Building a House Without Blueprints: When Interpretability Tools Work — and When They Don’t
Inside the Black Box: Cracking AI and Deep LearningShow Notes
This episode of Inside the Black Box: Cracking AI and Deep Learning explores a new theoretical framework that unifies sparse autoencoders (SAEs), transcoders, and crosscoders — and what it tells us about when mechanistic interpretability actually works.
We start by demystifying these tools and how they use sparse features to uncover internal concepts and computations in large language models, from DNA detectors to deception circuits in Claude 3 Sonnet. Then we introduce the linear representation hypothesis and the geometry of concepts as directions in activation space, along with the challenge of superposition when thousands of concepts must fit into limited dimensions.
Finally, we dive into Tang et al.’s recovery theorems, the compressed sensing roots of their approach, and why these results matter for using SAEs as a reliable “microscope” on model internals, especially in the context of fine-tuning and LoRA experiments. Along the way, we confront the uncomfortable possibility that the linear picture may break down at frontier scales — and what that would mean for the future of interpretability as a safety strategy.
More from Inside the Black Box: Cracking AI and Deep Learning
After thirty-odd podcast interviews about AI, Arshavir Blackwell found the same four questions coming up with very different hosts: Will AI take my job or kill us all? Why does AI writing all sound the same? What should I teach my children? Is it actually thinking? He talks about stepping outside the AI bubble, and about the two questions no host asked on their own: where your data goes, and what AI really costs.
A follow-up in the past-tense series testing whether U-shaped learning, model size, or measurement explains irregular verb behavior—and finding the language matters too.
This episode challenges the old idea of a single language box in the brain, using aphasia studies and hospital control groups to show how language processing is more distributed than once thought. It also explores how different languages rely on word order or verb marking in different ways, and why those differences matter when the brain is under stress.
Thirty years ago, the authors of Rethinking Innateness argued that grammar could come out of a learner with no grammar built in. Nobody could check it. Now there's an instrument. A reading of the Inside the Black Box essay on what Elman, Bates, and Karmiloff-Smith would make of transformer language models — attention as a lookup, structure nobody installed, and the part that isn't there.
Can a model learn past tense the way children seem to, if you remove the tokenizer from the equation? This final installment tests raw letters and raw sound, probes whether a U-shaped pattern can be forced, and shows why the answer points to distributed, frequency-driven learning rather than a discrete rule.