AI That Designs Its Own Chips: Ricursive's Anna Goldie and Azalia Mirhoseini

Summary

Anna Goldie and Azalia Mirhoseini, co-founders of Ricursive Intelligence, present their startup's thesis that AI should design the chips that train AI and that chip design itself should be automated to close a recursive self-improving loop. They describe AlphaChip's track record, a three-phase plan spanning AI-accelerated design tools, democratized custom silicon, and vertical integration into chips and models. They argue AI-driven customization can unlock massive performance gains and a Cambrian explosion of workload-specific chips. In Q&A, they discuss the economics of specialized chips and why AI-generated layouts look organic.

• Ricursive Intelligence is building AI for chip design and chip design for AI.
• AlphaChip, a deep RL chip-layout system, was used in Google TPUs, Axion CPUs, Pixel phones, AV chips, and by MediaTek.
• Phase one aims to accelerate chip design with tools up to 100,000x faster than today's commercial software.
• Phase two aims to democratize chip design through a designless platform for custom silicon.
• Phase three would vertically integrate into Ricursive's own chips and models.
• The founders argue customization and AI co-design will unlock major performance gains and a Cambrian explosion of chips.
• Q&A covers the cost and scale of specialized chips, compute as a design knob, and AI-generated organic chip layouts.