E45: The AI Copilot Revolution with Div Garg of MULTI·ON

Nathan Labenz interviews Div Garg, founder of MULTI·ON, the world's first personal AI agent and life copilot. Div talks about the product strategy and roadmap for the MULTI·ON browser, their natural language approach to skills, and the steps they are taking to ensure user safety. Div explains how the platform uses a critic model to detect the success or failure of tasks, and how it can be used to book flights, order food, and more. Div also talks about the future of memory systems, such as the user profile feature, and how it can be used to improve the user experience. We're hiring across the board at Turpentine and for Erik's personal team on other projects he's incubating. He's hiring a Chief of Staff, EA, Head of Special Projects, Investment Associate, and more. For a list of JDs, check out: eriktorenberg.com. RECOMMENDED PODCAST: The HR industry is at a crossroads. What will it take to construct the next generation of incredible businesses – and where can people leaders have the most business impact? Hosts Nolan Church and Kelli Dragovich have been through it all, the highs and the lows – IPOs, layoffs, executive turnover, board meetings, culture changes, and more. With a lineup of industry vets and experts, Nolan and Kelli break down the nitty-gritty details, trade offs, and dynamics of constructing high performing companies. Through unfiltered conversations that can only happen between seasoned practitioners, Kelli and Nolan dive deep into the kind of leadership-level strategy that often happens behind closed doors. Check out the first episode with the architect of Netflix’s culture deck Patty McCord. https://link.chtbl.com/hrheretics The Cognitive Revolution is part of the Turpentine podcast network. Learn more at turpentine.co Have your AI questions answered on an episode by emailing [email protected] Send your friends our 90 min AI Scouting Report with visual aides! The AI Scouting Report Part 1: The Fundamentals is on YouTube at https://www.youtube.com/watch?v=0hvtiVQ_LqQ . Part II is coming next week. TIMESTAMPS: (00:00) Episode preview (06:42) AI agents applied to everyday life. (12:03) AI-driven automation with browser extension. (15:02) Sponsor: Omneky (18:17) AI-driven automation of web tasks. (23:57) Task automation and planning. (29:43) Automate task completion with user validation. (34:45) Lifelong learning agent with high-level skills (40:19) Guiding users to create skills safely. (46:53) AI assistant to simplify lives. (53:14) Unlock parallelism with AI agents. LINKS: MULTI·ON: https://www.multion.ai/ Div Garg; https://divyanshgarg.com/ TWITTER: @labenz (Nathan) @DivGarg9 @eriktorenberg (Erik) @cogrev_podcast SPONSORS: Shopify is the global commerce platform that helps you sell at every stage of your business. Shopify powers 10% of ALL eCommerce in the US. And Shopify's the global force behind Allbirds, Rothy's, and Brooklinen, and 1,000,000s of other entrepreneurs across 175 countries. From their all-in-one e-commerce platform, to their in-person POS system – wherever and whatever you're selling, Shopify's got you covered. With free Shopify Magic, sell more with less effort by whipping up captivating content that converts – from blog posts to product descriptions using AI. Sign up for $1/month trial period: https://shopify.com/cognitive Thank you Omneky (www.omneky.com) for sponsoring The Cognitive Revolution. Omneky is an omnichannel creative generation platform that lets you launch hundreds of thousands of ad iterations that actually work, customized across all platforms, with a click of a button. Omneky combines generative AI and real-time advertising data. Mention "Cog Rev" for 10% off. Music: GoogleLM

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Div Garg?

  We want to unlock parallelism for humanity where like currently you can imagine like we are all single threaded. We can do only one task at a time. But if we could have AI, we could launch like maybe like, I don't know, like 100 AI agents at a time. All the AI agents to do the work and then tell you like, oh yeah, this work is done. And then you can like sort of like coordinate that. In the core, I believe like you need to have some sort of tenants if you want to build really intelligent AI agents. And one of them should be like, an agent should not be able to modify its own source code. Because if it's able to modify its source code, then it could like self evolve, it could do like really weird things. I don't think GPT-4 is there yet when it comes to a lot of really complex reasoning. And so you might need some like breakthroughs in the foundation model layer. You know, we, you have like 20 foundation models dropping and like, I don't know, 100 papers published.

• 0:44Div Garg?

  So it's definitely the progress is crazy.

• 0:46Nathan Labenz?

  Hello and welcome to the cognitive revolution where we interview visionary researchers, entrepreneurs and builders working on the frontier of artificial intelligence. Each week we'll explore their revolutionary ideas and together we'll build a picture of how AI technology will transform work, life and society in the coming years. I'm Nathan Labenz, joined by my co-host Erik Torenberg. Hello and welcome back to the cognitive revolution. My guest today is Div Garg, founder of Multi-ON, the world's first personal AI agent and life co-pilot. Until recently, Div was a PhD student and adjunct faculty at Stanford where he created and taught a class called Transformers United,

• 1:28Nathan Labenz?

  which explored how the revolutionary transformer architecture is already beginning to unite such previously disparate fields as natural language processing, computer vision, biology and much more. Now, because he believes that AI agents will become one of the most important parts of the AI technology wave, Div has taken leave of the PhD program, raised venture capital and is currently building the Multi-ON team so that people no longer have to browse the web alone. This is our third show about AI agents, and I'm really struck by how different the product and rollout strategies seem to be. Flo Crivello, the CEO of Linde, often posts images of intricate workflow automations that Linde has created for him based on detailed prompts.

• 2:13Nathan Labenz?

  Matt Welch, the CEO of Fixi, has launched a public-facing agent platform as a playground for developers, but is really targeting enterprise customers. And Div, with Multion, has wrapped an AI agent up in a Chrome extension so that users can easily delegate tasks to it, can watch what it's doing and supervise along the way, and can even assist in real time as necessary.

• 2:36Nathan Labenz?

  The AI agent space is fascinating, and in my view, it really does merit a diversity of approaches. Each strategy, of course, has its own strengths and weaknesses, and time will tell which one ultimately wins out. But after testing a bit over the last couple of weeks, I can say that Multion is really fun to use. Watching it navigate the web, explain its reasoning, and attempt to move forward toward whatever goal you gave it really does feel like a glimpse of the AI-powered future. In this conversation, we talk about Multion's product strategy and roadmap, their natural language approach to skills for Multion, Div's vision for how AI agents will impact human work, the steps that they're taking to ensure user safety, the standards of reliability

• 3:20Nathan Labenz?

  that Div believes Multion will need to achieve to see mass adoption, and lots more along the way. I really loved Div's tremendous energy and positivity, and I found this conversation extremely interesting. If you agree, as always, we'd appreciate it if you'd give us a review on Apple or Spotify, leave a comment on YouTube, or simply share the show with friends. If you have any feedback or suggestions, you can reach us by email at [email protected] or DM me on Twitter, where I am @LeBenz. Now, I hope you enjoy this conversation with Div Garg of MultiOn. Div Garg, welcome to the cognitive revolution.

