I Can See Everything

Substack · Watch/Read · (gathered 2026-10-21)

• Zvi Mowshowitz analyzes Nick Cammarata’s “I Can See Everything” essay, which claims that large language models (LLMs) are far more interpretable than previously thought, with most neurons representing a singular concept (“One Neuron, One Central Concept”) rather than being “polysemantic” (packing unrelated concepts). Zvi acknowledges the work’s importance but cautions against overblown optimism.
• Cammarata argues that apparent polysemanticity is often “recursive generalization” (related offshoots) or misinterpretation, contrasting with the prior belief in “superposition.” However, critics like Eliezer Yudkowsky argue that even if individual neurons are interpretable, this is a “primitive playground” that doesn’t solve the “hard problem” of understanding how an AI mind works or predicting its behavior for alignment.
• The “Colorado-Tornado” neuron example challenges Cammarata’s claims, showing a neuron strongly activating for two seemingly unrelated concepts, suggesting the “cleanliness” might be overstated and the theory may not hold up universally.
• Zvi warns against the “Most Forbidden Technique” (training on interpretability signals) to prevent AIs from obfuscating their internal states, and notes that interpretability breakthroughs are “dual use,” potentially advancing capabilities as much as safety. While a positive update for neuron interpretability, it’s the “beginning of the work,” not a solution for alignment.