---
format: "aidr-story-markdown/v1"
id: "076c86c65c6f2bf67ff314a0ebe40819ff730d0ab157ed49e03441dcb5afa429"
canonical_url: "https://aidr.today/076c86c6?lang=en"
title: "Goodfire Identifies \"Neuron Communities\" for Arithmetic and Date Recall in LLMs"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-17T17:37:38.000Z"
category: "Research"
topics: ["llm"]
source_urls: ["https://marketbrief.now/ai/update-goodfire-identifies-neuron-communities-for-arithmetic-and-date-re-f4f74e18","https://huggingnews.com/ai/update-goodfire-identifies-neuron-communities-for-arithmetic-and-date-re-f4f74e18"]
summary: "Goodfire has identified \"Neuron Communities,\" small groups of neurons that collaborate to perform specific tasks inside large language models. The research startup found that these small clusters are responsible for common behaviors such as arithmetic, recalling dates, or producing text in all capital letters. The finding comes amid an ongoing dispute with researchers at OpenAI, Google DeepMind, and Anthropic over how legible AI neurons actually are. Critics have challenged Goodfire's previous claim that neural networks are functionally sparse, arguing that a 2.9% increase in model loss—which Goodfire called modest after ablating 92.2% of the weakest neuron activations in Alibaba's 35B parameter Qwen3.6 model—would actually represent a massive loss in training compute."
---

# Goodfire Identifies "Neuron Communities" for Arithmetic and Date Recall in LLMs

> [Open the canonical story](<https://aidr.today/076c86c6?lang=en>)

**Published:** 2026-11-17T17:37:38.000Z
**Category:** Research
**Topics:** llm

## Summary

Goodfire has identified "Neuron Communities," small groups of neurons that collaborate to perform specific tasks inside large language models\. The research startup found that these small clusters are responsible for common behaviors such as arithmetic, recalling dates, or producing text in all capital letters\. The finding comes amid an ongoing dispute with researchers at OpenAI, Google DeepMind, and Anthropic over how legible AI neurons actually are\. Critics have challenged Goodfire's previous claim that neural networks are functionally sparse, arguing that a 2\.9% increase in model loss—which Goodfire called modest after ablating 92\.2% of the weakest neuron activations in Alibaba's 35B parameter Qwen3\.6 model—would actually represent a massive loss in training compute\.

## Sources

- [Story source](<https://marketbrief.now/ai/update-goodfire-identifies-neuron-communities-for-arithmetic-and-date-re-f4f74e18>)
- [Story source](<https://huggingnews.com/ai/update-goodfire-identifies-neuron-communities-for-arithmetic-and-date-re-f4f74e18>)