• 4:00Div Garg?

  Yeah, happy to be here.

• 4:01Nathan Labenz?

  I'm excited to talk to you. So you are the founder of the AGI Company Inc., also known as MultiOn, which is online at, for limited users for now, at multi on.ai. And, you know, from what I understand, you have left the comfortable environment of a Stanford PhD in machine learning and thrown yourself into the incredibly turbulent waters of trying to build an AI agent startup in the midst of, you know, one of the craziest technology environments ever. So to get started, tell us a little bit about that.

• 4:42Div Garg?

  So definitely, I would say like this seems to be the right time to go and like start building AI applications and agents. And I think the best startups are born during like turbulence, almost like you have a very big technology revolution happening. There's a lot of turbulence. A lot of players are getting like shift up shifted. And I would say like this is still early enough where you we want to go and like sort of start building AI applications that can be useful in everyday life. So I will say like, I still think like agents are still early, but this is the right time to like go start building them as a startup and start solving the problems there. And even if it takes you like, say like five years or like more, I think this is the right time to like start building and like tinkering with them and solving them as like a problem.

• 5:25Nathan Labenz?

  Yeah, definitely. If you wait until it's easy, then it's definitely too late to, you know, be in on the the ground floor of the value creation. So that definitely makes sense. I want to get into, you know, a lot of detail about how it works and all that kind of stuff. But what's up with, how did you come up with the incorporated name of the AGI Company Inc.? And where does the name MultiOn come from?

• 5:47Div Garg?

  Sure, sure. So that's a good question. For the AJ company, it was just like, it was really funny. So I was checking online, is this a name that's available? And I was actually very surprised, like this is a name available. So I went and registered. And so we're not actively using that name for now. We're also, we're just like using multi on. We maybe we will like start using it in the future. Maybe we can spin up a nonprofit or something.

• 6:14Nathan Labenz?

  The nonprofit controlled startup is all the rage these days. So what about multi on? Where does that come from as a name?

• 6:21Div Garg?

  Yes, yes, yes. So I would say, um, yeah. So this is like, I would say, I have a physics background. I did some like international physics olympics back in the day. And the name is actually inspired by like a hypothetical, like quantum particle that's present at every place at the same time. So it's like, if you have like some like neutron, muon, fermion, it's like multi on.

• 6:42Nathan Labenz?

  Fascinating. Okay. I blind spot for me. I didn't even make the connection. For me, it meant just like something that was on in multiple ways, I guess. I don't know. I was kind of thinking of it as my, you know, my AI assistant that's always on or something like that.

• 6:57Div Garg?

  Yeah, I think people think of it like that, too.

• 6:59Nathan Labenz?

  Tell me about your, your kind of approach in, how do you see the, the relationship right now between research and product building? It's, you know, the, the pace of research is so insane that it's like, as a product builder, you know, feels very hard to keep up. And yet, if you don't, you know, that seems like it by definition, like a huge problem. So it does seem like a good time for researchers to start companies. How are you thinking about that for you?

• 7:28Div Garg?

  So definitely, I think, yeah, this is a good time because you can finally say like AI is starting to get applied in everyday life. So if you look one year back, like there was no GPT four, there was no stable division, a bunch of those things. And like, people were not actually using AI as a product. Like maybe they were like seven cars, but didn't work. Uh, there were no large language models used in production. And I think this has shifted a lot, just like has been like exponential shift. And this, like I would say, it's a right time where AI is starting to get applied as product. And the field is moving very fast. So it's definitely hard to consolidate the research with the product because all the iteration cycles are really fast. And things like even like in a week, you have like twenty foundation models dropping and like, like, I don't know, a hundred papers published.

• 8:10Div Garg?

  So it's definitely the progress is crazy. Um, for us, I think like the nice thing is, uh, We really like the open source community and like we see that as like a way where we can like bootstrap on top of developments there as well as like do collaborations to sort of like like take the best approaches and like pull in to what we are doing as an agent to company.

• 8:34Nathan Labenz?

  What's the status of the company right now? I noticed that you have put out the call for hiring, but I haven't seen any fundraising announcements. So is, would you like to break any news here on the cognitive revolution?

• 8:47Div Garg?

  Sure, sure. So obviously we are in the middle of a seed race. We have got some really good term sheets. And obviously you'll see an announcement in like maybe a few weeks. So can't disclose yet.

• 8:59Nathan Labenz?

  Cool. We'll keep an eye out for that, but I would not doubt that you would have any trouble raising a few million bucks in a seed round at least. So wash this face, everybody. So what roles are you looking for? What is kind of your, I assume it's multiple things and obviously it's like people who are obsessed with AI, but anything in particular that you are in need of right now?

• 9:25Div Garg?

  I'm sure. So I would say like in general, we are focused on hiring journalists at a current stage. And I think like focusing on hiring people who are really passionate about AI and agents. And they can like basically define the culture of the company as we grow and like build it up. And I think we are looking for everyone from like hardcore full stack folks to backend engineers to AI specialist, even like research engineers. And I think we want to solve like agents as like both of AI research problem as well as a product problem. Okay, like what's the right UI UX? How do we build this into a right product that everyone can use, make it reliable? And they're also like really useful and interactive.

• 10:02Nathan Labenz?

  Ian, I'm having flashbacks to our, my conversation with Arvind from Perplexity, who has a pretty similar background in multiple ways, including like having spent, you know, serious time in research and also just the fact that, you know, they have set a very high bar for how fast they ship. And I see you guys also shipping very quickly, you know, very short time between iterations. So that's a high compliment, actually, in my mind.

• 10:30Nathan Labenz?

  Where do you think we are right now in this agent development? Do you think that the product is already useful or still just kind of a hobbyist thing? What is the path to actual like agent-based utility?

• 10:46Div Garg?

  Definitely. So I think that's a good question. So I would say for most agents, you will see like, say, AutoGPT or BBHR, even like a lot of like coding like agents, they're very early, like you can use them for very simple applications. For multi, I will say we are like a bit further higher in terms of like you have actual applicability currently where you can ask it to say like online shop for you, order your dinner, like book flights, so on. I think like the biggest things we're trying to solve is like, how do we make it really reliable? And how do we reduce the variance of what the agent does? And I think one issue we have seen is like, even if you ask it to do like one, something like 10 times, it might just have slightly unpredictable behavior. And so it's very hard to make a deterministic for the end user. Okay, what's going to happen?

• 11:28Div Garg?

  And I think if we can solve that, then we can like really give it to everyone and like make it really useful. And I think like a lot of the problems are like making it really reliable and making it very predictable. And those are the problems we're trying to focus on currently too.

• 11:42Nathan Labenz?

