AI is getting smarter. But do we actually know how they work?

Stanford researcher Aryaman Arora takes us inside the science of AI interpretability, a field of research where we try to understand what’s happening inside language models.

We discussed Anthropic’s research of a hidden internal neural workspace in large language models, Golden Gate Claude, user awareness research at Transluce, and why different models are starting to develop increasingly distinct personalities.

We also talked about the various research traditions, what sparse autoencoders and causal interventions can reveal, and why we can actually edit a model’s internal thoughts.

Aryaman Arora is a PhD researcher at Stanford, advised by Christopher Potts and Dan Jurafsky. His research focuses on AI interpretability, language models, and linguistics, and he has also worked at organizations including Transluce and Redwood Research.

Timestamp:
00:00 What is interpretability research?
03:11 J-Space: Do models have an internal reasoning space?
07:34 Do models have internal reasoning space?
08:56 Can chain-of thought tell us what a model is thinking?
12:07 Interpretability, alignment, and trusting AI
16:10 Sparse autoencoders: translating AI’s internal representations
19:34 Two competing approaches to interpretability
23:32 The compute problem
24:11 Linguistics
30:06 Interpretability benchmarks
31:28 Gold Gate Claude: can we edit an AI’s thoughts?
34:09 Transluce and user awareness in frontier models
36:09 Weird chat and discovering strange model behaviors
37:02 Is it all about edge cases?
38:36 Can interpretability explain hallucinations?
40:27 H3 paper
43:05 Academia vs. industry
48:15 What’s next

About Valley101:
Valley101 is an independent media that tells deeply reported stories about the people, companies, and technologies shaping the future. We cover AI, semiconductors, startups, and the global technology industry through documentaries, explainers, and conversations with the people building it.
Subscribe for more stories from Silicon Valley and beyond.

Contact: video@sv101.net