Sami Petersen

[Elliot Lipnowski @ElliotLip]

Really cool project. Innovative use of "standard" game theoretic modeling ingredients to tackle a novel, important application.

[Zach Mazlish @ZMazlish]

I'm going to be helping @pawtrammell teach a course on the economics of transformative AI this summer at Stanford — if that sounds of interest to you, applications are open!
https://t.co/DMYyJu9oBN

[Global Priorities Institute @GPIOxford]

GPI Research Affiliate @pawtrammell is giving a summer course on the Economics of Transformative AI at Stanford this year. More info and applications here: https://t.co/6Li9aqeFc6. Applications close 21 February.

[Sami Petersen @SamiCPetersen]

@gwern I’ll briefly note on this point that we don’t need training to be hard to distinguish from testing or from deployment.

[Sami Petersen @SamiCPetersen]

New draft!

We develop a technique for safely delegating to a strategic AI that may be misaligned with your objectives—without assuming you can restrict its choices, punish it, or otherwise control it once deployed.

How? Sequential information design with imperfect recall. 🧵

[Global Priorities Institute @GPIOxford]

Imperfect Recall and AI Delegation, the new working paper by Eric Olav Chen, @Alexghersen and Sami Petersen is now available to read here: https://t.co/WFM2kTkqgb

[Sami Petersen @SamiCPetersen]

@gwern Great points. The mechanism isn't easily applicable to tasks like those. It’s intended for task setups that we can design, deliberately, to satisfy the requirements. Our discussion in the draft is limited; we’ll expand on it soon on AF/LW/EA. We’d love to get your thoughts there.

[Marginal Revolution @MargRev]

A new paper on the economics of AI alignment https://t.co/UaHvIE4xlu

[Sami Petersen @SamiCPetersen]

To see how we exploit these dynamics to achieve the results above, check out the working paper!

https://t.co/UuNvyKln7h

[Global Priorities Institute @GPIOxford]

Imperfect Recall and AI Delegation, the new working paper by Eric Olav Chen, @Alexghersen and Sami Petersen is now available to read here: https://t.co/WFM2kTkqgb

[Sami Petersen @SamiCPetersen]

This lets us partly screen for misaligned types (only misaligned ones can fail a test) and, even if we don’t succeed, we discipline them in deployment (even misaligned ones play nice with positive probability).