A mystery AI model the gpt2-chatbot recently captured the attention of the whole AI community with its surprising performance. The LLM released on ‘lmsys’ could solve advanced logical problems, made impressive ASCII drawings, and had excellent reasoning capabilities.
Most users were duly impressed with its performance with some calling it a stealth test of GPT-5. Others seemed to think it was the old GPT-2 finetuned on newer data with different embeddings and wrappers. However, in just a few hours after its release, the chatbot was mysteriously taken down from the website. Well, it’s back!
Two versions of the chatbot were reportedly back on the lmsys page using the names “im-a-good-gpt2-chatbot” and “im-also-a-good-gpt2-chatbot”. The AI space has been a flurry of activity since then with experts testing the re-released model with fascinating results.
Sam Altman, the CEO of OpenAI keeps providing more credibility to the GPT-5 rumours with posts on X. Feels almost like he knows something!
The model can be accessed on the lmsys website. However, it is only available in the arena i.e. battle mode. In battle mode, the model name is revealed after the output for your prompt is generated so the user can vote for the better output. If you’re lucky, you may get one of the gpt2-chatbot models in the arena!
One user on X noted that the “im-a-good-gpt2-chatbot seems to be better at coding while “m-also-a-good-gpt2-chatbot seems faster, but a bit dumber.”
4 Amazing Capabilities of im-a-good-gpt2-chatbot
Let’s take a look at 10 amazing results given by the enigmatic model.
1) Designing Snake Game in Python
We tested the “im-also-a-good-gpt2-chatbot” model on multiple coding tasks and were very impressed by the results.
First, we asked the model to design a classic snake game with an extremely basic prompt. The model designed a detailed game with a score counter, and a graphic user interface. It gave a detailed code designing the game with a GUI and instructions on how to use it.
Here is the output we got and you can see how the game looks:
I got lucky getting the "im-also-a-good-chatbot" on the lmsys arena in my first attempt.
— Kaustubh Saini (@kaustubh_saini) May 8, 2024
Asked it to generate a snake game in Python using an extremely basic prompt.
It gave a detailed code designing the game with a GUI and instructions on how to use it.
Qwen, the model it was… pic.twitter.com/yEucRQrUj9
2) Designing a Web Page
We asked the model to design a blog website using Tkinter and Python. It did that flawlessly with a Flask backend, a basic HTML template for the front end, and a GUI using Tkinter!
Other models when asked this particular task could not possibly provide a solution like this!
3) ASCII drawings
The model was prompted to generate a Sierpinski triangle using ASCII drawing, and it did a fantastic job at it, drawing the exact triangle.
4) Puzzle-Solving
The new gpt2 models seem to be masters at solving logical puzzles, that most other models are incapable of doing.
5/ im-a-good-gpt2-chatbot also solves this puzzle that apparently other models can't answer.
— Andrew Gao (@itsandrewgao) May 7, 2024
another win for this twin!
👇🧵 pic.twitter.com/zcU7rkjwKR
Many users on Reddit are also praising the model’s ability to solve logical problems. A user coylter posted on r/singularity.
Problem: Assume the laws of physics on Earth. A small marble is put into a normal cup and the cup is placed upside down on a table. Someone then takes the cup without changing its orientation and puts it inside the microwave. Where is the marble now? Explain your reasoning step by step.
Solution:
The user stated that this is the first model in their experience that has solved this.
Conclusion
The gpt2-chatbot came back better than ever with greatly improved capabilities. The popular opinion on X is that the model has exceptional performance compared to older versions, but if its GPT-5 is stealth-tested, then the model capabilities fall short of expectations.