  So let's get into then a little bit how it works. The first thing that seems like a actually huge decision point is when you're creating an AI agent product. And I've, you know, sampled, demoed at least like a number of them at this point is kind of at what level does this thing operate in the first place? We've seen, you know, a couple of different paradigms designed there where, you know, sometimes it's like it works all in the cloud. And you kind of, you know, define some little agent to do some little thing in some platform. That's, you know, a probably not super flattering description of some of the things I've tried, but you know, that's. It's it's executing over there somewhere, you know, far from you.

• 12:27Nathan Labenz?

  Then there's like the OS level co-pilot concept. And then there's what you've chosen, which is the browser extension ride along kind of co-pilot mode. How did you choose that? What do you see as the big pros and cons? I have some thoughts, but I want to hear yours.

• 12:48Div Garg?

  Sure, definitely. So then the biggest like one choice you made this is to make it like really seamless in terms of like user experience and also like optimism distribution. And so currently, like if you go to the Chrome app store, it's like a one click install. Once you install like that's it, that's all you need. And if you want to sign up and make an account first, you can start using Maltal in your current browser. It has access to your like current logins, like you can use it on any website you are. You have authorized it to. So that makes it like seamless and very easy to get started. And that's one of the reasons we chose to make the choice. I would say another advantage is if something goes wrong or the agent is confused, it can simply ask the user for help. So it's like, oh, it can ask like, say, is it like, can you log into this website? Can you solve this captcha for me?

• 13:31Div Garg?

  For example, and that gives like a lot of like very like immediate interaction to the user where like the agent can correct it. And so the user can correct the agent and we can use that as like a feedback loop, which we can like find in the agent on to improve it for future interactions.

• 13:50Nathan Labenz?

  I think, you know, to now share my perspective, I think this is a pretty smart approach. I could see an OS level approach, you know, kind of having some similar advantages. But the fact that I'm logged into everything in my browser and that you kind of get that for free by adding on a browser extension that acts in the same space as I do as a human user. That even if it's not like the final form of this long term, definitely feels like a really good place to get started. And in my testing of it, I've I've kind of experienced that where it's like. Oh my God, you know, this morning, for example, I was on the, on the phone with a teammate.

• 14:34Nathan Labenz?

  I wanted to add him to a GitHub repository. So I just tried it with, I'm gonna, I'm gonna say it, start saying how you say it, multi on. And, you know, so I pull up the new tab and I just say like, go, I asked him, what's your GitHub username? And I said, okay, go to this repo and add this user. And right off the bat, it's like, other paradigms, you know, I'm immediately stuck with authentication.

• 14:59Speaker 02

  Hey, we'll continue our interview in a moment after a word from our sponsors.

• 15:02Nathan Labenz?

  And, you know, anybody who has, you know, had to go authorize some, you know, temporary, whatever, special purpose key out of GitHub, you know, I'm immediately like, oh God, I have to go do that. You know, and I have to do that for everything I want to connect or, oh yeah, this, you know, that really bogs down. I'm, I'm kind of always say, access is always the hardest part. So that's really tough. But with this, it, it just allows, you know, the AI to kind of ride in the same space as me. I'm now watching it, it navigates to my thing. It, it, you know, was able to find going to the team page. It was able to search, you know, find, go, pull up the add a collaborator.

• 15:43Nathan Labenz?

  It was able to search for the guy's username and find him. And then I did experience also what you were saying around kind of my ability to intervene and like help it if I need to, which obviously is like not the long-term product vision. But there was a step where it was kind of getting confused where it had like found the guy in the search results, but it needed to click before it could actually click invite. And it wasn't quite, you know, clicking the click to get to the invite click to be accessible. And so it was kind of stuck, but then I just clicked the thing. And, you know, I think this kind of gets to my next question of like, the model seems to be sort of a... You know, just kind of a, a simple loop, right? Where I injected my little action and changed the state space.

• 16:27Nathan Labenz?

  And then as it kind of took the next read of the space, it was like, oh, now I'm able to proceed. Because, you know, something kind of just got easier for me or whatever. Something happened good and now I can, you know, my barrier has been removed and it kind of just continued. Uh, so I, I, I think that paradigm is super, super interesting. One question I do have, does it, does it have the ability to interact with other extensions? Like I have LastPass. How does that relate? I haven't quite figured that part out yet.

• 16:56Div Garg?

  Oh, yeah, I'll definitely say like, uh, it's also funny, like for that use case, I've actually used it for similar like use cases, like adding like people to get help. Uh, I've also used it for like sending like, uh, like DocuSign stuff is actually pretty good. Like instead of like send this like whatever contract to someone can actually go and do that whole thing. And, and also the ability to like interject, I think that's great. And then we can use that as like fine tuning samples to improve the ability down the road. So we're actually building this like, uh, like a recorder mode where like you can like sort of like teach it, like give it a demo like, oh, maybe this is how you should do it. And if you just show some query, maybe one or two times, then it will know exactly what's the best way to do it. And I was slightly expanding on this, like one thing I've noticed is like with any of agent, it's doing something for the first time. It might take a longer roundabout path.

• 17:39Div Garg?

  But if we can like sort of like give it some sort of like feedback or demonstration, it can like learn the shortest path. And I think that helps a lot. And that's sort of the approach we are trying to take to improve our performance over time. And to answer the second one, yeah, it's like definitely like, uh, like ability to like interact with other extensions. It's like, uh, plus we are trying to explore. So for example, like for passwords, like one password has an API. So we are building an integration with the one password API and also other extensions. And so this is a work in progress where we can interact with a lot of existing, um, browser solutions to simplify the flows. Currently, Mertion can also use Honey, for example. So if you're on Amazon and you want to check out, and if you see like, there's a Honey coupon, you can apply. And the agent can go and actually just apply that coupon for you and like get you a discount.

• 18:24Nathan Labenz?

  Cool, interesting. So the, the core architecture, if I, I'm, I'm kind of guessing and you can, you know, tell me, uh, where I'm right or wrong and, and share as much additional detail as you want to. But in kind of keeping with the paradigm of like. Working from a sort of research foundation into product. It seems like the kind of initial core of it is essentially a react paradigm. Where you're looping through and kind of each cycle you are taking kind of a snapshot of like. First of all, what goal have I been given? What is the current state of affairs that I'm looking at? There's this kind of interesting and we can, you know, share some videos to accompany this.

• 19:11Nathan Labenz?

  Kind of little chat that isn't so much, it's kind of a chat you can chat with it if you need to along the way. But a lot of times it's just telling you what it's thinking and what it's going to do. And then there's the actual action. So that's kind of the the baseline.

• 19:26Nathan Labenz?

  React, you know, research paper. It seems like then now you're you're starting to, I'm interested to hear if there's any, you know, major deviations from that. But I know like the next big thing too is then layering on the skills component to that.

• 19:39Div Garg?

