Engineering
Subquadratic attention gets its first serious test
Subquadratic attention is a sparse LLM design now showing a 56.2x speed test win at 1M tokens. Treat it as promising, gated infrastructure.
#long context
1 story tagged long context.