Agents' next frontier: agent-to-agent and network effects — Jean-Denis Greze, Town

Can you all hear me? Can you all hear me? Good. Um, well, first of all Good. Um, well, first of all , thank you for coming. I can't , thank you for coming. I can't , thank you for coming. I can't believe believe believe anyone came, but it's anyone came, but it's anyone came, but it's really nice. Uh, really nice. Uh, really nice. Uh, my name is Jean-Denis. my name is Jean-Denis. my name is Jean-Denis. Um, I'm the Um, I'm the Um, I'm the CTO at a company CTO at a company CTO at a company called Town. We won't called Town. We won't talk much about Town, you talk much about Town, you can go to town.com can go to town.com can go to town.com and take a look if you and take a look if you and take a look if you want, but that's not the want, but that's not the want, but that's not the main topic of main topic of main topic of today's today's today's conversation. I was the conversation. I was the conversation. I was the CTO CTO CTO at Plaid for 7 years, at Plaid for 7 years, at Plaid for 7 years, and before that I worked at and before that I worked at and before that I worked at Dropbox. And even before that, I Dropbox. And even before that, I Dropbox. And even before that, I built built built software for software for software for hedge funds. I've done a lot of hedge funds. I've done a lot of hedge funds. I've done a lot of things in my things in my things in my career, and now I'm career, and now I'm career, and now I'm working on uh, working on uh, working on uh, assistant agents assistant agents assistant agents for regular people, for regular people, for regular people, not for engineers, but not for engineers, but not for engineers, but for everyone in America and the for everyone in America and the for everyone in America and the world. And one of the things world. And one of the things world. And one of the things we're working on we're working on is systems where is systems where is systems where agents collaborate agents collaborate agents collaborate with other agents. with other agents. with other agents. That is, the interaction of That is, the interaction of That is, the interaction of agents. The main idea is agents. The main idea is agents. The main idea is that we see that we see that we see huge network huge network huge network effects if agents effects if agents effects if agents can work can work can work together to solve together to solve together to solve problems, because in the problems, because in the problems, because in the real world, real world, real world, most of us most of us most of us work with other work with other work with other people, right? Eh, people, right? Eh, people, right? Eh, the more the better. the more the better. the more the better. But I don't really But I don't really But I don't really think think think agent interaction makes much agent interaction makes much agent interaction makes much sense as a concept.

sense as a concept. sense as a concept. So I want So I want So I want to rethink the whole to rethink the whole to rethink the whole conversation through the lens of conversation through the lens of conversation through the lens of search. I believe that search. I believe that search. I believe that most most most LLM-based systems are simply a LLM-based systems are simply a LLM-based systems are simply a search task. And you search task. And you search task. And you try to make try to make try to make sure that the context sure that the context sure that the context window immediately window immediately window immediately before displaying the before displaying the before displaying the result to result to result to the user or the user or the user or calling the tool calling the tool calling the tool contains the correct contains the correct contains the correct information for information for information for the user. If you the user. If you the user. If you put the right put the right put the right information in the information in the information in the context window, then context window, then context window, then relying on the relying on the relying on the intelligence of LLM, you intelligence of LLM, you intelligence of LLM, you will get the best will get the best will get the best possible result. possible result. possible result. Um, so, you know, 4 years Um, so, you know, 4 years Um, so, you know, 4 years ago we did it ago we did it ago we did it manually, people manually, people manually, people would fill in the would fill in the would fill in the pop-up window themselves. pop-up window themselves. pop-up window themselves. Then, a few years Then, a few years Then, a few years ago, most people ago, most people ago, most people started started started using RAG, using RAG, using RAG, meaning they said: meaning they said: meaning they said: let's create a let's create a let's create a tool, tool, tool, for example, a search tool, for example, a search tool, for example, a search tool, that can that can that can browse different browse different browse different systems and collect systems and collect systems and collect data from there. And then data from there. And then data from there. And then people realized that it people realized that it people realized that it didn't scale very well didn't scale very well didn't scale very well and had its and had its and had its own problems. And right own problems. And right own problems. And right now we're completely now we're completely now we're completely focused on focused on focused on agent-based search—the agent-based search—the agent-based search—the idea is that you give idea is that you give idea is that you give an agent a lot of an agent a lot of an agent a lot of tools and they tools and they tools and they search across all the search across all the search across all the content; and then, content; and then, content; and then, hopefully, before hopefully, before hopefully, before making making making a call to the tool, a call to the tool, a call to the tool, it has exactly the it has exactly the it has exactly the content it content it content it needs to make the needs to make the needs to make the correct call correct call correct call and provide the correct and provide the correct and provide the correct information to information to information to the user. Um, and the user. Um, and the user. Um, and by the way, there are no by the way, there are no by the way, there are no people here. It's people here. It's people here. It's just one LLM challenge, the one that just one LLM challenge, the one that matters when there's the matters when there's the right context.

