Ricker

@0xRicker

Why moving the expensive LLM out of routine agent decision loops makes systems 193x faster. Autoregressive token generation belongs at the reasoning layer, not the tool socket.

OrcaRouter

@orcarouter

When the decision space is bounded, searching it removes 1–2 orders of magnitude of inference overhead. JEV testing proves non-autoregressive classifiers beat greedy sampling.

Andrej Karpathy

@karpathy

Revealed latent demand (no thinking, single token, low latency intelligence) that was under-invested into because of higher intelligence. Swarms freeze without fast reflexes.

Will Depue

@willdepue

Sure it's just a classifier, but it's a zero-shot classifier with frontier-ish intelligence. Real-time agent reflex layers are where all latent infrastructure demand is flowing.