Vision models are beginning to reveal neurons and circuits that represent human understandable concepts

Vision models are beginning to reveal neurons and circuits that represent human understandable concepts showing how ideas emerge from input words interact with other ideas and ultimately help generate actions Laniakea is just one init away and the path forward depends on building better models than anywhere else while clanker startups push intense 80 plus hour work weeks and agentic coding becomes a transformative paradigm where clanker agents act as true collaborators rather than autocomplete tools Clear instructions and direct conversation with the machine often outperform elaborate prompts and techniques like sparse autoencoders from signal processing are helping identify cleaner combinations of neurons with the discovery of 30 billion features marking a major step forward while still representing only a small fraction of what remains to be understood