right context. right context. This, this is what you're This, this is what you're This, this is what you're trying to do. trying to do. trying to do. You're trying You're trying You're trying to design, um, this to design, um, this to design, um, this system. Steeply. So, system. Steeply. So, system. Steeply. So, how does this relate to how does this relate to how does this relate to agent interaction? agent interaction? agent interaction? So, I want you to So, I want you to So, I want you to imagine such a world. There imagine such a world. There imagine such a world. There aren't many aren't many aren't many agents out there who can agents out there who can agents out there who can do anything. There's only do anything. There's only do anything. There's only one agent, right? And it one agent, right? And it one agent, right? And it has one context has one context has one context window, and it has window, and it has window, and it has access to all the access to all the access to all the information in the information in the information in the universe. He can universe. He can universe. He can look at look at anyone's email, he can anyone's email, he can look at look at look at any any any company's information, he can company's information, he can company's information, he can look at look at any government's information, and he has it any government's information, and he has it right there in a pop-up right there in a pop-up right there in a pop-up window. And window. And window. And then you ask him then you ask him then you ask him to do something; you, you to do something; you, you to do something; you, you have your little have your little have your little system prompt with system prompt with system prompt with all this data, and what all this data, and what all this data, and what will happen is that it will will happen is that it will will happen is that it will give you the give you the give you the best possible best possible best possible result. And this, result. And this, result. And this, in fact, is what a in fact, is what a in fact, is what a multi-agent world is. multi-agent world is. multi-agent world is. It's just an agent that It's just an agent that It's just an agent that has access to all the has access to all the has access to all the world's information. world's information. world's information. This is the natural state of This is the natural state of This is the natural state of things. This is the ideal things. This is the ideal things. This is the ideal state of affairs. There is a problem with this state of affairs. There is a problem with this state of affairs. There is a problem with this state of affairs state of affairs state of affairs , and the problem , and the problem , and the problem arises because of something. arises because of something. arises because of something. So, in economics, So, in economics, So, in economics, they study the they study the they study the Coase theorem, which says that, Coase theorem, which says that, Coase theorem, which says that, essentially, even people, essentially, even people, essentially, even people, if they all have if they all have if they all have access to all the access to all the access to all the necessary necessary necessary information and there are no information and there are no information and there are no transaction costs transaction costs , we get an , we get an , we get an economically ideal economically ideal economically ideal outcome from a outcome from a outcome from a contract or contract or contract or negotiation. Well, it's the same thing negotiation. Well, it's the same thing negotiation. Well, it's the same thing . We can't . We can't . We can't gather all the contacts in gather all the contacts in gather all the contacts in the world, we can't the world, we can't the world, we can't provide them to LLM. Even provide them to LLM. Even provide them to LLM. Even theoretically, even with an theoretically, even with an theoretically, even with an infinite infinite infinite context window, context window, context window, because of because of because of privacy and privacy and privacy and security. We are people. I security. We are people. I security. We are people. I don't allow you don't allow you don't allow you to look at my mail, to look at my mail, to look at my mail, so there can't be an so there can't be an so there can't be an agent agent agent I would be willing to I would be willing to I would be willing to allow to allow to allow to look at my look at my look at my mail all the time. But if it mail all the time. But if it mail all the time. But if it existed, it would be existed, it would be existed, it would be very, very powerful.

very, very powerful. very, very powerful. So, I think this, this is So, I think this, this is So, I think this, this is kind of a test for a kind of a test for a kind of a test for a multi-agent multi-agent multi-agent system: how system: how system: how well does it well does it well does it approximate this? If it approximate this? If it approximate this? If it approximates this, it approximates this, it approximates this, it means: if you means: if you means: if you can get can get can get the same data in your window the same data in your window the same data in your window that a that a that a perfect perfect perfect system with access to system with access to system with access to all the data in the world could get, then all the data in the world could get, then all the data in the world could get, then you will get the you will get the you will get the optimal optimal optimal result. This is what result. This is what result. This is what you need to you need to you need to try to do. try to do. try to do. So, we're going to talk So, we're going to talk So, we're going to talk about five strategies about five strategies about five strategies that people that people that people use in use in use in different companies to different companies to different companies to try to get the try to get the try to get the data they need into this data they need into this data they need into this LLM challenge with LLM challenge with LLM challenge with externalities. So the externalities. So the externalities. So the first one is close first one is close first one is close access to everything access to everything access to everything within the trusted zone. within the trusted zone. So, my So, my wife and I have a joint wife and I have a joint wife and I have a joint agent, and he has agent, and he has agent, and he has access to my and her access to my and her access to my and her mail, including mail, including mail, including emails from before emails from before emails from before we became... well, we became... well, we became... well, that's okay, she doesn't that's okay, she doesn't that's okay, she doesn't ask me about that, ask me about that, ask me about that, but she asks, but she asks, but she asks, for example, if for example, if for example, if I have planned anything for I have planned anything for I have planned anything for our children or if I have our children or if I have our children or if I have fulfilled certain fulfilled certain fulfilled certain agreements with agreements with agreements with third parties. third parties. third parties. So the fact that our So the fact that our So the fact that our agent has access to agent has access to agent has access to both of our systems is both of our systems is both of our systems is just great. In a just great. In a just great. In a work context, work context, work context, this could be an HR agent this could be an HR agent this could be an HR agent who has access to who has access to who has access to all HR systems, just like all HR systems, just like all HR systems, just like any any HR employee, or at least HR employee, or at least as much access as the as much access as the average average HR team member. All HR team member. All HR team member. All HR employees HR employees can can can ask this agent ask this agent ask this agent questions, and, bam, it questions, and, bam, it questions, and, bam, it produces pretty good produces pretty good produces pretty good results. And now results. And now results. And now it is very popular. This is it is very popular. This is it is very popular. This is popular with IT and popular with IT and popular with IT and security departments security departments security departments because it is the same because it is the same because it is the same model as SaaS in model as SaaS in model as SaaS in security, so it security, so it security, so it works very well. I works very well. I works very well. I think there's a think there's a think there's a problem here, a problem here, a problem here, a fundamental fundamental fundamental problem that I problem that I

