Nick Cammarata at Goodfire published a 26,000-word blog post. He claims neural networks may be easier to understand than previously thought. His work suggests neurons group concepts by shared patterns, not random tangles. He demonstrated this on an older image model, Inception v1. He now plans to test modern large language models. Some researchers call it a major breakthrough. Others, like Stanford’s Chris Potts, remain skeptical. Jacob Hilton at Alignment Research Center warns the public may overestimate the progress. Neel Nanda at Google DeepMind says neurons are only one piece of the puzzle. Cammarata admits his findings do not solve everything. He argues the AI safety community needs optimism. He believes full understanding is possible with enough time and tools.

• 26,000-word blog post
• 0.01% of top-tier LLM inner workings understood