What you'll learn

• 1 (00:03) **Context: AI Safety Landscape** - The host frames the episode around a shift from dreary AI safety headlines to a potentially major positive breakthrough.
• 2 (01:02) **The Paper That Changed the Conversation** - Nick Camarada's "Neural Networks Are Already Clean" becomes an instant talking point, with OpenAI's Rune calling it a potential major moment in scientific history.
• 3 (02:58) **The Core Problem: AI is Grown, Not Built** - The host explains why interpretability matters: AI systems are trained, not programmed, making their inner workings a black box.
• 4 (04:03) **The Race to Interpretability** - The field's history from 2017 to 2025 is charted: initial excitement, then despair as models appeared too tangled to read.
• 5 (05:20) **The New Claim: AI is Mostly Readable** - Nick's essay argues that the inside of AI is mostly readable already, and we've been looking at it wrong.
• 6 (05:47) **The "Cat and Car" Detector: A Case Study** - The host explains the famous detector that lit up for both cat faces and car fronts, which was the go-to proof of scrambled AI insides.
• 7 (07:50) **Three Fixable Mistakes** - Nick proposes three causes for the apparent scrambling, all of which are fixable.

+ Full timestamped outline available in the app