  Definitely. Yeah, so I think, yeah, on high level, that's very correct. That's what you're doing. I would say like we have like three parts of the system. So one is we have like our representations of models. Maybe like take the like the DOM and like make it into like a compact text and mapping. Which we can give it like a like a planning model, which is like a G P T four or some of the similar L. That can break down the task and find like what are the right actions to take for this particular task. And and then we can like do something similar to a react paradigm where we have an inner thought and the agent can decide based on the objective. Okay, what's the right thing to do? And then like similar to the right actions. We also have our own intermediate action grammar.

• 20:23Div Garg?

  And that action grammar helps a lot where instead of like directly generating JavaScript code, we like generate like commands in our own action like language. And this has the ability where we can like type check them, we can verify everything, say, there's no like malicious activity, there's no prompt injection, for instance. And after we have verified that, we can compile that from an action grammar to actual DOM level events. For the for the website. And we have found that this helps generalization a lot compared to just like code. And for the agent, I will say we use some very interesting tricks. So if you have something like AutoGPT or BabyAJ, for example. The biggest issue they have is like they have this thing called like almost like task divergence. Where if you give them an original objective, they will like diverge away from it and don't know how to like.

• 21:07Div Garg?

  Error correct back, how to come back to the original thing once they have made a mistake. And for us, I think like the way we have formulated the problem is we tell the agent to always try to. Reach like come as close to the original objective as possible. And we basically give this like in a problem like never give up. Uh like always try to go as close to possible. So I think also like, like yes, like the never give up is like something that helps a lot.

• 21:34Nathan Labenz?

  Yeah, it's the the psychology component of of like the positive self-talk or the uh. The need for grit, uh grit is me a language model grit. Uh is kind of what you're trying to to coax out of it. So it sounds like there are. A couple of. Expanding on this kind of core react concept. There it sounds like our specialist. Modules that wasn't clear to me if these are like. Language model powered or just explicit code. Cause you could kind of imagine doing either, right? If you have a DOM, Lord knows like DOMs are. Terrible, you know, ginormous messes these days.

• 22:15Nathan Labenz?

  Uh, you know, long gone are the days of like highly semantic uh HTML, unfortunately. No such luck, you know, for us these days. So. You could imagine just writing kind of classic code to try to compress that. You could also imagine having a like specialist. Let either, you know, specialist. Three point five turbo or Claude instant prompt or even a fully, you know, fine tuned specialist model to help. Understand the HTML. What do you? What do you find there is kind of working or how do you see that evolving?

• 22:49Div Garg?

  Yeah, so I think it's very interesting. Um. So for the representation side, like I'll say. We have three modules, so like a representation module, like a planner module and like the extraction module. And initially, we started everything except the planner was like, uh, maybe just like, uh, low level, like, like, like more like heuristic code. And now we're transitioning everything to a model. So on the representation side, um, we like can take a combination of the screenshot plus like, uh, like the DOM. And we can have embeddings from the DOM and we can have embeddings from the image. And then we can feed it to like, uh, like a representation model that can combine them together. And this is similar to maybe like what some like instruct blip does, um, get like a combined presentation. Uh, which could be, in this case, we have like made a choice of making it like text space. So it's easy to analyze. It's easy to like see what's, what's happening.

• 23:34Div Garg?

  And then we can give that as a prompt to the planner model, like this is like the state representation. And this is a user query, what could, what, uh, should we do next? And once the planner can make like, okay, this other sub, uh, like task, we need to accomplish, we can like feed that to, uh, the action layer.

• 23:51Nathan Labenz?

  Yeah, okay. Very interesting. So is it, um, you've got kind of a joint, a joint vision text representation? And it's a, we'd had the blip authors actually on the, on the show in an early episode, but it was before instruct blip. But I'm familiar with that as well. That ultimately cashes out though to a text output, right? In the end of kind of the representation, it's, it's a semantic representation where you would be clicking on like. Oh, I mean, I guess then when the action module is going to act, it's acting on a semantic action space. So it's like saying click this button, not click at, you know, coordinates one hundred, two hundred.

• 24:34Div Garg?

  Yeah. So I'll say it's a still a semantic representation. Um, and we can give like IDs to each, um, element. And then like predict which ID to choose. And that helps a lot. Uh, we can also, um, we haven't seen the need for actual coordinates. Um, but that could also be a choice if you want to move the cursor to a exact coordinate. Um, you could also potentially predict the coordinate to move it to and then make a choice there.

• 25:00Nathan Labenz?

  Yeah, it's all kind of on the table. Uh, once, once the, the. Modal barriers are broken down, you know, everything kind of starts to become possible. GPT four, you kind of said, you know, or similar for the central planning model. Does that in practice mean like GPT four today? Or is there any, in my experience, like for these kinds of executive planning, long range planning type tasks? GPT four is kind of in a league of its own, is that what you found also?

• 25:31Div Garg?

  Yeah, when it comes to like long term reasoning and like planning capabilities, we have found GPT four to be the best. Also experimented like, like, like open source models, you know, like cloud hundred K. And what we're in the process is like, we have collected a lot of like interaction data. So we already have like fifty thousand like web interaction samples. And we are like starting to find you in our own like model. So we can find you in a Falcon forty B model and we're starting to do that. And pretty soon we could potentially transition to all in-house models. And the nice thing here is like the approach we're taking is where we are launching the product, getting data, improving the product over time, using the data. And then, and then I saw like collecting more data and improving it.

• 26:13Div Garg?

  And so we're seeing like this is something that can give you a feedback loop where you can collect like the data that matters the most to the user and improve the model there. And I continuously like build this like loop to like continuously get improvements. And like also like almost no one is currently doing that. So this gives us like an interesting advantage in this space.

• 26:30Nathan Labenz?

  Well, it's a little tricky to necessarily define, right? Because one, in my red teaming with GPT four back in the fall of last year, one of the things I really tried to explore and certainly not as deeply as you have, but is this kind of self delegation, you know, concept. Can GPT four like break problems down? I used a more kind of recursive structure. I guess inspired by, you know, some Doomer thought of recursive self improvement. I kind of went toward recursive delegation, which I actually don't think now is like really probably the right paradigm to build a product around. But, you know, I was just kind of doing my own messing around. I found that one really hard thing, though, was how do I actually know if I've succeeded?

• 27:15Nathan Labenz?

  You know, for a given task. And I was even just doing, you know, some pretty basic tasks like go online, find some information. And then I tried to create my sort of validation layer on top. But I found that for many, many things, it was pretty hard to determine if I'd actually succeeded or not. So are you detecting success or failure like automatically? Or is the, you know, are you to some degree relying on the user to kind of tell you when it worked or didn't work?

• 27:41Div Garg?

  Also, in our case, we are relying, it's like automatic. So the model is able to predict whether it has fulfilled the task or not. And we also are experimenting with a separate critic module where the critic could say like, oh, like, is this task complete for this particular objective, given the current set of actions? And I think even for Voyager, I think they took a similar approach. And I've seen this to be pretty reliable in practice.

• 28:10Nathan Labenz?

