Transcript: Michael Maximilien on Reliable AI — Interview with Alexy

So, my name is Michael Maxmillion. I go by Max, sometimes Dr. Max. But I used to work at IBM research for a pretty long time. And I just left IBM and to start my own company and of course uh NAI company. It's uh my last name.ai. So, maxmillion.ai. And uh yeah so I'm a founder you know when I when I did uh graduate school 20 years ago um one of the topic I focused on was on AI agents uh as a matter of fact my PhD work was on that uh and that was 20 years ago. So for fast forward 20 years, uh two or three years ago, um you know, of course, agents has become kind of like the the hot thing this year, but it slowly was creeping up. And uh on AI, we started at IBM a group called AI for quantum that essentially helped uh the quantum division at IBM use AI. And probably the thing that I'm most proud of is what I'm showing today at the conference. It's a it's a tool called WeCave CLI. It's an open source tool that basically standardizes uh vector databases. Uh so it allows you to to basically manage you know in development test and production your vector database and it does it in such a way that it's it's super fast and it's compatible with different vector database right now and it makes it very easy for you to sort of experiment with your vector database your embeddings the queries the results that you're getting and so on so that when you build your agents you can have the best data uh for So just talking about this weave CLI. So one of the problems that you end up with when you build agents like for instance a rag agents so retrieval augmented generation is that you may be able to query it on a collection of data and get different results. Okay. And now sometimes that's good and but sometimes most of the time you want it to be pretty precise you know. So in the cases where you want your agent to have uh at least not elucinate a lot or have uh more predictable answers then you have to work hard. So one of the things that I try to do with weave CLI is to allow you different ways for you to do your embedding. So you can experiment with that. You can experiment also on how the data is being chunked. So divide it into smaller pieces because that will impact the results that you get and you can also uh change the algorithms that uh the agent will use or at least the vector database will use to search. So when you give it a query the search you can change the the the parameters and because we're supporting different databases you can run the same test across different database. Reliable AI for me is if what you're building you want it to be more precise then you can use something like wave CLI to experiment before you go into production so that when you go in production the results are more predictable. So predictability I think is is is what reliability is for me. It's a big part of it. Well I think certainly better data better processing of that data just like I'm trying to do with weave is is important. um security is is super important. So having much more secure uh protocols. So like for instance I know MCP added security uh to make it a little bit um more you know I guess uh uh like to to catch it up with what exists in the web right now. I think that's important. I think because agents have agency by definition. I think a big part of what we need is a system for uh reputation across agents because like for instance eBay is a great example of this, right? So if you think of every seller and buyer on eBay as agents, they each have agency, they each have goals. The reason eBay works and has survived so long is because of the reputation system that they established. So basically the ranking mechanism and the fact that sellers know that if they advertise a product on eBay and they end up doing something, you know, like sending you crap, they're going to get a bad rating which will immediately impact their reputation and when somebody goes to buy the next product, they will see immediately that something happened. I feel like agents will have to get to that too. There's going to need there's going to need to be a way probably not exactly the way eBay works, but there needs to be a way for agents to have some reputation mechanism across them so that you know that when you're talking to this agent, it has a reliable uh um way of operating that it's not advertising giving you false advertisement in terms of what it can do. So I think all of this is going to be important but that's mostly for the future. So obviously for the immediate present, it's about figuring out how to get the data as kosher and as correct as possible. It's about having the basic security features that we enjoy in the web so that you can make sure that the communication is is is secure and and that your data is not being leaked and so on so forth. So those are the first steps but eventually we're going to need to get to uh reputation mechanism I think. So first thing is I would say it will be a little bit different than what we have right now. I'm actually in my company uh changing my stack. So I started with the basic Python stack that everybody uses and then start you know adding some UI on top of it. So JavaScript and stuff like that and I got rid of all that. I think Python is great for for experimentation. It's great also for maybe training but it's not good enough for production. I think for production you need better language, better system, better libraries. So I switched to Golang uh which I think is a proven uh environment for the cloud. U you know Kubernetes is built in Golang. My stack now is Golang for the back end and on the front end I use TypeScript which is essentially JavaScript but typed and that gives you a lot of advantages makes it a little bit harder for some cases but basically my stack is go TypeScript