 Space Program

鈥淲e are not (training GPT-5) and won鈥檛 for some time,鈥� said Altman. 鈥淪o, in that sense, it (the letter) was sort of silly.鈥�

However, while this statement alludes that the development of GPT-5 is not underway, improvements are being made on GPT-4 continuously to expand its capabilities. Altman stressed that work on GPT-4 too does have safety implications that need to be looked at and addressed.

鈥淲e are doing other things on top of GPT-4 that I think have all sorts of safety issues that are important to address and were totally left out of the letter,鈥� Altman said.

According to The Verge, these comments by Sam Altman bring to light a significant challenge in AI regulation which is the apparent difficulty when tracking progress in AI development. Altman鈥檚 claim that they are not training GPT-5 is rendered virtually useless by his next claim that they are 鈥渄oing other things on top of GPT-4.鈥�

James Vincent of The Verge labels this as 鈥渢he fallacy of version numbers鈥�. This is the idea that numbered tech updates such as GPT-5 reflect a linear improvement in capability from previous versions. In this case from GPT-4 to GPT-5 which is not the case at all!