PSSA: A non-transformer language model written from scratch in Rust

PSSA is a non-transformer language model written entirely in Rust from scratch that uses a plastic state-space architecture instead of the transformer's attention mechanism. In benchmarks on the same corpus and with matched parameters, PSSA achieved significantly better performance than a transformer, reaching lower training loss (3.98 vs 4.43) and better generalization on unseen data while generating text roughly 12 times faster on CPU. PSSA's key advantage is its linear computational cost that grows with sequence length, compared to transformers' quadratic cost, achieved through a fixed-size recurrent state and episodic memory lookup rather than re-reading the entire context at each step.