A d/acc vision for decentralized ai by Vincent Weisser | Devcon SEA

[Music] welcome to the stage Vincent thank you and um thank you everyone for coming so very excited um to talk a bit about a DC vision for decentralized AI so I think in very few Fields like the kind of like philosophical um principles of like d are more important than than fori where I think like really we're like kind of like standing at the crossroads like this is from vitalik's um blog post on this where really there's this um two cams more recently obviously some of you seen it probably on on Twitter or an X um on both kind of like the kind of like Diesels or like this the hardcore safety is who want to stop Ai and think like um it's extremely dang ous and like speak of like 99% probabilities of um us dying as well as like the extreme optimists who are basically um talking about like nor risk whatsoever and I think like it's extremely important to take like the synthesis of those views because I think extreme optimism and extreme pessimism kind of lead to inaction um and I think the more realistic view is that like there's a lot of like um positive things and I think also a lot of risk and I think actually quite importantly even the people talking about the risks they they often um obviously propose things like for example alignment and safety but then it ends up in this kind of like totalitarian almost like one world government views of like heavy compute governance like nation states uh enforcing things behind closed doors and and kind of fantasizing about like the desert project behind closed doors to build super intelligence with our Bros like that's that's kind of I think like a really the dangerous almost like proposal from like the people building the technology like the closed Frontier um AI Labs which are obviously predominantly based in the US and maybe China um so I think there's kind of like a middle ground and and I'll talk through the different four aspects of DC and we'll try to like bring them back to to AI so I think on the one side there's like the frenal progress and the idea that um progress is great but you need to make progress on on things um that make the world a safer place um kind of the principle of decentralization and of open architectures and open source um of democracy and of Defense so I'll go through um one by one so kind of like on the differential uh progress point I think it's quite important that basically like uh this point at like not all technological progress is equal and um there's some progress which basically U makes the world a safer place and it's is great to put in place um beforehand and basically um accelerate the frally like basically meaning that um there's some things which if if we solve them and a lot of the things we covered today like bios safety would be a great example um it's one of the things also the um AI people are most worried about is bios safety so if uh we would wouldn't have to worry about any pandemic if we would have like all the vaccines and all the detection um then potentially there would be less risk from AI as well um so we we kind of really need to um accelerate some of those defensive capabilities that we discussed today um to have like a safe kind of like um ground for or so um like AI progress more broadly and I think really importantly also like enhance kind of like human agency the second one um is the centralized and open source architectures I think this one is actually um extremely important and um I think the the problem a bit is that like the camp taking very short timelines and like AGI very seriously um is usually the the close Source Camp of people like the my eyes of the world and then a lot of the open source people like examplified maybe by llama and Yan leun um are kind of almost like AI pessimist I would say like they're they basically they don't really uh take the possibility seriously that we'll get like very fast progress over the next like 5 to 10 years so they also don't take the risks as seriously potentially um but I do think that like there's a lot to learn from like open source and decentralized Technologies and it's actually incredibly important I think not to um yeah like build AI in a vacuum and like single Labs um but rather I think like building them decentralized and open source makes it way easier to be safe because everyone in the world can like audit can can see the code right and and can um help improve it and make it safer and also it's like you don't have to repeat the effort right it's like right now every AGI lab is doing safety but they're doing it in a vacuum like they're doing the safety basically five times like they're not not coordinating on um sharing their research as much as they could um so I think really the the core of this is also like that the natural tendency of AI is really towards concentration right like you need a lot of compute um as also IMA touched on so I think there's basically the natural driver is towards massive centralization and and also towards kind of like nation states capturing it for their own interests um which I think is is potentially some of the like the biggest risk from it is basically is like um almost like very heavy-handed governments um enforcing kind of like uh AI governance and policy and kind of like making it very hard to for individuals to participate um so there's different elements to this um like I think there's also obviously kind of like now more and more things happening at the intersection of of crypto and Ai and decentralized AI a lot of it is still very nent so um yeah I think there's a lot of things to do I think one thing that I think is actually crucial is that I think even um like crypto and decentralized AI could contribute to things like verification of of specific safety properties for example um to actually like basically Advance um the technology safely the the other piece is like democracy and I think one interesting element there is that like a lot of the proposed solutions from the iabs is like um obviously AI governance and policy but like it it usually happens in a board of five people that can then kind of like make those decisions or behind closed doors in some small government it's not really like a global Democratic input into AI like it's really like limited to a small set of people so I think actually it's extremely important that like basically Ai and AI development is like accountable to the broader public and and takes like Democratic inputs uh not just from like specific countries or specific stakeholders but kind of like across the population and really then it preserves kind of like those Democratic Values um as well as kind of like distributes the benefits broadly right it's like I don't know how many people really trust in in in like kind of like the um the changing narratives of some of the the big closed Labs like openi that obviously promise to spread everything uh like spread the the benefits from the technology I think like this promise got weak weaker and weaker over time so I think um at this point I think there's very few people who actually trust them to spread the benefits of the technology um if they attempted and and and kind of like shareholder wealth maximization entities um the other fourth piece is defense I think basically it's quite important to build um this like protective capabil ities um because I think there there is obviously very concrete risks which are also solvable so it goes back to like the frenal progress It's like we'll get all the upside from AI like if we actually um like um build it in in the open like open source if we uh make it safe to use and and easy to use for people um so I think it's actually quite important that um yeah we also take the aspect of Defense seriously yes I think there there's like different impact areas to to highlight for me is like oh source and decentralized AI but also using actually um AI to advance science like all the areas we heard about today be it neurotech or or bios safety could be Advanced with AI and and I'll touch on briefly before wrapping up um like how that basically led me to the conclusion to work on those areas so like I um I'm building a platform called Prim which is kind of like a peer-to-peer computer and intelligence protocol really with a goal to make it like way more ible for everyone to accelerate human progress and um some of the things we did recently was for example training large Foundation mod of 10B model size like distributed across like Europe Asia and um the US um as well as like advancing some um differential like progress so the First Foundation model we built was like a metagenomic um Foundation model that basically discovers pandemics and waste water so it's like was a project with team at MIT and um it's it's a good example of like the French acceleration because like wasn't very expensive to build a model but it helps us now to potentially discover the next pandemic extremely early um say at an airport um so I think there's all of those almost like opportunities also for like open source Ai and decentralized AI to contribute differentially to the progress that like makes um like AI progress go smooth and I think the other thing is really is like there has been a bunch of like interesting work obviously on making those system safer and more aligned but I think there hasn't been enough of that kind of work in the open like in with open source so I think that's also another huge opportunity for like decentralized AI so that also concludes my talk so um the conclusion is like DC provides like a really valuable framework for AI and if you think um like exponent like pretty rapid AI progress is like out of us then it's worth taking seriously like the the huge benefits AI could potentially bring but also the huge risks and how kind of like the DC philosophy but also crypto could play a part in like contributing to kind of like a positive um outcomes for this so I think it's like yeah like incredibly important that basically um AI doesn't become like captured by like small um groups or individuals or nation states but rather that it's like um kind of like openly accessible and um with Democratic inputs to everyone um so that's kind of I think like what D can contribute to