Capability Is Leaking to the Edge at an Alarming Rate

2026-08-29

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

A latent flow transformer generating 128x128 face images on a $5 RP2350 microcontroller. CritICL using small language model failure modes to improve large model reasoning at inference time — essentially, the small model's mistakes are teaching the big model how to think better. These aren't separate trends; they're the same force moving in opposite directions. Capability is compressing downward (tiny models on microcontrollers) while also being amplified upward (small models boosting large model inference-time performance). The CritICL paper is particularly interesting because it inverts the usual scaling narrative: instead of 'make the model bigger,' it's 'make the model smarter by showing it what smaller models get wrong.'

For builders, the practical takeaway is that your deployment target is no longer a constraint that limits your ambition — it's a design parameter you can optimize around. If you can't run the 70B model, run the 7B model and use its failure modes to guide a better inference strategy. The ceiling is your imagination; the floor is a five-dollar microcontroller.