Demo: training models to read K/V vector breadcrumbs left by other instances
Sauers_ · x · 2026-11-11
Sauers demonstrates that you can train models to read 'neuralese' K/V vector breadcrumbs dropped by other model instances — instances communicating via KV-cache vectors instead of natural language. Idea credited to @astarchai; implementation details not yet shared.
More from coding & agent
- Fish Audio launches Live 3 full-duplex voice model, builds a CRM voice agent in 5.8 seconds — thisdudelikesAI · 2026-11-11
- Running 6 concurrent ds41-flash agents: persistent enough to need pi-durable coordination — dbreunig · 2026-11-11
- OpenAI roundup: architecture details of four GPT-Live voice agent builds — pbbakkum · 2026-11-11
- Ex-Apple keyboard engineer Ken Kocienda open-sources Tug, a macOS AI coding app in beta — Scobleizer · 2026-11-11
- Carbon tops Abundant Labs agent benchmark with 14/20 vs Opus 5.5 and GPT-6 — amankhan · 2026-11-11
- VS Code Copilot falls back from Computer Use and Python to PowerShell + COM to build an Excel workbook — dfinke · 2026-11-11