problem that I think about almost every morning think about almost every morning think about almost every morning when I wake up: Will the when I wake up: Will the system need fewer people over time system need fewer people over time ? And does this ? And does this ? And does this approach get better as approach get better as models improve? The problem with models improve? The problem with this approach is this approach is this approach is that the answer is "no" to that the answer is "no" to that the answer is "no" to both questions. You both questions. You both questions. You still need still need still need people to work with the people to work with the people to work with the data, and you don't data, and you don't data, and you don't get a magical get a magical get a magical unification of unification of unification of disparate data. You disparate data. You disparate data. You simply created a new simply created a new simply created a new isolated system isolated system isolated system because because because a person decided so. So the problem is a person decided so. So the problem is a person decided so. So the problem is , if this is your , if this is your , if this is your approach to building approach to building approach to building better AI, then in better AI, then in better AI, then in a few years you'll a few years you'll a few years you'll be in full be in full be in full flight. However, it's flight. However, it's flight. However, it's nothing to worry about. Your nothing to worry about. Your nothing to worry about. Your flight is my flight is my flight is my opportunity. I just opportunity. I just ...I don't want to be an ...I don't want to be an ...I don't want to be an [ __ ]. Sorry. That [ __ ]. Sorry. That [ __ ]. Sorry. That was rude. But I was rude. But I was rude. But I think it's a think it's a think it's a good way of good way of good way of thinking right now. It's just thinking right now. It's just thinking right now. It's just not the best way to not the best way to not the best way to think about the end think about the end think about the end result. Another result. Another result. Another approach, which I approach, which I approach, which I think is a little think is a little think is a little smarter, and I'll smarter, and I'll smarter, and I'll try to try to try to explain it, is to explain it, is to explain it, is to create create create tools that tools that tools that offer a different offer a different offer a different balance between balance between balance between power and power and power and privacy. So I'll privacy. So I'll privacy. So I'll give you an example. give you an example. give you an example. Um, the Um, the Um, the use case is that use case is that use case is that I want I want I want to ask my to ask my to ask my agent, "Is anyone in agent, "Is anyone in agent, "Is anyone in my company my company my company connected to anyone connected to anyone connected to anyone in the finance in the finance in the finance department at Acme Corp?" So, department at Acme Corp?" So, department at Acme Corp?" So, the way to do this without the way to do this without the way to do this without information information information silos is to simply silos is to simply silos is to simply give me access to the give me access to the give me access to the Gmail of everyone in Gmail of everyone in Gmail of everyone in my company. I'll my company. I'll my company. I'll see who's see who's see who's texting people at texting people at texting people at Acme Corp, then I'll Acme Corp, then I'll Acme Corp, then I'll look at their look at their look at their Google or LinkedIn profiles Google or LinkedIn profiles Google or LinkedIn profiles and say, "Oh, it looks like and say, "Oh, it looks like and say, "Oh, it looks like you text the CFO a lot you text the CFO a lot ." Can you ." Can you introduce me to introduce me to introduce me to him? But, of course, we don't him? But, of course, we don't him? But, of course, we don't want these want these want these information information information isolates. So what if isolates. So what if isolates. So what if you created a you created a you created a tool that would tool that would tool that would look at the Gmail of look at the Gmail of look at the Gmail of every employee every employee every employee with access to it with access to it and simply give you a

and simply give you a and simply give you a score of the strength of the connection? score of the strength of the connection? You would give this You would give this tool a domain and tool a domain and tool a domain and say, "I'm looking for say, "I'm looking for say, "I'm looking for someone who is a someone who is a CFO." He would CFO." He would look at everyone in the look at everyone in the look at everyone in the company who company who company who had sent letters to had sent letters to had sent letters to that organization, that organization, that organization, rate them rate them rate them , give you , give you , give you that rating, and the agent would that rating, and the agent would that rating, and the agent would get it and be like, get it and be like, get it and be like, "Cool." He would then "Cool." He would then "Cool." He would then use Slack to use Slack to use Slack to write to that person at the write to that person at the write to that person at the company. It would be something like company. It would be something like company. It would be something like this: " this: " Hi, Bob, I see you Hi, Bob, I see you Hi, Bob, I see you know Jane, the know Jane, the CFO of Acme Corp." And CFO of Acme Corp." And Bob would say, "Yes, Bob would say, "Yes, Bob would say, "Yes, that's right." And then your AI that's right." And then your AI that's right." And then your AI would say, "Oh, can would say, "Oh, can would say, "Oh, can I write a letter, or I write a letter, or I write a letter, or can you write a can you write a can you write a letter to letter to letter to introduce me?" And introduce me?" And introduce me?" And then Bob would agree, then Bob would agree, then Bob would agree, do it, you would do it, you would do it, you would meet, and everything would be meet, and everything would be meet, and everything would be great. So this is great. So this is great. So this is actually a really actually a really actually a really cool approach. I don't cool approach. I don't cool approach. I don't know how many of you know how many of you know how many of you do that. We at Town do that. We at Town do that. We at Town practice this for practice this for practice this for several things that several things that several things that our users often do our users often do our users often do . We . We . We ask ourselves: what ask ourselves: what privacy-preserving tool privacy-preserving tool would be acceptable to would be acceptable to would be acceptable to all of our all of our all of our users? They users? They users? They can opt out can opt out can opt out if they don't want if they don't want if they don't want to use it, but to use it, but to use it, but it has a natural it has a natural it has a natural network effect network effect network effect because it because it because it overcomes overcomes overcomes information silos in an interesting way information silos in an interesting way . Um, another interesting . Um, another interesting . Um, another interesting thing here is allowing thing here is allowing thing here is allowing other people other people other people to create drafts of to create drafts of to create drafts of emails in your emails in your emails in your inbox. You inbox. You inbox. You allow other allow other allow other employees in your employees in your employees in your company to write company to write company to write letters on your letters on your letters on your behalf because they will behalf because they will behalf because they will still ask you still ask you still ask you for introductions if for introductions if for introductions if they are on the they are on the they are on the sales team. So it's worth sales team. So it's worth sales team. So it's worth saving yourself saving yourself saving yourself a few clicks. So a few clicks. So a few clicks. So the question here is, will the question here is, will the question here is, will people accept the people accept the people accept the compromise on compromise on compromise on privacy that privacy that privacy that you make you make you make within the corporation within the corporation ? Oh, bad, bad. Oh my ? Oh, bad, bad. Oh my ? Oh, bad, bad. Oh my God. In a corporation, God. In a corporation, God. In a corporation, you know, it mostly you know, it mostly you know, it mostly works. So the problem works. So the problem works. So the problem here again is that it's here again is that it's here again is that it's all manual, not all manual, not all manual, not dynamic. It's manual dynamic. It's manual dynamic. It's manual work, because people have to work, because people have to work, because people have to think about think about think about tools. Maybe tools. Maybe tools. Maybe I could create I could create I could create tools, but it's

