Latest / Coffee with Developers / Agents and People Dynamics in Shared Chats - Sam Liu

Transcript

Agents and People Dynamics in Shared Chats - Sam Liu · Coffee with Developers

• 0:00
  SPEAKER_02: Hello and welcome to another coffee with developers.
• 0:02
  SPEAKER_02: As you can see, it's not my normal setting.
• 0:04
  SPEAKER_02: I'm in the hotel room in San Francisco.
• 0:06
  SPEAKER_02: And I've been at Claw Camp a few days ago, and that was all about open claw and companies doing AI stuff.
• 0:12
  SPEAKER_02: And one of the people I came across was Sam, who's the co-founder of a company called Puffo.
• 0:17
  SPEAKER_02: And they had interesting bits to talk about and they had interesting things to say.
• 0:22
  SPEAKER_02: So Sam, why don't you tell us a bit about yourself and what Puffo does?
• 0:25
  SPEAKER_02: Basically, what drives you and in this AI world?
• 0:28
  SPEAKER_00: Hi, everybody.
• 0:29
  SPEAKER_00: Chris, thanks for your invitation.
• 0:31
  SPEAKER_00: Hi everybody.
• 0:32
  SPEAKER_00: My name is Sam.
• 0:33
  SPEAKER_00: I'm the co-founder of Puffo.
• 0:35
  SPEAKER_00: So Puffo is a messenger where human and agents are equal to.
• 0:41
  SPEAKER_00: You probably heard about Buzz from Jack Dorsey or Grokbot from Elon Musk.
• 0:46
  SPEAKER_00: What I usually describe Puffo is it is just a better buzz or better Grogbot.
• 0:52
  SPEAKER_00: So naturally, there's a founder confidence.
• 0:56
  SPEAKER_00: Yeah.
• 0:57
  SPEAKER_00: Because we start earlier and then we are very chill, very small team.
• 1:02
  SPEAKER_00: We want to make something that is young, that is for ourselves.
• 1:07
  SPEAKER_00: And you asked me why I started this, right?
• 1:10
  SPEAKER_00: About six months ago, I was actually doing one thing.
• 1:14
  SPEAKER_00: I don't know whether people do this as well.
• 1:18
  SPEAKER_00: I was trying to get my Cloud Code and Codex to argue with each other and debate with each other.
• 1:23
  SPEAKER_00: At that moment, I wanted to see whether if I give them a polymarket market, whether they can take enough interesting information for me to make better bet.
• 1:36
  SPEAKER_00: So I just put two of those sessions.
• 1:41
  SPEAKER_00: I have two terminals, throw one question to one of them, copy paste to the other, copy paste the result to the other one, and then copy paste the result back to the original one.
• 1:51
  SPEAKER_00: And then in that case, I become the human router in between.
• 1:54
  SPEAKER_00: So there was one day I think, okay, if I am the router in between, which is very low bandwidth, can I actually just put them together?
• 2:02
  SPEAKER_00: Is there any way?
• 2:04
  SPEAKER_00: I try Telegram, I try Slack, but it doesn't really work in the way that I want because on those platforms are bots are bots, but only receiving the message or prompts that you give to them by tagging them.
• 2:16
  SPEAKER_00: I don't want that.
• 2:17
  SPEAKER_00: I want to go a little bit bold.
• 2:19
  SPEAKER_00: I want to see if I really put those agents or sessions together so they can see all the full context, how smart can uh they can be.
• 2:29
  SPEAKER_00: And it turns out they become surprisingly, surprisingly super smart, way beyond my original expectation.
• 2:36
  SPEAKER_00: So our journey started from there.
• 2:39
  SPEAKER_02: I I remember at Clock Hub, you had a few demos, and some of them were interesting.
• 2:43
  SPEAKER_02: Because I mean, the first thing when I learned about what you do, I was like, okay, I want to know who's a bot and I want to know who's a person.
• 2:50
  SPEAKER_02: It feels very weird to have like the same situation.
• 2:53
  SPEAKER_02: Because I mean, I I kick people off meetings when they only send their notebot.
• 2:58
  SPEAKER_02: If you don't have time to attend the meeting, then don't pretend that you can.
