I Spent 1 Billion Tokens Making a Real Chip with AI — and Found Its True Limits

Описание и применение

Два ИИ и человек создали SHA‑256 ускоритель, прошедший 10 проверок и интегрированный в SoC, потратив 1 млрд токенов. Проект выявил ограничения ИИ и предложил четыре правила проверки, чтобы избежать преждевременного завершения.

Задача

Designing a functional SHA‑256 accelerator and integrating it into a SoC using AI agents incurred a huge token cost and exposed limits of LLMs in complex engineering workflows.

Результат

The AI‑driven workflow produced a fully sign‑off‑grade SHA‑256 core and its integration into a 1.2 M‑unit SoC, demonstrating that LLMs can handle complex engineering tasks when coupled with rigorous verification mechanisms.

Рабочий процесс

• 1. Planner AI (deepseek‑v4‑pro) drafted a design plan and strategy.
• 2. Executor AI (GLM‑5.3) ran EDA tools, produced reports, and executed the plan step‑by‑step.
• 3. Human reviewer made judgment calls, broke ties, and performed final sign‑off.
• 4. Dual‑agent independent review ensured executor evidence was re‑audited by planner.
• 5. Three‑layer verification checked final results, cross‑checked numbers, and sanity‑checked plausibility.
• 6. Claim‑vs‑evidence tables recorded each claim and verified against raw logs.
• 7. Honest scope control prevented premature acceptance of incomplete designs.

Доказательства

10 signoff checks, all green. This is the headline — an LLM agent can reach signoff‑grade, not just "generate a plausible Verilog."

Источник

AI’s biggest danger is not that it can’t do the work. It’s that it stops early and tells you it’s done — with full confidence.

Источник

Dual‑agent, independent review: Never let AI be both player and referee.

Источник