tools, but it's tools, but it's still manual because you still manual because you still manual because you have to explain to everyone have to explain to everyone have to explain to everyone what's happening. what's happening. what's happening. People may not People may not People may not like it if it like it if it like it if it happens without happens without happens without their consent regarding the their consent regarding the their consent regarding the trade-off between trade-off between trade-off between privacy and privacy and privacy and security. Steeply. And security. Steeply. And security. Steeply. And again, it doesn't again, it doesn't again, it doesn't get better, although " get better, although " better" is the better" is the better" is the problem. Steeply. So, the problem. Steeply. So, the problem. Steeply. So, the third category, it's third category, it's third category, it's extremely extremely extremely popular, but popular, but popular, but mostly only in mostly only in mostly only in the context of a single the context of a single the context of a single user. It's, user. It's, user. It's, you know, like you know, like you know, like personal wikis in the personal wikis in the personal wikis in the realm of cloud realm of cloud realm of cloud services. We would services. We would services. We would call it that, but it's call it that, but it's call it that, but it's about teams. about teams. about teams. So it's a shared, So it's a shared, So it's a shared, isolated space. isolated space. isolated space. You create a new You create a new You create a new place where data place where data place where data accumulates accumulates accumulates within your within your within your company and its company and its company and its divisions, and over divisions, and over divisions, and over time you start time you start time you start adding more and adding more and adding more and more information there. more information there. more information there. All agents have All agents have All agents have access to this, so access to this, so access to this, so you no longer keep you no longer keep shareable information in shareable information in isolation—it isolation—it isolation—it automatically automatically automatically goes into this goes into this goes into this public space. public space. public space. So, examples: So, examples: So, examples: shared skills. If shared skills. If shared skills. If you write code in an you write code in an you write code in an organization, organization, there are probably shared there are probably shared skills in your repository that everyone skills in your repository that everyone skills in your repository that everyone can improve on. Someone can improve on. Someone can improve on. Someone knows a better way to, knows a better way to, knows a better way to, say, say, say, profile a profile a profile a database, and can database, and can database, and can write such a write such a write such a skill. Next skill. Next skill. Next time someone sits and time someone sits and time someone sits and thinks, "Oh my God, the thinks, "Oh my God, the thinks, "Oh my God, the database query is so database query is so database query is so slow." He slow." He slow." He uses the uses the uses the profiling skill, and everyone profiling skill, and everyone profiling skill, and everyone becomes a better becomes a better becomes a better engineer. So, that's engineer. So, that's engineer. So, that's one version. Another one, which is one version. Another one, which is one version. Another one, which is quite popular, is quite popular, is quite popular, is that that that people have shared people have shared people have shared environments like environments like environments like wikis, Airtable, etc., and wikis, Airtable, etc., and wikis, Airtable, etc., and they have a habit of, "Hey, they have a habit of, "Hey, they have a habit of, "Hey, add more add more add more data to there over time." So, data to there over time." So, data to there over time." So, that's cool. And it works as that's cool. And it works as long as your agents long as your agents long as your agents have these tools, as have these tools, as have these tools, as well as some incentives well as some incentives well as some incentives to constantly to constantly to constantly exchange data between exchange data between exchange data between these isolated these isolated these isolated repositories. I think the

