The New GPT2 and All You Need to Know!
A new mysterious chatbot appeared out of nowhere on public chatbot websites. Reportedly, it’s GPT-4 level.
Anew noise has been circulating the AI community for the past day about a new language model called gpt2-chatbot (Not gpt-2-chatbot, apparently it’s important). There was no information as to where it came from and who made it, however, there is more to know about it now, and here’s a recap:
✨This is a paid article. If you’re not a Medium member, you can read this for free in my newsletter: Qiubyte.
It’s GPT-4 Level
Everyone seems to agree on one thing: the new gpt2-chatbot is in the same ballpark as gpt-4-turbo. This unexpected level of capability was what circulated talks about gpt2-chatbot in the first place. However, there is no way to be sure unless there are benchmarks and leaderboard updates that show where it stands.
One interesting observation by Andrew Gao is that gpt2-chatbot has better agentic abilities than GPT-4, that is being able to chop up big tasks into smaller ones and plan a way to solve them.
If this is the case, we could speculate that it might be GPT-4, fine-tuned by OpenAI for AI agent purposes. Aside from that, it’s not as powerful and impressive as we expect GPT-5 to be, as some like to believe. Seems like we have to wait longer to see that happen.
Inmy own few prompts, I can also confirm that it’s something-to-keep-an-eye-on level. My test was to have gpt2-chatbot play Flappy Bird for me (I tested this with GPT-4 and it failed but follow me to read the article any day now).
The model responds similarly to GPT-4: it can represent the game in text but fails to follow the game’s rules and mechanics. The pipes aren’t moving left as they should be in the Flappy Bird game, and new pipes aren’t randomly generated. So in this test, it’s almost an exact function as GPT-4-turbo.
Other users have different takes. Some say that it’s superior to GPT-4-turbo, especially in coding tasks…
However, for some reason, many people are using it to create ASCII art :D
Where Did it Come From
Now we can almost be certain that it’s from OpenAI, as pointed out by Sam Altman:
There is a bit of OpenAI history with another GPT2 that goes back to 2019.
GPT-2 is a large transformer-based language model with 1.5 billion parameters, trained on a dataset. of 8 million web pages. GPT-2 is trained with a simple objective: predict the next word, given all of the previous words within some text.
This is part of an OpenAI blog post dating back to February 2019 that addresses a new model called GPT-2. This is the predecessor to ChatGPT, released in 2022. It was a text-completion model, and not a chatbot, which means it worked based on completing sentences and not according to the prompt-response model we use today.
Could it be that the released gpt2-chatbot is GPT-2 retrained on a new dataset? Being only 1.5 billion parameters, gpt2-chatbot is only as fast as GPT-4, and the capabilities are not at a 1.5B level. So this seems very unlikely.
Another possibility is that — just as the name suggests, GPT2 is really a second version of the common GPT models we know today. This could be due to a new method of training, a novel architecture, or essentially any innovation that is big enough to call it Not GPT, but GPT2.
How To Try
Right now you can try this chatbot for FREE (by the time you read this, it might not be available anymore). Bear in mind that the workload on this chatbot may be too much. I have experienced errors when trying to work with it. There might also be messaging limits applied, so be wise in your messages (at least avoid “hi”). To test it out:
go to
https://chat.lmsys.org/
head over to “Direct Chat”
select “gpt2-chatbot”
It’s Already JAILBREAKed
JAILBREAK is a method of using prompts to bypass the system restrictions that keep a model “aligned” and safe to use. Well, they’re not so safe as proven by Pliny the Prompter 🐉 who has been able to jailbreak almost any LLM to date. Only hours after release, it is already jailbreaked, giving Meth recipes and planning world domination.
It’s safe to wait for the release of benchmarks/more information to conclude where this new chatbot stands against other peers. I hope it’s a new GPT paradigm, but even if not, it’s still satisfying to have another player in this league!
🌟 Join +1000 people learning about
Python🐍, ML/MLOps/AI🤖, Data Science📈, and LLM 🗯
follow me and check out my X/Twitter where I keep you updated Daily: https://twitter.com/itsHesamSheikh
Thanks for reading,
— Hesam