The mystery of GPT-2

Day 121 / 366

There is a new chatbot going around, which has everyone speculating about it’s origins.

There is a website called LMSYS Chatbot Arena where users can compare various LLMs. A few days ago a mysterious model named ‘gpt2-chatbot’ was added there. It wasn’t a big deal at the time, because at first it just looked like a fine-tuned version of the old GPT-2 model.

But slowly this started gaining traction on social media, when users reported that it was outperforming GPT-4 in terms of logical reasoning. This is where the speculations began that it might be a hidden release of the new GPT 4.5 or GPT 5 model.

This model's coding, as well as mathematics skills, are way better than that of GPT-4. It also did well in things where LLMs generally fail, for example generating ASCII art.

To add fuel to the fire, OpenAI’s CEO Sam Altman tweeted this today —

I think this is definitely some sort of a marketing tactic. Given the hype that has been created around this model, I am sure we will hear from OpenAI or some other company soon claiming responsibility for creating it.