Vitalik: Restricting AI capabilities by purpose is better than restricting them by user identity.

Vitalik stated on the X platform that excluding high-end offensive network capabilities from the training data in Kimi 3.1 is acceptable. Differentiating by purpose is less negative than differentiating by user identity. Society needs time to harden systems, examine old code, and formally validate new code. If the first wave of major AI risks is not properly addressed, society could move towards a highly restrictive model environment, widespread KYC and surveillance, where users might be stratified based on credentials and nationality, receiving different levels of model access.