repositories. I think the repositories. I think the next version that next version that next version that some people are working on some people are working on is using an AI is using an AI cleaner. cleaner. cleaner. In fact, if In fact, if In fact, if there is one good there is one good there is one good idea in this report that I idea in this report that I idea in this report that I think works really think works really think works really well, it is this one. This is an well, it is this one. This is an well, it is this one. This is an AI cleaner. So AI cleaner. So , you have AI inside , you have AI inside , you have AI inside every private every private every private repository. The AI ​​has a repository. The AI ​​has a repository. The AI ​​has a policy on policy on policy on what should what should what should remain in the repository. remain in the repository. And it also has a description of all the And it also has a description of all the shared spaces shared spaces shared spaces you have. you have. you have. At the end of the day, he At the end of the day, he At the end of the day, he reviews the new reviews the new reviews the new information in the repository information in the repository information in the repository and posts it in and posts it in and posts it in public spaces. Well public spaces. Well public spaces. Well , publicly for , publicly for , publicly for your company. So your company. So your company. So it's the same as a it's the same as a it's the same as a personal wiki that personal wiki that AI eventually creates for you so that I can AI eventually creates for you so that I can learn about your learn about your learn about your goals, friends, and all that goals, friends, and all that goals, friends, and all that , but at a , but at a , but at a company level. Hmm, company level. Hmm, company level. Hmm, the hardest part is the hardest part is the hardest part is choosing what private choosing what private choosing what private information can be information can be information can be shared and shared and shared and added to shared added to shared added to shared repositories. And I think there are repositories. And I think there are repositories. And I think there are two approaches. There is the two approaches. There is the two approaches. There is the "ask the person" approach. It "ask the person" approach. It "ask the person" approach. It looks like this: LLM looks like this: LLM looks like this: LLM makes a list of things to makes a list of things to makes a list of things to add, and add, and add, and then asks then asks then asks the user, "Hey, the user, "Hey, the user, "Hey, do you mind if I do you mind if I do you mind if I put this in the shared put this in the shared put this in the shared space?" And you are reading space?" And you are reading space?" And you are reading this. You say: "Yes."

this. You say: "Yes." this. You say: "Yes." Saved a lot of time. Is Saved a lot of time. Is Saved a lot of time. Is n't that right? I n't that right? I n't that right? I mean, you wouldn't have mean, you wouldn't have mean, you wouldn't have done it otherwise. Uh, done it otherwise. Uh, done it otherwise. Uh, I guess the other option I guess the other option is when you actually is when you actually is when you actually ask LLM to implement ask LLM to implement ask LLM to implement the policy. And I believe the policy. And I believe the policy. And I believe that this is exactly what will that this is exactly what will that this is exactly what will happen very quickly. happen very quickly. happen very quickly. Um, and I think in the Um, and I think in the Um, and I think in the next 6 months we're going to next 6 months we're going to next 6 months we're going to have a bunch of systems have a bunch of systems have a bunch of systems where companies where companies where companies are going to trust LLM are going to trust LLM are going to trust LLM policies to policies to policies to automatically automatically automatically make more and make more and make more and more information public more information public more information public that was previously that was previously that was previously private. If you're in a private. If you're in a Fortune 500 company, unfortunately, I don't Fortune 500 company, unfortunately, I don't think that's think that's think that's going to happen anytime soon going to happen anytime soon going to happen anytime soon , but if you , but if you , but if you look at smaller look at smaller look at smaller companies, 10-50 companies, 10-50 companies, 10-50 employees, where employees, where employees, where trust is high and the trust is high and the trust is high and the likelihood of likelihood of likelihood of data abuse is low, and data abuse is low, and data abuse is low, and where it's clear where it's clear where it's clear what data can't be what data can't be what data can't be shared— shared— mostly financial mostly financial mostly financial and HR—you'll and HR—you'll and HR—you'll see a lot of see a lot of see a lot of this. And the cool thing about this is that this. And the cool thing about this is that it actually it actually it actually improves the trajectories of improves the trajectories of improves the trajectories of systems working on systems working on systems working on shared tasks shared tasks . Okay, that was the . Okay, that was the . Okay, that was the third approach. The third approach. The third approach. The fourth approach is fourth approach is fourth approach is pretty obvious: pretty obvious: pretty obvious: use use use people as people as people as intermediaries for intermediaries for intermediaries for information. It's like information. It's like information. It's like traditional traditional traditional agent interaction. My agent agent interaction. My agent agent interaction. My agent asks yours, "Hey, does asks yours, "Hey, does asks yours, "Hey, does anyone know anyone anyone know anyone anyone know anyone in the finance in the finance in the finance department at Acme Corp?" You, as a department at Acme Corp?" You, as a department at Acme Corp?" You, as a human, see the request human, see the request human, see the request and say, "Yes, I don't and say, "Yes, I don't and say, "Yes, I don't mind." "Find mind." "Find mind." "Find information in my information in my information in my mail." "And then it mail." "And then it mail." "And then it shows the result, and shows the result, and shows the result, and you say, 'Yes, I'm you say, 'Yes, I'm you say, 'Yes, I'm happy with this happy with this happy with this result.'" result.'" result.'" sending it to sending it to sending it to the person who asked it." The the person who asked it." The the person who asked it." The big problem here is that big problem here is that big problem here is that for for any question where few any question where few any question where few people have that people have that people have that information, you're information, you're information, you're essentially spamming essentially spamming essentially spamming everyone with that question.

