{
  "id": 18002502,
  "title": "deconstructed the Qwen3.5-0.8B update example",
  "url": "https://urgent.news/2026/11/03/deconstructed-the-qwen3-5-0-8b-update-example",
  "topic": "world",
  "section": "World",
  "published": "2026-11-03T18:37:38.000Z",
  "source": {
    "name": "Simon Willison from Simon Willison’s Newsletter",
    "slug": "simon-willison-from-simon-willison-s-new",
    "url": "https://www.seangoedecke.com/llms-are-not-superhuman-at-reading-and-writing-neural-network-weights/"
  },
  "original_language": "en",
  "account": "A recent paper claims that large language models (LLMs) like GPT-6 Astra can directly interact with neural network weights. The paper demonstrates that Astra can identify the games a small network was trained for by analyzing its weights, and even modify those weights itself. The standout result is that Astra successfully improved a 0.8 billion parameter Qwen language model's mathematics capabilities by editing its weights. This capability has the potential to enable recursive self-improvement, a significant advancement in AI research. However, the paper argues that this is not as impressive as it seems, as even a human ML researcher could perform these tasks. The paper highlights that identifying what game a model was trained for based on its weights might seem impressive, but it is not as remarkable as it appears. A heatmap in three lines of Python code can effectively convey this information without executing actual Python code. The most impressive example in the paper is fixing a 0.8 billion Qwen model's ability to handle mathematics. Qwen had a problem where it would get stuck in loops after processing 32 thousand tokens. To fix this issue, Astra wrote four pairs of reasoning texts, measured the weights contributing to the looping behavior, and then trimmed those weights to resolve the problem. This process is similar to the Heretic abliteration playbook, a technique commonly used by human researchers. The paper suggests that while the ability of Astra to make direct weight updates is impressive, it should not be considered a radically new capability. Instead, the best approach to updating model weights is to use the standard training process. The author concludes that the focus should be on improving training methods rather than on the ability of LLMs to directly manipulate model weights.",
  "summary": null,
  "key_points": [
    "Astra LLM can modify neural network weights directly",
    "Improves 0.8B Qwen model's math capabilities",
    "Technique similar to human ML researchers' methods"
  ],
  "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."
}