Today in AI

The interpretability crowd had a banner day, and it's about time we paid attention to what models are actually doing inside rather than just marveling at their outputs. Two papers in particular suggest we're peeking under the hood at exactly the right moment. First, researchers discovered that OCR-specific attention heads in vision-language models aren't merely reading text off images—they're general-purpose semantic feature detectors that align surprisingly early with language concepts. This matters because it reveals a deeper integration between visual and linguistic processing than architectures nominally suggest. Then there's the gut punch: Sparse autoencoders on Gemma-3 and Qwen3.5 show these same VLMs *already contain* the features needed to detect harmful memes, but suffer a "readout gap" where relevant internal signals get misrouted rather than missing entirely. The implications are uncomfortable—our models aren't ignorant, they're just structurally bad at using what they know, which may be worse for safety.

On the efficiency frontier, DeepSeek-v4.1 Flash is pushing KV cache compression to 4x without quality collapse, specifically targeting the long-horizon agent workflows that everyone assumes will dominate 2025 compute budgets. Meanwhile, AURA tries to solve the fascinating coordination problem of multiple autonomous agents sharing physical radio infrastructure in Open RAN networks—less flashy than chatbots, but as AI systems gain bodies in the physical world, stability guarantees start mattering more than benchmark bragging rights. The scaling research on how architectural interventions can modify scaling exponents also deserves notice; if we can genuinely bend the compute-performance curve through design rather than mere scale, the current CapEx arms race starts looking less inevitable. Rounding out, the tokeniser design decomposition confirms what many suspected—the search algorithm dominates the objective, a subtlety with outsized impact on every model's efficiency and even its apparent capabilities.

Bottom line: The most important AI progress today isn't new capabilities—it's finally understanding why existing capabilities fail and how to rewire access to knowledge models already possess.