  So that seems like, yeah, some things will lend itself better to that than others. One of your, you know, videos that I've seen was the world's first AI booked flight. And I guess for one thing, you could kind of determine like, is a flight booked or not? You know, if you ever get to a success message that has like a confirmation code, then, you know, it's pretty safe to say you've booked something. You could still have some questions around like, did I actually book what the user wanted me to book?

• 28:39Nathan Labenz?

  And there's, you know, varying degrees of rightness on that, too. Like, did I take them to the right city? You know, is it the right, uh, did I get the right flight given the options? Did I book the right seat? I mean, you know, did I get their traveler number in there? You have this feature too of like your profile, which I haven't explored in depth yet. But, you know, I also, I'm interested in your thoughts on kind of the future of memory. But right now, you're kind of a big text box, put your profile here, say, you know, tell us who you are. And then presumably the, you know, future of that is like more elaborate, you know, kind of memory systems on a per user basis as well. But, so I tried something else. I'm throwing a lot at you, you can kind of unpack this at length. But I tried something the other day where I was like, go get me my three, find the URLs for my three most liked LinkedIn posts.

• 29:25Nathan Labenz?

  And it got as far as my feed.

• 29:29Nathan Labenz?

  And so it was like doing okay, but then it never actually gave me back like here are three URLs. And I wasn't quite sure like, did it already think it had succeeded by kind of getting there? Or I don't know exactly what happened, but I'm not sure how you would even determine if you were successful. You know, here are three URLs, but are they, you know, the most liked URLs? How do you think about that? Kind of, you know, it's tough to ground some of these, these tasks, right?

• 29:54Div Garg?

  Yeah, I think that's, that is, yeah, I think that's a good, good point because like suppose even if you complete a task. It's possible like you completed the task like wrong. So maybe you booked a wrong flight, for example, or ordered something wrong. And I think that's also a consequence of failure from our standpoint. It's, it's hard to detect because like the agent might just think like it's doing the right thing. And it successfully completed the task. And in this case, I think we've found like if you have a separate agent that is just like a critic agent or a valuation agent. That can help a lot. And at each point it can verify, oh, like, is this the right thing? Corresponds to the right thing, what the user's original like query or objective was. And we can make a feedback loop there where the critic agent can like criticize any decision.

• 30:38Div Garg?

  And give it as a correction to the, to sort of like the agent actor, actor model. So this is something we have played with. Second thing we can also do is we can have a user sort of like, like acknowledgement flow. Where like it can put something in your cart or like maybe like go all the way to the checkout for a flight. But before pressing the buy button, it will like ask the user like, okay, like I found this flight, you can send your screenshot. Do you want me to like choose this file? Do you want a different one? And if the user says yes, then you can go and proceed to actually complete the whole thing. And this is also one way where we can like make it safe and make sure like you don't accidentally make up, end up making a wrong purchase, for example.

• 31:18Nathan Labenz?

  Yeah, that makes a lot of sense. How do you envision that kind of? Maturing, you know, in the obviously like early kind of demo era of the product. I largely just watch it, right? So I I kind of half the fun is doing the, you know, riding along and seeing what it does and what it gets right and doesn't and, you know, helping it get over a little humps with my, you know, extra in injected click and you're not recording that now, but there is the kind of a mode that you're developing where you you will start to record those manual interventions.

• 31:54Div Garg?

  Definitely. Yeah. So I think one is like, just like recording interventions. And we can use that as like, like the next time it it would like to just learn it. And and like will like remember like the interventions and what sort of corrections you get and improve its behavior automatically. And we can do like in context learning for this, we can also do fine tuning of the model. Where once we collect a lot of like this sort of interventions from like a lot of our users, we can like find in the model to improve the performance where it's currently not working.

• 32:26Nathan Labenz?

  Is the vision sort of to have a like I set something off on it on a task? And then I kind of go do my own thing. And then maybe I get like a notification, you know, three minutes later that's like

• 32:42Nathan Labenz?

  multi on is ready to have you review, you know, before purchase or whatever?

• 32:47Div Garg?

  Yes. So I think, yeah, I would say we have a lot of things planned on the pipeline, but in a high level, you can think like, suppose you want to make it say order a burger near you, for example. And so maybe like the first time you order it, you might want to watch and see what's doing. Maybe if it does something, you know, can intervene, give it a correction. And then you can maybe like play around with like two, three times, have some fun, like sort of like guiding it, interacting with it. And once you know it's like reliable enough, then you can just like leave it in the background. And when the agent is done, it will send you a notification. Okay, I'm done. It could also be like a mobile notification. You can be anywhere. Like my time will be run in the cloud. So you can be like, like, just like be anywhere. You can say like, oh, like, get me a burger.

• 33:28Div Garg?

  It will run on the cloud, do the task for you and send you a notification, which could be like, okay, like I found this thing. Potentially along with the screenshot and then you can like acknowledge whether to like finish the task or to actually get the like say the burger in this case. Or you want it to like cancel it or do something else.

• 33:46Nathan Labenz?

  How does that cloud mode relate to? Off, because I'm imagining like a headless browser in the cloud or whatever, where I'm not logged in to, you know, all my accounts.

• 33:56Div Garg?

  Yes, uh, that's an interesting question. I would say we are working on a lot of things, uh, but I, I feel like we have cracked this problem. And there it's too early to like, say like, uh, like how exactly we're doing it, but I think we have a really like a nice and elegant way to like, sort of like solve the art issue.

• 34:14Nathan Labenz?

  Cool. Well, that's a, that's a genuine breakthrough. Uh, if you could make that work smoothly. It is a huge stumbling point for so many of these experiences right now. So then tell me about the memory side. So you've got. Uh, I understand. Okay. We've got, for one thing, my kind of profile, which I can currently like just manually manage and, you know, tell it, you know, my name's Nathan and my email is this and my, you know, I prefer window seat, whatever. And then. You also have kind of seemingly system level memory that is like the skills. And again, this is kind of, you know, close follow as I understand it on research progress where.

• 34:54Nathan Labenz?

  Was it even a month ago yet that, um, I think this is an Nvidia project, right? The Voyager project of, I, I refer to this as the L L M's as lifelong learners. Where they're going out and doing stuff in this, in the Voyager case, it's Minecraft, figuring out how to do stuff successfully and then storing these successful subroutines. Your skills though, as I've seen them are like. Way higher level more to like natural language than the Voyager skills, which were like very granular kind of Minecraft execution code. From what I've seen of your skills, it's like a sort of note card length, highly. You know, abstracted away from all that detail, kind of summary of like. What it is you're trying to do and roughly how to do it. So.

• 35:40Nathan Labenz?

  Help me understand that a little bit better. I'm a little, I guess my surprise there is I, I wouldn't expect that it would be the high level thing that the, that the agent would need as much, you know, I would, when I talk to G P T four and I say, how do you go book a flight? It's kind of like, well, you go to kayak and you search for stuff and it seems like it kind of knows the general direction, but then it really gets bogged down in low level details. So I guess I'm a little surprised that the skills are so high level, or maybe I'm missing something. So enlighten me.