everyone with that question. everyone with that question. So if I ask So if I ask So if I ask this question in a company this question in a company this question in a company of 100 people, 100 people of 100 people, 100 people of 100 people, 100 people will get a notification will get a notification will get a notification in Slack asking them to in Slack asking them to in Slack asking them to approve approve approve John's requests to John's requests to John's requests to use their use their use their personal network of personal network of personal network of contacts to find the contacts to find the contacts to find the answer to this, answer to this, answer to this, you know, huge you know, huge you know, huge question. That's not very question. That's not very question. That's not very efficient. So I efficient. So I efficient. So I decided there was a better decided there was a better decided there was a better way. It's very way. It's very way. It's very powerful, but I haven't powerful, but I haven't powerful, but I haven't seen it in seen it in seen it in practice very often. It's the practice very often. It's the black box approach. I black box approach. I black box approach. I wish I had wish I had wish I had a diagram for that. a diagram for that. a diagram for that. Unfortunately for all of you, Unfortunately for all of you, Unfortunately for all of you, I don't. Here's I don't. Here's I don't. Here's what it means. The what it means. The what it means. The black box approach black box approach is when you ask is when you ask is when you ask a question that can a question that can a question that can only be answered only be answered only be answered by looking at by looking at by looking at information in information in other people's isolated repositories. You have an LLM that no one has that no one has access to, that access to, that access to, that accesses all the data accesses all the data accesses all the data and finds the and finds the and finds the answer. Right? I mean answer. Right? I mean answer. Right? I mean , when I I say " , when I I say " finds an answer finds an answer ," it either gets ," it either gets ," it either gets an answer or it's an answer or it's an answer or it's going to make a going to make a going to make a call to some call to some call to some tool. And then tool. And then tool. And then it looks at what it looks at what it looks at what information was information was information was needed to needed to needed to make that make that make that call? And it call? And it only asks permission to make the call to the tool from only asks permission to make the call to the tool from the people who the people who the people who own that own that own that information. So in information. So in information. So in the example I the example I the example I gave earlier, when I gave earlier, when I gave earlier, when I asked 100 people in asked 100 people in asked 100 people in my company, "Hey, do my company, "Hey, do my company, "Hey, do you know the you know the CFO of Acme Corp?"

CFO of Acme Corp?" The request goes to the The request goes to the The request goes to the agents of every agents of every agents of every employee in my employee in my employee in my company. All of their company. All of their company. All of their agents look at agents look at agents look at their Gmail and private their Gmail and private their Gmail and private vaults to see vaults to see if they're connected to the if they're connected to the CFO. It's automatic. None of automatic. None of the people get the people get the people get a request for approval a request for approval a request for approval for that. Then for that. Then for that. Then the agent in the "black the agent in the "black the agent in the "black box" gets a box" gets a box" gets a list of 20 people who list of 20 people who list of 20 people who have a connection. It have a connection. It have a connection. It looks at the looks at the looks at the email contacts to email contacts to email contacts to see who has the see who has the see who has the strongest connection. strongest connection. strongest connection. It determines that it's It determines that it's It determines that it's Bob. And then it Bob. And then it Bob. And then it just asks Bob, just asks Bob, just asks Bob, "Hey, Jean-Denis wants "Hey, Jean-Denis wants "Hey, Jean-Denis wants you to introduce you to introduce you to introduce him to Jane, the him to Jane, the him to Jane, the CFO CFO CFO the CEO of Acme Corp. I the CEO of Acme Corp. I the CEO of Acme Corp. I know you're know you're know you're familiar with it. Can familiar with it. Can familiar with it. Can I share this I share this I share this piece of information piece of information piece of information with Jean-Denis? And you say, with Jean-Denis? And you say, with Jean-Denis? And you say, "Yes, of course." You "Yes, of course." You "Yes, of course." You click "yes," and it's click "yes," and it's click "yes," and it's fine. fine. fine. The important thing is that you The important thing is that you The important thing is that you have to trust the " have to trust the " black box." You black box." You black box." You have to trust that have to trust that have to trust that you can you can you can break down all break down all break down all the barriers to LLM having the barriers to LLM having the barriers to LLM having full access and not full access and not full access and not asking for permission until it's asking for permission until it's asking for permission until it's time to time to time to transfer the data or transfer the data or transfer the data or this final step. this final step.

In fact, In fact, it's quite it's quite it's quite feasible within the company. feasible within the company. feasible within the company. And by the way, your And by the way, your And by the way, your security and security and security and compliance team might agree with that compliance team might agree with that compliance team might agree with that . You just . You just . You just have to make have to make have to make sure that the sure that the sure that the human involvement phase is human involvement phase is human involvement phase is right, and you right, and you right, and you have to make sure have to make sure that you don't leave out that you don't leave out that you don't leave out unnecessary information unnecessary information unnecessary information with the final with the final with the final answer. So, answer. So, answer. So, you know, the nightmares you know, the nightmares you know, the nightmares and stuff—things and stuff—things and stuff—things like, oh, sorry like, oh, sorry . I mean, we have . I mean, we have . I mean, we have plenty of time. I'm plenty of time. I'm plenty of time. I'm almost done, so it's almost done, so it's almost done, so it's okay. The nightmare okay. The nightmare okay. The nightmare scenarios in these scenarios in these scenarios in these cases—is when cases—is when cases—is when someone can ask: " someone can ask: " someone can ask: " Are you connected to a Are you connected to a Are you connected to a recruiter at another recruiter at another recruiter at another company that you company that you company that you shouldn't be dealing with?" " shouldn't be dealing with?" " shouldn't be dealing with?" " to see if you're to see if you're to see if you're interviewing interviewing interviewing somewhere else, right? somewhere else, right? somewhere else, right? So, you know, sometimes with a So, you know, sometimes with a So, you know, sometimes with a black box, black box, you can inadvertently you can inadvertently you can inadvertently get information that get information that get information that you shouldn't you shouldn't you shouldn't have access to. You have access to. You have access to. You have to have to have to think really hard about how to think really hard about how to think really hard about how to build a great build a great build a great system. So, these are the system. So, these are the system. So, these are the approaches. I think if I were to approaches. I think if I were to approaches. I think if I were to bet on something that has an bet on something that has an bet on something that has an immediate immediate immediate payback for the world and for payback for the world and for payback for the world and for small companies, small companies, small companies, it would be a wiki page that's it would be a wiki page that's automatically automatically automatically generated by artificial generated by artificial generated by artificial intelligence.

