Image: reddit.com · Introducing the New Frontier of the GPQA-Dumb
Bongochat: a deliberately dumb chatbot trained on an RTX 5070
A tongue-in-cheek language model trained from scratch with Karpathy's nanochat on a home GPU in about three hours, with Claude Code running the training.
Credits
- Reddit poster TheOnlyVibemaster Source ↗
- Repository account ninjahawk Source ↗
Earliest dated source:
The full story
The maker used Claude Code to train a model from scratch with Karpathy's nanochat, locally on an RTX 5070, in about three hours, and released the weights on GitHub as Bongochat.
The results are played for laughs. Asked to solve the unified field theory, it repeats the word "theory" 50 times. On Math-500 it left almost every answer blank, it scored 0 out of 500 on coding, and it suggested graduate-level calculus for a simple addition problem, then forgot it had done so.
How it was done
Tools
Themes
Loading themes…
Sources
From a short post by the maker, so the details are as they tell them.
- Reddit: Introducing the New Frontier of the GPQA-DumbReddit poster: TheOnlyVibemasterPosted
The maker's post on r/ClaudeAI.
If it disappears: archived copy of Reddit: Introducing the New Frontier of the GPQA-Dumb (opens in a new tab)
- The code: github.comRepository account: ninjahawk
The project on GitHub.