A version of Muse Spark Open was released today, fulfilling their original promise, though the model underperforms their Spark 1.3. The Tesla Roadster demo is also well worth watching—while not strictly AI, there is nothing cooler than a real-life “flying car.”
Today, our headline is a rare deep dive into interpretability and alignment. We last spoke with Goodfire back in February:
The First Mechanistic Interpretability Frontier Lab — Goodfire AI’s Myra Deng and Mark Bissell From Palantir and Two Sigma to turning Goodfire into the standard-bearer for applied mechanistic interpretability, Member of Technical Staff Mark Bissell and Head of Product Myra Deng are trying to…
…as well as our conversation with Anthropic last year on scaling the “Bitter Lesson” of interpretability:
The Utility of Interpretability — Emmanuel Ameisen, Anthropic Thanks to everyone who joined us for AIE — full livestream replays of Day 1 and Day 2 are now live, along with 6 other breakout streams. You can binge-watch them all this weekend! We’ll have a more comprehensive recap once we get some sleep.
Now, Goodfire’s Nick Cammarata has published a major essay outlining his core thesis: “Neural Networks Are Already Clean”—arguing that many known cases of polysemanticity and superposition (which produce tangled “hairballs” of weights that resist human explanation) dissolve upon closer agentic analysis:
