Goodfire researcher Nick Cammarata believes studying AI neurons and circuits directly may offer a clearer path to understanding models than previously assumed. In an October 16 clarification, he expressed optimism but stressed that large language models are still far from fully mapped.
His essay, introduced the previous day, presents a working hypothesis: better tools and modern models may make their internal workings easier to investigate. A more detailed technical companion paper is still being completed. Cammarata calls for explanations to be tested experimentally.
Neel Nanda, Google DeepMind's mechanistic interpretability lead, sees useful ideas for practical debugging but not a decisive advance toward complete understanding. He worries that simplified neuron maps could miss important information, particularly in weak activations.
OpenAI researcher Adam Majmudar initially responded with strong optimism, then qualified his prediction. He meant controllability, not solving every alignment problem, and called his initial one-year timeline ambitious. This was his assessment, not a company commitment.
Cammarata himself cautioned against excessive expectations on October 16. He considers the approach promising but says further tools and new ideas may be needed. Whether it can deliver reliable control of large AI systems remains unresolved.