AI news
May 1, 2024

Is "gpt2-chatbot" Supposed To Be GPT-4.5 Or GPT-5?

Amysterious AI chatbot, gpt2-chatbot, appeared on the internet.

Jim Clyde Monge
Jim Clyde Monge

It’s been months since OpenAI released GPT-4, its latest and most capable multi-modal language model to date. However, many users have been raising concerns over its degrading performance over time. Even I had already canceled my ChatGPT subscription because I find other language models, such as Claude Opus, to be more powerful and reliable.

As a result, many are questioning when OpenAI will release the next version of the GPT series. The anticipation is growing, and people are eagerly awaiting the new capabilities and improvements that the next iteration will bring to the table.

Well, two days ago, a mysterious “gpt2-chatbot” appeared on LMSYS Chatbot Arena. This unexpected arrival has generated significant interest and speculation among AI experts and enthusiasts due to its impressive capabilities, which are believed to be comparable to, or even surpassing, OpenAI’s GPT-4.

Initially, the origin of the model was unclear. But Sam Altman posted a tweet that basically confirmed that the GPT2 is from OpenAI.

What is GPT2

The “gpt2-chatbot” has been observed to perform well in various tasks, including solving math riddles, exhibiting advanced reasoning skills, and demonstrating a human-like tone and mathematical competency.

For instance, X user Andrew Gao showed a demo where the model solved a very difficult math question in one shot.

That’s very impressive.

To be clear, GPT2 is not GPT-2.

GPT2 is a newer and more advanced version of the GPT series, while GPT-2 is an older model.

Here’s another example that showcases GPT2’s capability.

Respond with just a numerical answer:

When I tried the same prompt in Claude Opus, I didn’t get the same result. This suggests that GPT2 is more effective at recognizing and completing mathematical patterns compared to some other AI models.

Respond with just a numerical answer: 1+2=2 2+3=4 5+6=
Image by Jim Clyde Monge

I also tried it on Gemini Advanced, but it didn’t grasp the pattern as well.

Respond with just a numerical answer: 1+2=2 2+3=4 5+6=

Try GPT2 Yourself

Right now, gpt2-chatbot was removed from the LMSYS Org platform due to unexpectedly high traffic, suggesting that it may have been overwhelmed by the volume of users trying to interact with it.

Even when it was still accessible, it was slow and severely rate-limited — you only get 8 turns.

Is it GPT-5?

There is widespread speculation that OpenAI is preparing to release its next major model, GPT-5, this summer. With GPT-4 having been released just over a year ago, companies have been racing to develop and deploy better models to remain competitive in the rapidly evolving AI space.

The competition is intense, and everyone is curious whether GPT2 is a sneak peek of what’s to come with GPT-5.

At present, there is limited information available about the gpt2-chatbot. It’s impossible to say with certainty whether this is indeed GPT-5. The mystery surrounding this model is part of what makes it so captivating, and we’ll have to wait and see what details emerge in the coming weeks and months.

Final Thoughts

If OpenAI is indeed the company behind GPT2, they may have developed a more efficient method for fine-tuning language models. This new approach could have enabled them to train GPT-2, a 1.5B parameter model, to perform remarkably close to GPT-4, which is an order of magnitude larger and more costly to train and run.

However, OpenAI has not made any public announcement regarding this GPT2 model. In the coming weeks, the creator and origins of the gpt2-chatbot will likely become public. Stay tuned for that.