self-learning-model

Models whose weights change while they run. Predict → get surprised → rewrite a small, bounded part of yourself → never forget the base. Proven with falsifiable experiments (PROOF.md, p ≤ 4e-04).

pip install self-learning-model          # zero-dependency core
pip install "self-learning-model[qwen]"  # + the self-learning Qwen3-VL runtime

1 · Strands-expert Qwen3-VL-2B that keeps learning

A Qwen3-VL-2B post-tuned on the entire strands-agents codebase (strands probes NLL 4.85 → 2.22, 8/8 improved), plus a plastic layer that adapts at inference — with a provable off-switch.

from slm.qwen import StrandsPlasticQwen                     # needs [qwen] + HF_TOKEN

m = StrandsPlasticQwen.from_pretrained()                    # cagataydev/strands-qwen3-vl-2b
print(m.chat("How do I create a custom tool in Strands Agents?"))

for doc in your_stream:
    m.observe(doc, learn=True)   # predicts; if surprised, rewrites its fast weights
m.reset()                        # exactly back to the strands-expert base

Verified: continual out-of-distribution stream NLL 6.18 → 5.37 while strands expertise is untouched (Δ −0.01). Model: cagataydev/strands-qwen3-vl-2b (private).

2 · Zero-dependency core (Python + browser, bit-identical)

The mechanism in ~600 lines of stdlib Python, mirrored in vanilla JS (Δ≈1e-15):

from slm import SelfLearningLM
lm = SelfLearningLM.from_pretrained("cagataydev/self-learning-model")
lm.observe("novel text", learn=True)     # weights change — no training step

Live demo: docs/index.html — watch a transformer's weight matrix rewrite itself in your browser as it reads.

Is it really learning? Yes — proven.

Six skeptic attacks, six closed doors (PROOF.md):

Honest negatives included: single-prompt in-context (softmax attention already wins), emergent in-context learning at tiny scale, and the finding that ~⅓ of naive test-time-training gains are mere calibration — a confound we control for.

How it works

frozen base (instinct — never updated, can't forget)
  + tiny plastic LoRA (experience — updated every step at inference)
  + surprise gate (learn only when wrong) + EMA decay (bounded, stable)
loss = next-observation prediction error   ← the free label from reality

The stability–plasticity dilemma is resolved by the decay: without it the model learns hard but forgets (retention +7.09); with it, it learns and retains (+0.03).

More

Built on ideas from TTT (Sun et al.), Titans (Behrouz et al.), fast-weight programmers (Schmidhuber), DeltaNet — and @karpathy's atomic GPT.

MIT.