tgakathunderr / BIM-2

What is BIM-2?

BIM-2 is a biologically inspired language model built entirely in Python that learns sequences without backpropagation, GPUs, or pretraining. It uses Jaccard Surprise as its sole training signal, scaling Hebbian learning rates directly by prediction error. The system runs a two-layer cortical architecture with top-down feedback on CPU hardware, consuming under 200MB of memory. Users interact through a REPL where they can teach facts, correct mistakes mid-conversation, and query learned knowledge.

Why is it gaining traction?

The anti-deep-learning crowd finally has a working alternative. BIM-2 proves you can build sequence memory using only Hebbian plasticity and sparse distributed representations, not gradient descent. The correction mechanism is particularly clever: type "WRONG: liam RIGHT: zara" and the system applies maximum-strength learning to the correct response. The built-in benchmark that tests direct recall and 2-hop inference on fictional company data gives you measurable proof the model actually learned something. No GPU required means you can run experiments on a laptop during a coffee break.

Who should use this?

Researchers exploring biologically plausible learning algorithms will find this the most complete open-source implementation of HTM-style cortex with Hebbian plasticity. Neuromorphic computing enthusiasts who want to experiment with apical feedback and synaptic homeostasis without custom hardware. Anyone building toy knowledge bases where explainability matters more than accuracy. Not suitable for production language tasks or anyone needing state-of-the-art performance.

Verdict

BIM-2 is a serious research prototype with a well-documented architecture and working benchmarks, but the 1.0% credibility score and 10 stars reflect its niche appeal and early-stage status. The code is clean and the concepts are sound, but test coverage, documentation, and community support are minimal. Try it if you want to learn about biological learning or experiment with alternatives to backpropagation. Do not use it for anything that requires reliability or performance.