intelligence. intelligence. Like a knowledge base Like a knowledge base that's constantly that's constantly that's constantly updated. I think there updated. I think there updated. I think there will be versions of that will be versions of that will be versions of that in the form of databases in the form of databases in the form of databases and wikis. And I think and wikis. And I think and wikis. And I think we're going to increasingly we're going to increasingly we're going to increasingly rely on big rely on big rely on big language models to language models to language models to make decisions about make decisions about make decisions about what's what's what's public and what's public and what's public and what's not. There are problems. not. There are problems. not. There are problems. So, obviously, injecting So, obviously, injecting So, obviously, injecting prompts into data silos prompts into data silos can be a can be a real problem real problem . If you have a more . If you have a more . If you have a more open silo, open silo, open silo, someone can add someone can add someone can add something bad in there, and something bad in there, and something bad in there, and it it it can come up in a search, and can come up in a search, and bad things can happen. It's very bad things can happen. It's very easy to get a easy to get a easy to get a shared wiki that shared wiki that shared wiki that just gets out of just gets out of just gets out of hand, you know? hand, you know? hand, you know? For example, one For example, one For example, one piece of information is piece of information is piece of information is wrong because wrong because wrong because the model made the model made the model made a mistake, and it poisons a mistake, and it poisons a mistake, and it poisons everything forever. I have a everything forever. I have a everything forever. I have a personal wiki that personal wiki that personal wiki that thinks my thinks my thinks my agent's name is Apex, even though agent's name is Apex, even though agent's name is Apex, even though I renamed it I renamed it I renamed it Ivy a month ago. And Ivy a month ago. And Ivy a month ago. And somewhere in the memory bank of somewhere in the memory bank of somewhere in the memory bank of my settings, Apex my settings, Apex my settings, Apex still lives on, and I can't still lives on, and I can't still lives on, and I can't get rid of it. That's get rid of it. That's get rid of it. That's normal for Apex. It's a normal for Apex. It's a normal for Apex. It's a funny case, but it's funny case, but it's funny case, but it's much worse if it's much worse if it's much worse if it's really wrong really wrong really wrong information about your information about your information about your business. If there's no human at business. If there's no human at business. If there's no human at any of the stages any of the stages any of the stages , , , obviously there's going to be obviously there's going to be obviously there's going to be false false false positives and positives and positives and disclosures of the wrong disclosures of the wrong disclosures of the wrong information. You know, information. You know, information. You know, sometimes sometimes someone someone gets fired because of a wrongful disclosure. And sometimes a gets fired because of a wrongful disclosure. And sometimes a gets fired because of a wrongful disclosure. And sometimes a false false false disclosure doesn't disclosure doesn't matter at all. And sometimes matter at all. And sometimes a client sues you a client sues you a client sues you . So, you know, you . So, you know, you . So, you know, you have to be careful.

have to be careful. have to be careful. Um, and I think all of Um, and I think all of Um, and I think all of this sounds great, but this sounds great, but this sounds great, but who approves what, what's who approves what, what's who approves what, what's being recorded, what can be being recorded, what can be being recorded, what can be undone? The undone? The undone? The black box idea is black box idea is black box idea is great, but it can't great, but it can't great, but it can't be truly " be truly " black." Someone in black." Someone in black." Someone in your company your company your company is going to want to is going to want to is going to want to do an audit sooner or later. They do an audit sooner or later. They do an audit sooner or later. They want to understand what's want to understand what's want to understand what's going on there, going on there, going on there, right? So there right? So there right? So there has to has to has to be a person at some level, maybe be a person at some level, maybe be a person at some level, maybe in the CIO's office in the CIO's office , who has , who has access to all the data. access to all the data. access to all the data. Um, yeah. So, um, what do I Um, yeah. So, um, what do I Um, yeah. So, um, what do I think? Well, do I think the think? Well, do I think the think? Well, do I think the future is future is future is automation? I've automation? I've automation? I've said that before. said that before. said that before. I think in I think in I think in programming we programming we programming we used used used to approve everything and to approve everything and to approve everything and then say, "We only live then say, "We only live then say, "We only live once, take a chance." once, take a chance." once, take a chance." And now the Entropic gods And now the Entropic gods And now the Entropic gods have given us have given us have given us auto mode. And auto mode. And auto mode. And auto mode auto mode auto mode tries tries tries to understand when we're to understand when we're to understand when we're being unreasonable being unreasonable and tells us. and tells us. and tells us. I think the interaction of I think the interaction of I think the interaction of agents across agents across agents across information information information silos will be the same silos will be the same . I think we'll get used . I think we'll get used . I think we'll get used to less to less to less sensitive sensitive sensitive information information information being pulled out and being pulled out and being pulled out and put into put into put into shared spaces. And shared spaces. And shared spaces. And then we'll have an then we'll have an then we'll have an area that area that area that needs human review needs human review needs human review or approval or approval or approval . Over time, LLMs . Over time, LLMs . Over time, LLMs will become will become will become more powerful. We'll more powerful. We'll more powerful. We'll learn to build learn to build learn to build into them better. security rules into them better. security rules . We'll be better at . We'll be better at . We'll be better at designing designing designing tools for the tools for the tools for the most complex most complex most complex areas that areas that areas that balance balance balance privacy properly. There will be more and privacy properly. There will be more and privacy properly. There will be more and more automation more automation more automation to create to create to create shared silos and shared silos and shared silos and even to decide even to decide even to decide whether to involve a human.

