{
  "id": 10850785,
  "title": "PSSA: A non-transformer language model written from scratch in Rust",
  "url": "https://urgent.news/2026/09/30/pssa-a-non-transformer-language-model-written-from-scratch-in-rust",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-30T03:19:54.000Z",
  "source": {
    "name": "Hacker News",
    "slug": "hacker-news",
    "url": "https://github.com/Sparticle62ops/pssa"
  },
  "original_language": "en",
  "account": "PSSA is a language model that operates without using transformer architecture. It processes text one token at a time via a recurrent state-space layer, maintains a set of episodic memories, and modifies certain aspects of its own weights during execution. Developed entirely in Rust, PSSA does not rely on PyTorch, TensorFlow, or any other machine learning frameworks. It demonstrates superior performance compared to a transformer model when trained on the same dataset and hardware parameters. PSSA learns at a faster rate and generates text at a quicker pace. While a transformer computes scores for every pair of tokens in the context, PSSA retains a single, fixed-size state throughout the sequence, enabling it to complete tasks within fewer training steps. The performance difference between the two models is evident when evaluating their training loss on a 198,939-token sample that neither model has encountered before. Despite both models being subject to the same constraints, PSSA consistently achieves lower loss values, indicating its ability to generalize better. When generating text, PSSA completes 200 tokens in a fraction of the time required by the transformer model. However, it is important to note that these findings are based on training the models on a single entry-level GPU within a limited timeframe. To explore further possibilities and assess the impact of larger parameter scales or more advanced hardware, additional computational resources are necessary.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}