Korva is a scroll of the internet’s best ideas.
Publish every AI's training history
Verification between rivals
The proposal comes from the AI 2040 Plan A scenario. Because the US and China do not trust each other, any deal has to include verification, such as inspectors who visit each other's data centers and count the chips to rule out a hidden cluster. Data centers that serve customers keep today's privacy protections. Research clusters, where new AIs are trained, are made maximally transparent, with devices between the GPUs that log activity and publish it.
Removing the incentive
Today's race is a prisoner's dilemma: if I don't do the dangerous thing, the other company will. With everything visible, a company that starts researching a risky direction is seen at once, and everyone else can free ride on its research instead of paying for their own. Nobody gains an edge from the dangerous work, so nobody has a reason to start it. If something scary shows up, companies can stop together, since everyone can see that nobody is going on.
Opening the work
Publishing also opens the work to the whole scientific community, who can red-team proposals and run experiments on the AIs. A company is biased by its incentives, and a government auditor may be stretched across several companies, short on experience, or captured.
How have rivals verified each other's weapons?
The 1987 INF Treaty between the US and Soviet Union allowed inspectors on each other's missile bases and at factory gates. Ronald Reagan's phrase trust but verify became tied to it. Inspections under the treaty continued until 2001.
How could hardware logs resist tampering?
Secure chips can hold cryptographic keys that sign records, so any alteration breaks the signature. Computer security uses this idea already: a trusted platform module in many laptops measures what software loads at startup and signs the result. A 2023 paper by Yonadav Shavit on compute governance proposed using similar hardware in AI chips to verify training runs.
More from The Joe Rogan Experience
