DeepMind paper: chip sim waveforms as SQLite lets agents ace 95.33% of verification queries
rohanpaul_ai · x · 2026-11-16
A Google DeepMind and Columbia University paper introduces Back-to-the-Future (BTTF), an agentic framework for post-simulation chip debugging.
- Chip simulation dumps can exceed 10GB of text; no LLM context fits them and they offer no searchable interface.
- BTTF distills these unstructured dumps into a normalized SQLite database, then a multi-agent pipeline translates natural-language verification queries into schema-aware SQL, checks design rules, and audits answers while correlating signal anomalies with versioned RTL repos.
- On 150 expert-labeled questions it hit 95.33% execution accuracy with Gemini 2.5 Pro; specialized multi-agent setups scored 17.3% higher on quality audits than a single agent, at 1.96x latency.
- Practical takeaway: when agents drown in huge logs or traces, move the data into a queryable database instead of stuffing it into the prompt.
Related event: DeepMind Framework Automates Chip Debugging with 95% Accuracy(2 posts)→
More from coding & agent
- User demos /swarm: 12 parallel agents research, code and critique to build a 3D model — lishali88 · 2026-11-16
- Anthropic Launches Claude for Manufacturing With 17 Skills and 4 Connectors — emmanuelvivier · 2026-11-16
- Building an autonomous game-clip pipeline with Codex without burning usage limits — gondorzabitasi · 2026-11-16
- llama.cpp already supports /v1/realtime for local speech-to-speech chat with LFM3-Omni — ngxson · 2026-11-16
- AI agent swarm of 20 Claude models proves first nontrivial lower bounds for Naslund's crossing problem, formalized in Lean — burny_tech · 2026-11-16
- ywza: MCP server lets AI control your PC via accessibility tree — Important_Earth6615 · 2026-11-16