All topics / mechanistic interpretability
Topic
You are reading the free Skim layer. Read unlocks the synthesis and sources.
Mechanistic interpretability
The study of understanding complex abstractions in neural networks.
1
episodes
3
thinkers
5h
of conversation
3
books & papers
4
terms defined
The neighbourhood: mechanistic interpretability and the ideas it travels with. Drag to roam, click a star for the episode, click a neighbour to travel.
Drag to roam · scroll to zoom · click a neighbour to travel · click a star for the episode
From foundational to frontier
Climb the spectrum. The most accessible conversations come first.
Start here
ACCESSIBLECOREFRONTIER
The lexicon
Every term the guests lean on, in plain language. Read one in full, or filter to find it.
What the corpus says
The throughline across every conversation that touches this idea.
Dario Amodei predicts AI will reach PhD-level capabilities by 2026-2027, driven by scaling laws.
AI models like Sonnet 3.5 have shown rapid improvement, achieving a 50% success rate on SWE-bench.
AI systems could potentially reach ASL-3 by next year, indicating significant autonomy and risk.
Constitutional AI uses principles to guide model behavior, enhancing safety and interpretability.
Mechanistic interpretability in neural networks seeks to understand complex abstractions and deception features.
Voices on mechanistic interpretability
5 standout quotes from across the corpus.
Go read
3 books and papers cited across these episodes.
For the specialist
What experts find new
3 expert-level takeaways for a specialist reader.
At the frontier
Still unresolved
2 open questions flagged across these conversations.
The thinkers
Who takes this idea on, by how often they return to it.