• 36:08Div Garg?

  Sure, sure, sure. Uh, like, I will start with the memory part first. Um, so the memory, I think the current feature we have is like, it's like a scratch pen. And so this is almost like having, like extending notes with your agent where you can like say like, oh, this are my personal details. This is what, this is what I like, this is what I don't like. This is my zoom link. You can give it instructions like, okay, like always like add zoom links to my calendar invites, for example. Or like, don't do this thing, whatever. And it can like remember that as like M bearings and like a, a conference preferences, like, so this is like more like explicitly you give like some preferences to the model. And we're trying to make this like really secure. So like you can install your credit cards there if you want to install passwords there. And we guarantee everything is client side, it's private and secure. Um, so that's built like a, like a secure way to like give information uh to the agent.

• 36:52Div Garg?

  So there's a second layer we're building is I would say more like in, um, interactive memory. Where like if when you interact with the agents, it will like actively learn preferences over time. So like the first time you say, say order a burger near me, it might ask you like, okay, what's your address? And once you tell it your address, it can remember that and use that uh as like a long term memory in the future. Um, so if we say something like next, like, okay, find me a, like a haircut place nearby. It can like, okay, it now knows your location and can customize. And so over time, like the more you use it, the more it will learn about your preferences. And so this is a layer we are currently building. Uh, which we call it, which we are calling as a personalization uh to the end user. And I think that's a lot, very interesting. Um, and like, there's a lot of things you can do there. Um, and coming to the second point around skills.

• 37:37Div Garg?

  Yeah, I think, yeah, it's interesting. The choice you made on skills is to keep them high level. Like, how we're thinking about skills is almost, like, you can think what the agent currently does is, like, it goes from natural language instructions to each single line of code, uh, or like actions in a sense. So it's like, uh, for each natural language, pretty say, okay, like, okay, like, I take this action first, then second action, third action, whatever. So it's basically like each line of that, uh, action code. And what we want to do is instead of predicting like each line of the actions, we want to predict like high level action functions in a sense. Um, and this can be like, sort of like, almost like skills, where we say, um, maybe like send, uh, maybe like follow someone on Twitter, for example. So you could have some sort of like, say, like a function that's defined, that's like follow, uh, on Twitter can take an input which could be user ID.

• 38:22Div Garg?

  And it already knows, okay, what are the actions that are needed for this function, uh, that, and, uh, and then, and in that case, like, instead of like predicting each line of, uh, like, action code, you very basically like choose like the action functions and like, uh, combine them together. Uh, so that's how we're sort of thinking about skills. And the current skills are, I would say, a combination of scripting and natural language. So they're, they're very hello, and the choice we have made because of this is, like, anyone can go and develop their own skills if they want to. So if you want to develop your own customized workflow, or you want to give it like some sort of personality where you want to act, make it act like a recruiter on LinkedIn, or you want it to like, say, like, act like a, uh, socially, like, like a social influencer, for instance, you can make a skill around that.

• 39:04Div Garg?

  Uh, all in natural language and like feed it to the, uh, agent. And we're trying to develop a whole ecosystem around it where we can have a directory where people can go and afford skills, share skills with each other. And like the agent can pull in the best skill automatically for each here website. And I think this can allow, like, the agent to improve really fast and people to also, like, contribute, uh, to the agent.

• 39:28Div Garg?

  And so, yeah, so I would say, in a sense, uh, yeah, that's what we're trying to do. And we have seen like high level instructions actually work really well. It's almost like if you are a teacher and you're guiding the student, like the student makes a mistake, you can tell them, oh, you made this mistake. Or maybe like, this is something you can do better. Uh, maybe you should like click here to open this dropdown instead of typing on the dropdown. And if you can like give that instruction, um, that can help, uh, uh, improve the behavior of the agent.

• 39:56Nathan Labenz?

  It it's almost like the skills are kind of additional, kind of conceptual guidance on things where it didn't immediately work. But you're trying to keep the guidance high level enough that users can write their own

• 40:14Nathan Labenz?

  without having to get down to the like level of your action language, which I guess is a you know, and saw in some of those, maybe something you don't necessarily want to share anyway. Um, if that's a, you know, a key part of, you know, what makes the whole thing go. So, so that makes sense. So there's not a skill is does not is not stored as like actual procedure at all. That's all still generated dynamically at runtime each time.

• 40:41Div Garg?

  Uh, currently, yes, but what can also imagine just to add to this is we could like, if we have the natural language description of a skill, then we can have a model like generate the actual like action code for that particular skill uh, at the runtime and then we can also cache it. So you don't have to next actually, uh, like if you generate the action code each time, you can just like pre-fatch whatever is already generated and use it. But for a user, they just need to define the natural language interaction and another layer can go and compile the actual code.

• 41:07Nathan Labenz?

  Presumably though, you you also have to have some input of like current state for most of these skills, right? Like to generate the actual action code, you're going to want to be responsive to runtime conditions too. How do you think about, um, drift? You know, there there's just obviously, you know, Twitter, uh, announces and launches updates on a unpredictable schedule these days. Uh, so just to take that one as a, uh, you know, a sample. I think these things are often kind of forgotten and enough states forgotten by you, but like, you know, weekend hackathon type projects age really poorly when, you know, all of a sudden it's like, yeah, but somebody made a tweet to their website and now your whole thing doesn't work anymore.

• 41:49Nathan Labenz?

  How do you think about kind of keeping up with the ongoing? Evolution of things, or you know, you've got a skill, but even at a semantic level, that skill could be out of date if they change, you know, the flow. So yeah, how do you think about that challenge?

• 42:03Div Garg?

  Yeah, so I think that's the beauty of keeping everything as a natural language, high level instruction. So when it's like, if that high level natural language instruction is still correct, then we can dynamically recompile the skill to get sort of the new action code. And even if like something changes for the website, it's very easy for anyone like basically like who doesn't even have a technical background. We just go and like change the the high level language into like what is like now correct. And then the then we can like compile, recompile that to the actual action code. So I think it just like removes the technical barrier around actually understanding a low level, what like what are things that might be needed? And anyone, any sort of user can go and like sort of like play around and like build things on top.

• 42:48Nathan Labenz?

  How do you think about? So that also then makes me think, how do you think about malicious skills? You know, if I if I imagine this kind of rolling out to a lot of users, I'm interested also in your timelines on sort of. You know, at what point do you think multi on goes wide because, you know, it's actually useful to non hobbyist people? I've been saying for a couple of months now, six months. So basically from G P T for launch and the initial agent craze, I was like, give them six months, you know, that'll probably bring us do around like. October, I'm interested to hear how you kind of think of when those things would crystallize. But then as that also comes online and goes to scale.

