PSSA: an attention-free, self-updating LLM built in Rust
The pitch: linear cost, 12x faster CPU generation, and lower perplexity than a same-size transformer — at 1.5M params, in dependency-free Rust you can read.

Copy markdown
Attention-free, and it learns while it runs
PSSA drops attention for a recurrent state-space core, a 512-slot episodic memory with hyperbolic retrieval, and "plastic" weights that update mid-inference before folding back into the base model. Cost grows linearly with sequence length, not quadratically — the real draw for long-context and on-device work.
The numbers it reports (at toy scale)
On a 1.5M-param model over 12.7M tokens, the author reports held-out perplexity of 54.4 vs 83.8 for a same-size transformer, 24.1% vs 18.0% next-token accuracy, and 226ms vs 2,735ms to generate 200 tokens on CPU — roughly 12x faster. Self-reported and tiny, but every gradient is checked against a CPU reference to ~3e-8.
No PyTorch, GPL-3.0, cargo build — what it's for
It's hand-written linear algebra in Rust with a CUDA training path and a scalar CPU reference — no PyTorch or TensorFlow anywhere. A checkpoint ships, but the author is upfront that quality is poor at this scale. Read it as a clean, dependency-free blueprint for post-transformer ideas to run this weekend, not a model to put in production.