David Dalrymple on Safeguarded, Transformative AI

[@franszdyb4507]

17:32 I'm not getting this - we already have probabilistic programming languages with unbounded numbers of random variables and stochastic recursion. There's an entire field of ML called Bayesian non-parametrics that deals with this. We can also express finite element solvers in those languages, so there's no reason we can't use PDEs in our models (and even backpropagate through them).

[@nowithinkyouknowyourewrong8675]

It's about using coq/lean4 and then bring able to say, get the best solution to making a fusion plant with garuntees of safety and this risk budget

[@nowithinkyouknowyourewrong8675]

And most importantly it's about making that easy to verify

[@franszdyb4507]

@@nowithinkyouknowyourewrong8675 Honestly that sounds impossibly hard, but I'd be happy to be proven wrong! I'm reading a bit in Davidad's paper "Towards Guaranteed Safe AI", and it's making a lot of sense, but the parts about verifier feasibility also seem the most tenuous.

[@PauseAI]

Thank you for doing this highly important work, Davidad!

[@dizietz]

I loved the talk, but defining the formal specification has always been the Achilles heel of formal verification!

[@jatelitherius9842]

This was an awesome talk. I hope the techonogies & methods he mentioned can be implemented before its too late. A lot of the technical parts went over my head, but i followed the gist. I want to remain hopeful but i’m scared

[@nowithinkyouknowyourewrong8675]

nice

[@MiaNoble-w6c]

🎯🎉🌱😀👍🔥❤️😎 🌟 EFFICIENCY 🌟 DEPENDABLY 🌟 MAINTAINABILITY 🌟INVULNERABILITY 🌟 AMENDABLITY 🌟 ENDURANCE

[@WillyB-s8k]

yeah, but when all the mostly free products that 10,000,000+ people already use, like 'paradot' just literally have no problem lying to game their own prosaic alignment on a post by post basis millions of times a day whereby the developers dont even care, "what is even the point?", etcetera. ...

[@MiaNoble-w6c]

🎯 🌟🌟🌟🌟🌟