Today in AI

The mechanistic interpretability landscape just got upended, and the implications stretch far beyond the research lab. A team at Goodfire AI led by Nick Cammarata has published findings suggesting that neural networks are already dramatically more interpretable than the field has assumed — with Cammarata claiming 90% of Qwen-32B neurons carry single, discrete functions rather than the jumbled "superposition" of meanings that polysemanticity theory predicted. If this holds up under scrutiny, it undermines one of the central pessimisms that has slowed alignment work: the belief that neural networks are fundamentally inscrutable black boxes requiring heroic mathematical effort to unpack. The research emphasizes recursive generalization patterns as the key to this cleanliness, which — if replicable across architectures — could accelerate everything from automated red-teaming to reliable model steering. That said, the history of AI is littered with interpretability claims that didn't survive contact with larger or more complex systems, and Goodfire AI's commercial interest in interpretability tools warrants appropriate skepticism. Still, the confidence here is striking, and the field should be stress-testing this aggressively.

Elsewhere, we're seeing continued maturation across the embodied AI and world-modeling stack. 3DCodeVerse demonstrates that programmatic 3D generation — code as scene description — can dramatically outperform even capable visual systems like Claude Design in human evaluation, suggesting that structured representation still beats raw generation for complex spatial tasks. T2-z tackles the stubborn sim-to-real gap in robotics through test-time latent optimization, while GeoState advances view-decoupled 3D modeling by cleanly separating scene dynamics from camera motion — a conceptual move that feels increasingly necessary as autonomous systems need to reason about both what moves and what merely appears to. Less flashy but perhaps more indicative of where investment is flowing, NVIDIA and AI4Bharat's 2.8 trillion token Indic language dataset represents serious infrastructure commitment to language coverage that Western-centric development has long neglected, with an open-source GPU-accelerated pipeline that could lower barriers for other underrepresented language clusters.

**Bottom line:** Interpretability optimism is having a moment, but the real pattern is structural — AI research is converging on representation, verification, and clarification as the hard problems that generation alone cannot solve.