The Empire Strikes Back

Deepseek-1.5B proves that Chinese AI researchers have more ingenuity than US policymakers give them credit for. Smaller is sometimes smarter.

The assumption in Washington has been that brute force compute is the only path to AGI, and since we control the chips, we control the timeline. Deepseek challenges that. By optimizing architecture and training data quality, they've achieved performance comparable to much larger models with a fraction of the compute.

This suggests our export control strategy may be leaking. If the frontier is defined by efficiency rather than scale, then 4090s might be enough.