{
  "id": 14776531,
  "title": "Neural networks are clean",
  "url": "https://urgent.news/2026/10/15/neural-networks-are-clean",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-15T16:04:01.000Z",
  "source": {
    "name": "Hacker News Best",
    "slug": "hacker-news-best",
    "url": "https://already-clean.goodfire.com/"
  },
  "original_language": "en",
  "account": "Recent research has shed new light on neural networks, suggesting they are more \"clean\" and easier to study than previously thought. After extensive investigation into neurons and circuits within image and language models, the author has concluded that network mechanisms themselves are the legible building blocks, and the outlook for mechanistic interpretability may be far better than anticipated.\n\nThe prevailing notion of polysemanticity and superposition, which once led many researchers to believe neural networks packed concepts arbitrarily, has now been deemed misread. Instead, the observed phenomenon is mostly caused by a motif known as recursive generalization, which is less concerning.\n\nThis discovery opens up two potential approaches to mechanistic interpretability. The first involves direct study of the network's neurons and circuits, observing data flow and conducting experiments at agentic scale. The second approach employs a list of hypothesized features to extract them through increasingly sophisticated reconstruction objectives, often disregarding the original model.\n\nHistorically, the field began with the former approach but later shifted to the latter due to the perceived complexity of neural networks. A notable example was a neuron in an image model that seemed to have multiple unrelated jobs, such as responding to both cat faces and car fronts. The leading explanation for this polysemanticity was superposition, which proposed a new learning principle where a neural network stops grouping related things together and begins folding unrelated things together when running out of neurons.\n\nHowever, the author now believes that superposition was the wrong explanation for InceptionV1, a popular image model. Architectural issues in later layers caused them to overflow, while weak interpretability tools amplified these problems, making the late layers appear incoherent. Subsequent architectural improvements designed to enhance deep layers' learning capabilities inadvertently made them more interpretable.\n\nThe author's own experience with InceptionV1 exemplifies this. Initially, they found few interpretable neurons in the last layer. However, upon applying ordinary improvements and without altering the model's size or training data, 99-100% of neurons became easily interpretable, and the model exhibited overall legibility.\n\nThe author emphasizes that the canonical vision examples of polysemanticity and superposition were erroneous. In 2019, Chris Olah and the author observed a car feature appearing in unrelated dog neurons, interpreting it as the network hiding the car in a high-dimensional space for downstream neurons to retrieve. However, this was merely ordinary generalization and composition, not a hidden mechanism.\n\nMoreover, modern language models are far more interpretable than commonly believed. In a 2023 study, the median neuron in a small one-layer model received a score of 0 out of 14 on the rubric, indicating that the rater couldn't form any hypothesis. Another automated interpretability project found that approximately 90% of neuron poles in language models like Qwen3-32B had one main job, with a smaller proportion having two related jobs or being ununderstood.\n\nIn summary, recent findings suggest that neural networks are inherently more clean and legible than previously thought. By directly studying network mechanisms and implementing architectural improvements, researchers can now approach mechanistic interpretability with greater confidence.",
  "summary": "https://twitter.com/nickcammarata/status/2110757819427574056 Comments URL: https://news.ycombinator.com/item?id=50105647 Points: 241 # Comments: 73",
  "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."
}