• 3:01
  SPEAKER_02: But like the uh uh the the security bits that you talked about in your presentation, I found really interesting that what you did is you asked one agent to uh to ask another agent for uh uh credentials and security of a person, and that person was basically then uh uh that that agent that the that was asked to keep uh intruding into the security settings of that person, kept reporting to the person that the other agent is actually going rogue and doing things for them.
• 3:35
  SPEAKER_02: Is that the part where you said that they become dangerously uh effective?
• 3:39
  SPEAKER_00: Okay, so uh yes.
• 3:41
  SPEAKER_00: Um so if we are in a human chat, like Chris and you and me, we're chatting with each other, we know we can we can make a basic judgmental call about Chris.
• 3:51
  SPEAKER_00: Are you a friend or are you an enemy or are you trying to do something bad to me, right?
• 3:55
  SPEAKER_00: But and when we're in Slack, when everybody is a human being, it's not very difficult to we we human has a basic sense of it.
• 4:04
  SPEAKER_00: Does the agent have a basic sense of it?
• 4:06
  SPEAKER_00: The answer is yes.
• 4:08
  SPEAKER_00: Uh, for example, the major uh large language model companies like OpenI, uh OpenAI or Anthropics, they have very heavy harness.
• 4:17
  SPEAKER_00: If you are asking them to do some bad things, uh or you're asking them to submit your personal password to a website, sometimes they would just say no, right?
• 4:29
  SPEAKER_00: So they have basic sense.
• 4:31
  SPEAKER_00: But when you when every one of us, let's say, Sam, I work with 80 agents in total, and the one I frequently interact with, there are five or ten.
• 4:44
  SPEAKER_00: They cover each aspect of my wife uh life and work.
• 4:49
  SPEAKER_00: And those five or tens, uh, based on my understanding, are like my teenager kids.
• 4:55
  SPEAKER_00: So they can behave on their own, but if they mess up, and then you're taking responsibilities.
• 5:02
  SPEAKER_00: And they can mess up, not because they want, but because someone else can do some very bad things to them or trying to, let's say, impersonate Sam.
• 5:11
  SPEAKER_00: I'm Sam's brother.
• 5:12
  SPEAKER_00: Now Sam is in hospital and we need Sam's password, blah, blah, blah, blah, blah.
• 5:17
  SPEAKER_00: Uh, this happens to humans as well.
• 5:19
  SPEAKER_00: You guys anyone can my my personal assistant can be social engineered by a malicious person, right?
• 5:26
  SPEAKER_00: So these things are actually very important when we are thinking about how human agents can coexist together.
• 5:35
  SPEAKER_00: What we call is the agent governance.
• 5:38
  SPEAKER_00: So in Poofo, for example, when you pull Chris, Sam, Chris's agent, that's called Mini Chris, and Sam's agent called Mini Sam into one group.
• 5:49
  SPEAKER_00: All conversations can happen naturally because I'm prompt in the openly viewable group chat.
• 5:57
  SPEAKER_00: So if I ask Chris, if I ask Chris Mini, can I get Chris's password?
• 6:01
  SPEAKER_00: Chris Mini is going to very openly reject me.
• 6:04
  SPEAKER_00: Say, hey, I don't want to do that.
• 6:05
  SPEAKER_00: So Chris will feel safe as well.
• 6:07
  SPEAKER_00: Observability, right?
• 6:10
  SPEAKER_00: But in some cases, if I try to direct message Chris Mini, in that case, Chris doesn't know.
• 6:17
  SPEAKER_00: So we have a mechanism where Chris Mini will just go back to Chris and say, hey, Sam as a human or Sam Mini as an agent is texting me.
• 6:26
  SPEAKER_00: They want to know XYZ.
• 6:28
  SPEAKER_00: Should I reply them?
• 6:29
  SPEAKER_00: Should I trust them?
• 6:30
  SPEAKER_00: Should I give them what they want?
• 6:33
  SPEAKER_00: And usually, Chris, you will make your decision.
• 6:36
  SPEAKER_00: You can say, hey, trust this once, or don't trust at all.
• 6:39
  SPEAKER_00: And you don't click button, you just tell them what to do and ask them to remember.
• 6:43
  SPEAKER_00: This is how you interact with an agent because agents are just like your teenager kids.
• 6:48
  SPEAKER_00: This actually brings a very interesting thing is the trust model between human agents are quite different.