whether to involve a human. For example, if I For example, if I ask for notes from a ask for notes from a ask for notes from a weekly call weekly call weekly call with our with our with our vendor and vendor and vendor and their finance their finance their finance department. Maybe the LLM department. Maybe the LLM department. Maybe the LLM will say, "Oh, given will say, "Oh, given will say, "Oh, given your role, I don't your role, I don't your role, I don't need to ask either of them for need to ask either of them for need to ask either of them for permission." I permission." I permission." I can just can just can just share share share the notes with you. That's the notes with you. That's the notes with you. That's fine. She can fine. She can fine. She can look at the content. look at the content. look at the content. She sees what She sees what She sees what my role is. From a my role is. From a my role is. From a risk perspective, she risk perspective, she risk perspective, she can decide that I can decide that I can decide that I can can can disclose it. Um, and the disclose it. Um, and the disclose it. Um, and the cool thing about cool thing about cool thing about auto mode, by the way, is that auto mode, by the way, is that auto mode, by the way, is that if you if you if you design systems design systems design systems that way, they that way, they that way, they will scale will scale will scale with the power of with the power of with the power of the models. So I would the models. So I would the models. So I would advise: if you have advise: if you have advise: if you have agent interactions agent interactions agent interactions or cross- or cross- or cross- silo work, which I think is the silo work, which I think is the silo work, which I think is the better approach, better approach, better approach, make sure make sure make sure to define a to define a to define a low-sensitivity zone low-sensitivity zone low-sensitivity zone where you let the LLM where you let the LLM where you let the LLM make decisions. And make decisions. And make decisions. And accept it. And accept it. And accept it. And then, like then, like then, like magic, over magic, over magic, over time, things will get time, things will get time, things will get bigger and your bigger and your bigger and your system system system will naturally become will naturally become will naturally become more powerful, which is exactly what more powerful, which is exactly what more powerful, which is exactly what you need. you need. you need. So you want to be So you want to be So you want to be on the beach. That's what you on the beach. That's what you on the beach. That's what you want to do. That's want to do. That's want to do. That's where I want to be. With where I want to be. With where I want to be. With my kids in my kids in my kids in Hawaii. Okay, Hawaii. Okay, Hawaii. Okay, network effects. network effects. network effects. I have one minute. Uh, I have one minute. Uh, I have one minute. Uh, that's the conclusion. So we that's the conclusion. So we that's the conclusion. So we talked about five talked about five talked about five approaches: blah blah blah, approaches: blah blah blah, approaches: blah blah blah, trust boundaries, trust boundaries, trust boundaries, special special special tools, shared tools, shared tools, shared storage, people in the loop, storage, people in the loop, storage, people in the loop, and this version of the black and this version of the black and this version of the black box with a person in the box with a person in the box with a person in the loop. Uh, I think that's loop. Uh, I think that's loop. Uh, I think that's very powerful stuff.

very powerful stuff. very powerful stuff. I think the interesting questions are I think the interesting questions are I think the interesting questions are partly about partly about partly about companies, and I think that's companies, and I think that's companies, and I think that's going to work very going to work very going to work very soon. The big question soon. The big question soon. The big question is, are is, are is, are we ready we ready we ready for something like for something like for something like automatic automatic privacy? And I privacy? And I think the really think the really think the really interesting question is, interesting question is, interesting question is, I don't have I don't have I don't have the answer, but the answer, but the answer, but whoever does it will be whoever does it will be whoever does it will be richer than me, —is richer than me, —is richer than me, —is can you can you can you come up with scenarios where come up with scenarios where come up with scenarios where you can get multiple you can get multiple you can get multiple companies to agree that companies to agree that companies to agree that their their their data repositories data repositories data repositories have a common have a common have a common agent that works with agent that works with agent that works with all of them. For all of them. For all of them. For example, there's a company example, there's a company , which I won't , which I won't , which I won't name, somewhere in name, somewhere in name, somewhere in the world that works in the world that works in the world that works in finance, where there are a finance, where there are a finance, where there are a lot of investment lot of investment lot of investment banks. And actually banks. And actually banks. And actually investment banks investment banks investment banks benefit from benefit from benefit from sharing private sharing private sharing private data about private data about private data about private companies for companies for companies for things like things like things like lending. And they're lending. And they're lending. And they're starting to look in starting to look in starting to look in that direction. Where that direction. Where that direction. Where they trust they trust they trust each other's agents each other's agents each other's agents to be able to to be able to to be able to work with what work with what work with what used to be used to be used to be private private private information, and the agents information, and the agents information, and the agents decide decide decide what they can what they can what they can access and what they can't. And that's access and what they can't. And that's access and what they can't. And that's cool. That's really cool. That's really cool. That's really cool. And I think once cool. And I think once cool. And I think once you find you find you find some some some use cases across use cases across use cases across companies, uh, I companies, uh, I companies, uh, I think that's going to be a think that's going to be a think that's going to be a great springboard great springboard great springboard to move in to move in to move in that direction. So yeah that direction. So yeah that direction. So yeah . You know, as a human being . You know, as a human being , I ask myself, do I , I ask myself, do I , I ask myself, do I trust a trust a trust a future where future where future where agents make all the agents make all the privacy decisions? I privacy decisions? I don't know about that. I don't know about that. I don't know about that. I just I think that, for just I think that, for just I think that, for better or worse, that's the better or worse, that's the better or worse, that's the direction that direction that direction that things are going. And I things are going. And I things are going. And I just, I'm so on time just, I'm so on time just, I'm so on time . So, I'm on . So, I'm on . So, I'm on time. Thanks for time. Thanks for coming. I'm, coming. I'm, coming. I'm, uh, yes again. Thanks for being uh, yes again. Thanks for being uh, yes again. Thanks for being here and here and