The community of artificial intelligence experts is on tenterhooks with a mysterious AI chatbot called GPT2, a large linguistic model about which very little is known, which has been available for only three days, but which has sparked enthusiasm among researchers because it demonstrates a advanced reasoning and a great ability to deal with complex approaches. Although no one seems to know where it came from, everyone points to OpenAI, since the company’s own leader, Sam Altman, fueled that expectation by tweeting on Tuesday “I have a weakness for gpt2.” Are we facing a new great language model that will surpass what the GPT-4 that powers ChatGPT can do?
Everything is a mystery about GPT2. Its promoters gave the research community access to the LMSYS platform, an open source project to evaluate large artificial intelligence language models, so that experts can classify their capabilities. But GPT2 has only been available for three days. Its performance suggests it could outperform any of the currently existing models, which has sparked speculation about whether it could be GPT-4.5 or GPT-5, OpenAI’s next big model.
Sam Altman, meanwhile, plays the game of distraction. Putting OpenAI (Open AI, in English) as the name of the company sounds like the saying “tell me what you brag about…”. During a visit to Cambridge, England, the executive was asked by a journalist if he already knew when GPT-5 would be launched. “Yes,” he replied while smiling without saying anything else. Some researchers believe that GPT2 is a boosted upgrade of GPT-4. In fact, in a response, the model stated that it is based on the GPT-4 architecture.
In addition to its deeper reasoning than other existing models, GPT2 is an accomplished specialist in what is called ASCII art, those drawings made with characters, which come from the era of typing. You can ask it to draw anything and the results are very good, unlike what other AIs can do -see drawing below compared to Meta’s Llama-3-. He solves complex mathematical problems with ease and has good reasoning skills. Among experts, some are betting that we are facing a very good update of OpenAI’s GPT-2 model (with hyphen), launched in 2019.
We will have to wait, although the wait is always enlivened by news that appears. This week, Altman gave his vision of what the ultimate AI app will look like: “a super-competent colleague who knows absolutely everything about my entire life, every email, every conversation I’ve ever had, but doesn’t feel like an extension.” That is, something that does things for you, without having to ask yourself every moment. On the other hand, current OpenAI applications such as DALL-E, Sora and ChatGPT are “incredibly stupid” in the words of Altman, who knows something about this.