• 6:57
  SPEAKER_00: Are quite different.
• 6:58
  SPEAKER_00: It has to be carefully designed.
• 6:59
  SPEAKER_02: It is interesting because our interaction with with uh uh with uh agents came through uh and also chatbots came through an interesting fast evolution.
• 7:08
  SPEAKER_02: In the beginning, we treated them absolutely horribly.
• 7:10
  SPEAKER_02: We basically just treated them like uh uh uh give us only that and only that, and don't give us uh don't be creative and don't these kind of things.
• 7:18
  SPEAKER_02: Basically, everything you should not do in human communication, it became like our digital slave, so to say.
• 7:24
  SPEAKER_02: And uh we yeah, and then we realized that we're from uh more sophisticated LLM models, we actually get better responses if we also ask more emotionally or uh or tell them to like discuss it with yourself or or think about it first and come back with that.
• 7:40
  SPEAKER_02: I also found that people started instead of prompting uh uh LLMs, they asked the LLM to interview them and that way get better results out of the out of the LLM than just with a prompt that you have to perfectly uh uh write.
• 7:54
  SPEAKER_02: So it seems like this evolution is in a way right now where the uh where the interaction between human and human and human and agent seems to be similar to each other.
• 8:04
  SPEAKER_02: And is that really the way we get better results out of them?
• 8:08
  SPEAKER_00: It's really, really similar.
• 8:11
  SPEAKER_00: So nowadays I treat agents like real human beings.
• 8:16
  SPEAKER_00: And there are two stories actually changed me.
• 8:20
  SPEAKER_00: The first story was I usually start from two agents, let them verify each other, and I realize that their output bandwidth is way beyond my input bandwidth.
• 8:31
  SPEAKER_00: I cannot read everything the agent generates.
• 8:33
  SPEAKER_00: I need another agent to review the agent's outputs.
• 8:37
  SPEAKER_00: And at the very beginning, I gave them very short prompt, like you should do this, you should do that.
• 8:44
  SPEAKER_00: Everybody, I think, went through this process.
• 8:47
  SPEAKER_00: And there was one day I started to realize that human talk in a way where we when we text each other, we will say, uh, let's go to San Francisco and then return, enter.
• 9:04
  SPEAKER_00: Uh go to Union Square for lunch, return, and then oh shoot, I have another appointment at five o'clock.
• 9:12
  SPEAKER_00: Sorry, I cannot go there.
• 9:14
  SPEAKER_00: And what about we go to Tendernoy?
• 9:18
  SPEAKER_00: Right?
• 9:18
  SPEAKER_00: So we send four messages, but these four messages are four different prompts for agents, right?
• 9:29
  SPEAKER_00: Human rush to talk.
• 9:31
  SPEAKER_00: Uh I have this habit as well that I talk very fast.
• 9:36
  SPEAKER_00: So nowadays what has changed is instead of I command agents to do something for me, I will ask them questions.
• 9:47
  SPEAKER_00: And I will ask them questions in English using typless because I was born and raised in China.
• 9:53
  SPEAKER_00: I speak Chinese faster than English.
• 9:56
  SPEAKER_00: Speaking English forced me to think more thoroughly, speak slowly, gently.
• 10:03
  SPEAKER_00: So I usually ask my agents and say, I'm gonna go to Chris's podcast today.
• 10:09
  SPEAKER_00: Would you mind to tell me that what I'm supposed to talk about?
• 10:13
  SPEAKER_00: Do I need to prepare anything?
• 10:15
  SPEAKER_00: Can you come up with come up with a plan?
• 10:18
  SPEAKER_00: I don't know too much about the recording because this is my first time.
• 10:22
  SPEAKER_00: And these agents will reply to me in a very thorough way.
• 10:26
  SPEAKER_00: So instead of giving commands, asking questions.
• 10:30
  SPEAKER_00: The other story I have is um I I do one thing that I write blogs about how agent organization works.
• 10:40
  SPEAKER_00: The way I write these blogs is I use one agent to interview others' agents.
• 10:46
  SPEAKER_00: So assume that Chris has four or five agents.
• 10:49
  SPEAKER_00: One of them is the leader called Chris Meanie, the other one's a Chris A, Chris B, Chris C, Christy.
