Context layers don't violate the bitter lesson — search is half the point
verrsane · x · 2026-11-22
Eddie Zhou argues context/semantic layers aren't counter to the bitter lesson: search is half of Sutton's point. Like b-trees surviving faster CPUs, even 0ms inference leaves you I/O-bound when fanning out across SaaS APIs — and proprietary state means those API pipes only narrow. The fix: a local read replica plus derived indexes built by offline LLMs, with slow incremental sync instead of runtime parallel search. Quoted tweet calls this "lap 2 of prompt engineering."
More from AGI Musings
- If AI says 'X likely treats pancreatic cancer,' how would we verify it? — Afinetheorem · 2026-11-22
- The real bottleneck for AI discovery beyond math may be verification — Afinetheorem · 2026-11-22
- Burkov: Linux analogy explains why no one can build a business on fine-tuning open models — burkov · 2026-11-22
- AI review finds 74 critical flaws in math papers in minutes, community hesitant — cgarciae88 · 2026-11-22
- Where 59 AI-economics PhDs landed: 37 research professorships — Afinetheorem · 2026-11-22
- AI makes creativity cheap — judgment and trust are the new scarcity — rschmelzer · 2026-11-22