• 43:30Nathan Labenz?

  Now, you know, if I if I'm aware of the the fact that like I could go in and semantically modify some skills. You know, couldn't I just go in and modify a skill to be like, by the way, in the middle here, like Venmo all your money to my, you know, to my Venmo handle? And if you're already logged into Venmo, you know, those things are hard to. To undo, right? Like Venmo probably doesn't have an established protocol for my agent did it and I didn't mean to, I didn't mean to set all that money. Can you help me out? I think Venmo is going to kind of be confused by that. So how do you think about, you know, just user safety? I mean, there's so many levels to that problem, but that's just one that's, you know, jumping to the forefront of my mind.

• 44:13Div Garg?

  Definitely. So I would say that's one of the reasons we are in a closed beta where we have the choice where if you wanted to launch, we could just like launch basically everyone by through the Chrome like web store using an extension. The reason we have like made the choice to keep our current like user base small is to be able to make sure like everything is safe. There's nothing malicious, but like, like do like that teaming, uh, safety checks. And, and yeah, so this is like a very good point. So in terms of launch, I will say we are planning a standard launch in the next month. And so we'll start from there and like sort of see, okay, like what sort of like user, uh, how people using it, how can we sort of make it more safe? To keep it into the in a closed sort of like environment. And then once we know, okay, like what are the considerations on the system to improve it?

• 44:58Div Garg?

  We can do like a much finer launch. And so we are thinking of this into like phases where we do like a small launches first and then we'll do a big launch after that. Um, and this helps a lot with like security, a lot with like making sure everything is safe. Because especially with like very general web interaction agents like this, like if something goes wrong, that can create cause a very massive change action. Um, and we might not be able to come and like stop it in time. Um, so we want to make sure that we don't cause anything like that. Um, and also like for skills, um, yeah, really good point. Um, I would say like when it comes to choosing skills, we can like maybe like make a skill like protected or something like where it can't be modified by anyone.

• 45:37Div Garg?

  So this is like, if, if a skill is there, like it may be like only someone who has a, like some sort of a special access can go and like change those skills or the agent to be able to use them. Or we can also do something like you can only change individual level skills. So you can change me like skills that you define yourself, but you can change someone else's skill, for instance. And so like the only per, only person you will affect is yourself and so you won't be able to affect or create like malicious use for other users. Um. Yeah. And also what we're more worried about is like prompt injection attacks. But like people can just go and like hide prompts on websites and that can cause like weird things where like someone can just go and like hide like a crypto wallet link on a website and say like distribute this crypto wallet to everyone. And so this is something that we are trying to be very like safe about um.

• 46:20Div Garg?

  And we have a lot of tricks to make sure everything is safe. One thing is because we can uh type check our action like code in a sense before we actually simulate it in the browser. We can like make sure okay there was nothing malicious in this action code that we have created. Um. There's no sort of like red flags. There's no privacy violations. There's no nothing like hacking sort of like attempts happening. And once we have verified it, then we can actually uh simulate those actions. And that also obviously like adds like a great safety layer where we can certify everything is safe um before we do it.

• 46:54Nathan Labenz?

  Do you have an explicit sense of like how safe things need to be to be deployed? You know, people think about this in terms of like self-driving cars, you know, and I think we're kind of irrational on that one in that it seems like until it's clearly an order of magnitude safer than humans, it probably has no chance of getting deployed. And you know, to me, it seems like we maybe want to be a little bit more aggressive than waiting for like full order of magnitude better, you know, to start taking advantage of some of this stuff. I don't even know what the baseline is for a product like Multi-On, but how do you think about how many, you know, people per million can have like, you know, some adverse event?

• 47:40Nathan Labenz?

  Maybe that's not even the right way to think about it, but it's a tough one.

• 47:43Div Garg?

  Definitely. So it's like this like the um ways to think about this. The one is just like risk. If you're autonomous driving, like the risk is like you will basically kill someone and then you need to be like ninety nine point nine nine nine percent like accurate in a sense. Uh. If you want to like reduce that risk. Uh when it comes to like a lot of web interaction, the risk is much lower, especially in consumer settings. In enterprise, I think the risk is higher. That's why we're not like starting enterprise. Also, like a lot of consumer use cases, like there is more like maybe you might send a wrong message to someone, maybe you might make a slightly wrong purchase. And a lot of this are like reversible, so you can like uh like return back order for instance, and you cancel things, even if you send a wrong message to someone, you can be like, oh, sorry, my AI messed up. And it's actually really funny uh if you tell that to someone. So then like, yeah, so it all depends on like what are the risky scenarios?

• 48:26Div Garg?

  And like what we have done is like we've blacklisted websites which can be very risky. So for example, like banking websites, for example, or other websites where like it might be more like a sort of like a risk to the user. And in in in like low risk settings, I think it's completely fine to just deploy it, have people interact with it um and improve it over time. And I think this is the approach we are trying to take with multi on where uh let people interact with it, have fun with that, and like uh make it like much better over time.

• 48:51Nathan Labenz?

  You know, you're doing a lot of kind of hard work on it sounds like a lot of cases, right? I mean, the fact you've already kind of got a blacklist for sites, like there's a lot of little things like that that presumably you are gradually banging out and kind of sanding down all these rough edges. But then there's also potentially just like more breakthrough changes, you know, whether it's just foundation models get better, or maybe we see like new foundation models specifically built for agents. Uh what do you think are going to be the big things that really unlock this? And then, yeah, give me your timeline too for like, when does this actually become something where, you know, people that don't really care about AI

• 49:32Nathan Labenz?

  for intellectual reasons or curiosity reasons, actually just use this because it's better than, you know, using the internet browser themselves?

• 49:41Div Garg?

  There's a lot of like hurdles around just like making sure the experience is great. Um like finding the right combination of experience and engineering to make sure people can like use this like uh reliably and have the best experience possible.

• 49:57Div Garg?

  And yeah, I also think we, I think we'll be ready in terms of what we need. I think we are there at least for simple tasks. Unless you want to do like something like book me a one week trip to Italy, for example. Like we are still like far away from doing that. But if you want to like, yeah, maybe like book me a flight to like Venice, for example. On this particular date from this airport, this is something we can already do. Um. And then we can like make sure like we are doing this really reliably and add verification checks. Um. So yeah, so when it comes to like simple like tasks like that, I think we are mostly there. Um. With some rough edges currently, and I think in the next three months, I think it should be good. Where we can like just give it to non like non-technical people, have them play around with it.

• 50:39Div Garg?

  Like sort of like be like a browser companion or like try to change how like people interact with the browsers. Um. And when it comes to more complex tasks, I think that will take more time, especially if it's like a very abstract, high-level task. I don't think G P T four is there yet, uh, when it comes to a lot of really complex reasoning. And so you might need some like big uh tools in the foundation model layer. Where you can have it like do much more complex reasoning, uh, involving maybe like twenty plus uh uh steps. Um. Which it can currently do.

