MIT's 'AI telepathy' KV-cache transfer beats text by 40% accuracy at 4x speed, across model vendors
LucaAmb · x · 2026-10-29
MIT researchers built a mechanism for AIs to communicate by directly passing internal states (KV-cache) instead of text — reportedly 40% more accurate and 4x faster, and it works across vendors (Llama 3 ↔ Qwen 3 ↔ Gemma ↔ Granite, etc.). davidad predicts that by 2028Q1 this will end CoT monitoring as we know it, improve inference-time scaling curves, and make context-window size far less important.
More from AGI Musings
- "If AI devalued your CS degree, you're doing CS wrong" sparks UK grad job debate — georgemillo · 2026-10-29
- Richard Sutton's 26-word Bitter Lesson gets a scaling-first rewrite — samsja19 · 2026-10-29
- AI Twitter debate: if AI can replace you, you weren't doing anything novel — wavefnx · 2026-10-29
- Arvind Narayanan: why Claude's 'superintelligent' writing style backfired — random_walker · 2026-10-29
- Nature: AI is making data-analysis and modelling jobs obsolete, hands-on roles safer — mdancho84 · 2026-10-29
- "Everything white collar was trivial all along": an AI practitioner on agent-fleet burnout — teodorio · 2026-10-29