Show HN: Watch an LLM program a neural net's weights instead of training it (github.com)

A new project from MIT CSAIL introduces a groundbreaking approach to programming neural networks, allowing users to manipulate a neural net's weights directly using a language model, without traditional training methods like gradient descent. Dubbed “Bit,” this untrained multi-layer perceptron can be loaded in a web browser to demonstrate various behaviors by blending its weights with pre-programmed “smart pills.” Users can adjust the weight mixture to see real-time changes in Bit's performance across different games, showcasing a novel interactive model for AI manipulation. This innovation is significant for the AI/ML community as it challenges conventional training paradigms by enabling direct weight programming. The project leverages the Opus framework, which facilitates brewing custom weight configurations, while the sophisticated integration with Claude’s AI model allows for fine-tuned adjustments and immediate feedback. Developers can enhance their neural networks locally, share their custom solutions, and explore various game environments that employ the system. By reimagining how AI models can be constructed and modified, this project opens new avenues for experimentation and creativity in machine learning.