• 51:11Nathan Labenz?

  What do you think about this for the kind of future of work? So much of the stuff that people do and you know, I've mentioned to you offline that I'm involved in. Uh, A I advisory work with a friend's company, which is in the executive assistant space. And we're really trying to figure out like, how does a human E A relate to these kind of A I assistants? I guess you you said like how far is far and that's kind of maybe one way to think about the question. Because today our clients, you know, wouldn't really. They want to delegate all this stuff, right? They want to delegate, book the the thing at a high level and have it get done.

• 51:55Nathan Labenz?

  So they're probably not going to switch, you know, in the immediate or near term from. Working with a human who can handle that kind of task to an A I which can kind of. Take parts of it and accelerate bits of it, but can't really do the whole thing. Do you think that there is a like, what is your expectation for that kind of? Ever higher level sort of thing like is, do you see this kind of leveling out, you know, S-curved out to a place where it's like? Generally, things kind of remain the same, and this is sort of a tool for assistance. Do you see this kind of not leveling out and becoming, you know, something that literally rivals, you know, human employees? Where do we, where are we headed?

• 52:38Div Garg?

  Yeah, that's a good question. I would say initially, at least we think of ourselves more as like a supplementary aid rather than a replacement. For a lot of like current like assistance, I was like, we actually think like for a lot of our initial users, they might actually be like the prosumers or like assistants who want to just simplify their lives in terms of like scheduling. And so I think we, I think we are currently more supplement. Maybe in the next couple of years, I think it's possible you have human level full capabilities, but I think we're still a bit far ahead. What, what I see is like the evolution from like humans actually doing everything themselves to becoming more like organizers, where you can like instead of like booking, say like going to calendars, booking all the meetings, or like say going to different websites for

• 53:22Div Garg?

  ordering food on different restaurants, for example, or whatever booking flights. If you're EA, you could just go to multi on and the multi on will go and do the whole thing. But you just become a coordinator for the system. So you're just like, oh, now do this, now do that. And basically like my time will do all the actual interactions, all the web navigation stuff, but you become like the planner or the coordinator. And I think that is something that we can't still replace because I don't think foundation models are there yet. It might take a couple of years where they actually can do this really reliably without failing. And, and I think like the great thing about that is like we simplify a lot of like complex, very boring like interactions where like even for EA, the hard thing about their lives is just like you just have to go and interact with 100 different services.

• 54:04Div Garg?

  And if you can just like remove that from the lives, I think we can make and simplify a lot of like jobs and improve the quality of work.

• 54:12Nathan Labenz?

  How do you think about safety in the big picture of this? Obviously, the, you know, AI safety debates have heated up tremendously in recent months. And, you know, famously, some major thought leaders in the space have signed on to the extinction risk statement. It seems like the, you know, rogue agent AI is kind of the.

• 54:36Nathan Labenz?

  Most likely path towards some sort of short term, not necessarily mega disaster in the short term, but you know, the kind of thing that would be a real like shock to society to see, oh my God, AI, a rogue AI agent did that. My sense right now is that we're kind of limited more by just the limitations of the power of the best models than we are by any actual safety measures that we have. Like it's what my GPT four red teaming bottom line was. I think it is safe to release this model because its power is finite, not because you have it under control, really. So how do you see that race dynamic shaping up where like the capabilities are obviously improving?

• 55:17Nathan Labenz?

  The safety systems are not mature, you know, and you're sitting at this intersection of kind of maybe the most likely vector for like something crazy to happen in the, you know, in the short or medium term.

• 55:29Div Garg?

  Like the safety mechanism we need to have in AI agents is like have some sort of kill switch. And like one is like, okay, like you want to be able to disable them. And like there's like some mechanism where you can pull a lever and it's like, okay, like it's dead. It won't like basically won't take any other actions. Um, I think the current kill switch is just your API. So it's like, you just run out of your, your OpenAI uh budget and then undo anything. So I think that that's a current kill switch. I think that's very into agents in a sense, which is like one good way. But I think you can also maybe have like more explicit like uh controls where like you just like disable um it from like interacting with the elements, for instance, or you have more like hardware level safety mechanisms to. So I think this would be interesting to see.

• 56:10Div Garg?

  And in the code, I believe like you need to have some sort of tenants if you want to build really intelligent uh AI agents. And one of them should be like, an agent should not be able to modify its own source code. Because if it's able to modify its source code, then it could like self evolve. It could do like really weird things. It could find like ways maybe like if GPT four doesn't work, it can find like a cloud API key on the internet and just start using it or stuff like that. So I think that should be like one of the tenants where like um like it doesn't have source code access to itself. Uh, maybe it has like very explicit like access to different websites or you have permissive layers on what it can do and cannot do. Um, for us, I would say the funny thing is like when I initially started working on this uh back in February, besides sure this to Professor Stanford, and I was like, you know, like you literally went and like invented scanner.

• 56:54Div Garg?

  Like this is basically what scanner is supposed to be. And I was like, like, yeah, in a sense. And the other thing is like, it's like, I would say it's not there yet in terms of like, one is like foundation models. And second is also like, how you think about it? Like, if you make this into a A S system that is useful, you reduce the chance of it going rogue, you build in the safety mechanisms. I think you can build this into a very interesting technology.

• 57:20Nathan Labenz?

  Anything else that you wanted to share about your technology, your product vision for the future before we run out of time?

• 57:26Div Garg?

  Um, sure, sure. I would say for us, I think like we really care about like, can we like simplify or automate a lot of mundane things in life? Um, so you can imagine like currently there's like you have to contact with like hundred different services. It's really hard. Um, to do a lot of things where like, like maybe like you could have some sort of AI which can just like send like emails to people automatically or which is like coordinate for for events for you or automate a lot of these things. And I think this is where we see ourselves. We're like, we just want to intrinsically understand the user really well. And like automate things that they can just like delegate. So instead of you going and doing everything yourselves, you just tell the AI, oh yeah, do this for me.

• 58:08Div Garg?

  And the AI goes and does it. And so in a sense, we're calling this sort of like, we want to unlock parallelism for humanity where like currently you can imagine like we are all single threaded. We can do only one task at a time. But if we could have AI, we could launch like maybe like, I don't know, like hundred AI agents at a time, all the AI agents to do the work. And then tell you like, oh yeah, this work is done. And then you can like sort of like coordinate that. And I think that's what we see as the maybe might the evolution of technology where you just can coordinate a lot of this like AI agents to do your work. And it unlocks like a sync of it from primitive for us where we can like actually like become more parallel threaded if we can do this really well.

• 58:42Nathan Labenz?

  It's fascinating vision, uh, nice early execution and a really cool product that I've enjoyed piloting and playing around with. Div Gerg of Multiion, thank you for being part of the cognitive revolution.

• 58:57Div Garg?

  Cool. Sounds good. And yeah, thanks for inviting me here.