AGI with Safety Guarantees - Berkeley 2024

[Music] so I am fono my PhD is in completely unrelated stuff but I do remember what it was like to be an expert in a thing at once and I will be talking about AGI with safety guarantees this is not really alignment and you will see why governance and social Choice Theory and things like this are highly relevant the current idea of sending AI agents out in the world leads you to potentially being able to improve systems around the same time that you can modify or over modify those systems irresponsibly or outpace any kind of human oversight that we have and so there's this intrinsic challenge about the fact that Society built up of review Systems and governance systems and putting more agents into the world doesn't necessarily strengthen those because we want the review portion to be feedback based on or built from Human input and so the solution that we're pursuing at Al Computing is making tools to set rules that AI has to prove that it's obeying there's this concept of guaranteed safe AI That's slowly building in uh interest adoption notoriety um where AI systems should have a separate auditable World model that they work with and you should be able to specify as a user your goals for the AI system in terms of that world model and then any action proposed by the AI agent is validated against those goals in this World model and this isn't something like this isn't completely revolutionary outside of the scope of what people are proposed in other spaces this is how your the FAA validates uh autopilot on aircraft for instance there are people in the research space who are proposing this architecture or things like it saying we should have these three criteria and most of this has been scoped to research what we're doing is trying to build specific prototypes that Advance this type of architecture with things that people can benefit from and use in a very short term we could have an architecture where specification is its own step generation is its own step and review is its own step and humans can look at each of these gray boxes and say this is a thing is this what I want and you have tools where you don't have to trust an AI system but it's easier to review something like a solution specification or evidence than it is to understand the process of generating the solution and understand the solution itself in this workflow with co-pilot you have to know how to code at least well enough to be able to interpret is this does this code do what I want whereas here you should be able to say is this specification what I want do I have sufficient evidence that the code meets the specification one nice thing is that this specification process is completely compatible with existing governance processes in the same way that people vote on laws you could vote on specifications and these shouldn't necessarily be like this is the thing to optimize for it could just be this is Baseline what we Define as the things that no AI system should violate lots of other free parameters that people should be able to choose but most alignment techniques now like rhf treat an organization or in institution or a government body or a regulator as a set unstructured set of humans and you can do alignment with a bunch of individuals but there's no coherent way in which you can say Okay I want each of these people's voices to be represented equally but if you do that through a democratic process you can run your sortition or your quadratic voting or whatever you care about to pick a specification and then incorporate that into the outcome our very long-term goal is to have something that looks like an international standards body for specifications that can help people build tools to set these um make it very easy and rewarding for any individual or company or government that's working on AI to incorporate specifications or use this architecture in order to make sure that when you say I want to build some like nanotechnology that it's not self-replicating if you're building some medicine it's not going to destroy the environment if you're building code it's not going to go over and try and take over the internet not that these things are particularly possible now but I think it's reasonable to assume they could be very possible if not likely in the near future uh Atlas Computing is nonprofit We are doing this because we think that this architecture provides the most safety if everyone can use it and I even though I find it very likely that we might set up or spin out a for-profit I want an a strong legal and governance incentive that if a company comes to us the next day that is a competitor to that company and says we want to do this too that we have to help them and this to me is what the future of programming and especially AI uh accelerated programming should look uh that is a link to these these slides specifically and I'm very reachable that's my cenly so feel free to grab 30 or 15 minutes with me [Music]