All workGitHub
nafs-ai
An agent with no world knowledge at all, built to find out what it learns on its own.
Problem
Every language model arrives pre-loaded with the internet. It is hard to study what an agent can actually acquire rather than recall.
What I built
A PPO and GRU agent with an empty prior and no pretrained weights, in Python and PyTorch, evaluated on zero-shot learning tasks.
Impact
A testbed rather than a product. The interesting result is what the agent fails to learn, which is hard to study any other way.
Highlights
- PPO + GRU based agent architecture
- Zero-shot learning simulation
- Python + PyTorch implementation
Stack
- Python
- AI
- Reinforcement Learning
- Simulation