Perception Efficiency Is the Next Optimization Frontier

2026-09-03

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

Why it earned a slot

Across three completely different domains, the same pattern is emerging: the bottleneck isn't compute, it's attention. Language models can now control their own attention, reading only the fraction of the KV cache that actually matters instead of scanning everything. ShallowStream applies the same principle to video — index shallow, answer deep — for streaming understanding. Discriminative world models for web agents take a third cut: instead of generating full future states, they sample candidate actions and rank predicted outcomes, which is fundamentally more efficient than full rollout. The common thread is that the next generation of gains isn't coming from bigger models or more compute; it's coming from smarter perception — doing less work to extract the same signal.

This has major implications for edge deployment, real-time systems, and any context where latency or memory matters. The models that win won't be the biggest; they'll be the ones that see least and understand most.