newsfilter
← Interviews
80,000 Hours3h 9m

What the hell is going on inside neural networks? | Chris Olah

Chris Olah, Rob Woodland

Interview

Chris Olah leads Anthropic, a research organization founded to empirically investigate the internal mechanisms of large language models to ensure their safety and reliability. By reverse-engineering neural circuits and identifying abstract features such as multimodal neurons representing protected attributes, the team aims to detect dangerous failure modes and unknown unknowns that traditional testing misses. This approach relies on scaling laws and standardized tools like Microscope to build reproducible interpretability methods for increasingly powerful AI systems.