20VC: Anj Midha on Investing $300M into Anthropic | The Early Days of Anthropic & How 21 of 22 VCs Turned it Down | The Four Bottlenecks to Compute | What the China Has Smashed and Why We Should Be Worried

Anj Midha argues that while scaling laws in AI are not saturated, performance gains depend heavily on domain-specific context feedback loops, especially in underexplored areas like material science. He identifies four key bottlenecks to AI progress: context feedback, compute, capital, and culture—with culture being the most critical, as it enables algorithmic innovation. Midha shares how his lab, Periodic Labs, uses physical robots and real-world validation to close feedback loops and train models on proprietary scientific data inaccessible on the public internet.

You'll gain a rare, grounded perspective on the real bottlenecks in AI development from someone who's built labs, invested in Anthropic, and sees beyond the hype to where true progress happens.

1. 01The biggest bottleneck in advancing AI is not algorithms or compute, but access to high-quality, domain-specific context feedback data—especially in fields like material science where data is locked in labs and facilities.
2. 02Culture is the foundational bottleneck: mission-driven, flexible teams attract top researchers and naturally produce algorithmic innovation without being wedded to specific architectures.
3. 03Real-world performance of AI models, especially in science and coding, lags far behind benchmark results, necessitating real-world RL environments and operational traces to surface hidden failure modes.