Speaker 2
And now in 2024, we're still early on, but I think the word of the year is agents. So give us a high level overview of what's the latest on the research side as far as agents. And I guess Nester for the sake of our listeners, maybe give a kind of a very rudimentary definition of agents. Sure.
Speaker 1
Yeah. So the definition would be an autonomous or semi-autonomous agent that could or system that could operate within different environments to accomplish goals. But I think in practical terms, think of something that could book vacations for you that could do online shopping, like a tool that could schedule things for you. Like maybe you open up your Outlook one day and you say, I want to do a boxing class at 8 a.m., but one of my clients wants to meet at nine. Can you figure out a way that I could both box today and meet this client? And then the agent kind of does the scheduling for you. And that could obviously be really exciting because you could imagine a lot of productivity gains, not just in work, but kind of around the home if you have these tools that are helpful in these ways. I think what the research is suggesting is that existing models are somewhat capable, but it still seems they kind of lag behind in terms of creating these systems that are really functional and deployable in the real world. But it does seem like the AI community as a whole is much more interested in this question. One of the things that you saw in 2023 was the release of new benchmarks like Agent Bench, which was a new benchmarking suite for Agentic AI. Another one was ML Agent Bench, which was trying to test whether agents could be good computer science research assistants. And they found that on some tests, they could actually be pretty useful, but on other tests they couldn't really be. And then there was, of course, developers developments that came in the last year, like Voyager, which is a model that was, I think, based on GPT-4 that was able to play Minecraft at a very high level. Minecraft, of course, is a very complex, open-ended video game. So the fact that you had an AI system that could do well in this environment speaks really well to the kind of developments that are occurring in the space of Agentic AI. I mean, it seems to me kind of my take looking at the literature is that were maybe still not right at a moment where we have like a chat GPT for agents, like a functional agent that could help you do what you need to do. But it wouldn't surprise me if that's around the horizon because the research community really seems to be prioritizing that kind of research. And you know, the pace of AI, even surprises someone like myself that kind of works in the space, right? So who knows if there isn't somebody at an open AI or an anthropic that's cooking up the next new agent that we could all use? Or it could be, it could just be plugins, right?