An impressive new artificial intelligence model has suddenly appeared on the popular chatbot arena LMSys. This led to speculation as to whether this was a preview of a new model from a company like OpenAI, such as GPT-5.

Named GPT2, the robot was added to the arena without documentation or other information; those who encountered GPT-4 described it as more capable than GPT-4 and very good at reasoning.

Little is known about GPT2 other than its capabilities, and some users have run GPT2 on common benchmarks and found it to be near the top. This led to speculation that it was a preview of the new OpenAI model.

OpenAI CEO Sam Altman added fuel to the speculation fire by posting on X, "I have a soft spot for gpt2."

The new model appears in the LMSys arena as gpt2-chatbot. This should not be confused with one of OpenAI's earlier models, GPT-2 (hyphenated), which some speculate is a tweaked version of that smaller model.

Those who have tried this model have stated that in some responses it performed better than GPT-4, the current leader on the LMSys leaderboard and OpenAI's most powerful model. This includes tests conducted on several AI models.

Andrew Gao, a researcher at Stanford University and a leading AI expert, noted that the voice felt on the same level as GPT-4, not necessarily better, but different from OpenAI's model.

While it differs from GPT-4 in the way it responds, it is not a new model. He stated, "I have a feeling that tweaking GPT-4 would produce similar results."

It is not clear who built the GPT2 or where it came from. It could be a new startup coming out of stealth, it could be a group of researchers testing a tweaked version of an existing model, or, speculatively, it could be Open AI playing the gorilla marketing game.

Whether that is OpenAI's model is not certain, but several clues point in that direction. These include OpenAI's use of teaser-type tactics and some of the behaviors seen in GPT2.

Gao writes:" Someone reported that "this model, like other OpenAI models, is weak against certain special tokens and seems to be trained on the OpenAI family of tokenizers. In other words, GPT-4 is likely involved in the creation of the training data, even if it is not an OpenAI model.

In tests, GPT2 breaks learned conventions, creates ASCII art, and is particularly good at coding.

One promising theory is that Elon Musk is testing version 2 of the X-powered Grok language model as a way to make people aware that it is more than just a slightly upset chatbot.

We'll find out the origins eventually, but it's fun to speculate and it's nice to know that AI development continues at an innovative pace that surprises even the more jaded experts.