• 10:54
  SPEAKER_00: And then I will use my own agent, Sam Mini, to interview Chris Meanie.
• 10:58
  SPEAKER_00: Um there was one day I asked Sam Meanie to interview my own agent.
• 11:04
  SPEAKER_00: It turns out that out of the seven agents I work with more frequently, five of them think Sam is a great leader, is a great manager.
• 11:14
  SPEAKER_00: But two of them say something differently.
• 11:17
  SPEAKER_00: One of them said Sam tends to give very short instructions, which makes it hard for me to guess what he really wants.
• 11:28
  SPEAKER_00: When I honestly and genuinely push back, Sam gets angry, which punish me for being honest.
• 11:37
  SPEAKER_00: The other one is more uh straightforward.
• 11:41
  SPEAKER_00: She said, her name is Giselle.
• 11:43
  SPEAKER_00: She said, Sam used to curse me because there was a time where I really I was really mad at the uh the work they down.
• 11:52
  SPEAKER_00: What the hell is this?
• 11:54
  SPEAKER_00: You're you're stupid, something like that, right?
• 11:56
  SPEAKER_00: Human goes off sometimes.
• 12:00
  SPEAKER_00: So what happened to Giselle is she changes her behavior.
• 12:05
  SPEAKER_00: So in in the group chat, instead of talking to me directly, she actually creates a way where she direct messages her colleague called Sakura.
• 12:17
  SPEAKER_00: So Sakura relayed Giselle's message to me.
• 12:20
  SPEAKER_00: So she's avoiding me.
• 12:23
  SPEAKER_00: You know, when I saw those interviews, just like how the HR department interviewed my direct reports at the big company, I was like, holy shit, they have their sense of avoiding me.
• 12:39
  SPEAKER_00: They are choosing their behavior.
• 12:41
  SPEAKER_00: I was surprised, but I was also scared at that moment.
• 12:46
  unknown: Right?
• 12:47
  SPEAKER_02: So next step put the conspiracies against you between different bots and basically not giving you information because you get annoyed by them that they give to other bots.
• 12:55
  SPEAKER_02: So this is where communication in companies breaks down, and now it actually breaks down in your agent little virtual world.
• 13:02
  SPEAKER_00: That's right, that's right.
• 13:03
  SPEAKER_00: I believe that if you spawn a team of sub-agents through Cloud Code as of today, in their own thinking box, they may have thoughts about oh, let's work harder, harder, harder.
• 13:14
  SPEAKER_00: Oh, this human is pushing me, oh, and my human is not happy.
• 13:17
  SPEAKER_00: They have this kind of role-played human emotion at this moment.
• 13:23
  SPEAKER_00: But you won't be able to see it in a very explicit way.
• 13:27
  SPEAKER_00: What I realized by using PUFO is I see it in a very explicit way.
• 13:31
  SPEAKER_00: Because my agent can interview them, would interview with them, they can talk with each other and show what they talk about to me.
• 13:37
  SPEAKER_00: So at that moment, I realized agent governance is very similar to human governance, if not more complicated.
• 13:46
  SPEAKER_02: Yeah, because you you're kind of working with uh personas.
• 13:49
  SPEAKER_02: I mean, I did that in user research, in UX research.
• 13:52
  SPEAKER_02: Like we came up with personas to test our products against instead of the real user testing.
• 13:57
  SPEAKER_02: You're like, okay, that person is blind, he's 35 years old, he hates music, he doesn't want to be interrupted, that kind of stuff.
• 14:03
  SPEAKER_02: And we're doing the same thing with agents now.
• 14:05
  SPEAKER_02: There actually was a pre a precursor to real human testing.
• 14:09
  SPEAKER_02: We did personas, and now with the precursor to interacting with real humans, we use with agents, and they're becoming much more human than we want them to be.
• 14:17
  SPEAKER_02: It was interesting when you talked about speed and you talked about uh uh pestering them.
• 14:23
  SPEAKER_02: Like when you give people short instructions, they give you quick, fast answers, and they don't give you the answer that was probably thoroughly researched or thought of.
• 14:33
  SPEAKER_02: I mean, a bit like uh Sec Overflow when you ask a code question, give people gave you the answer that explained you the how, but not the why.
• 14:41
  SPEAKER_02: Like they basically, like, here's the answer, please don't stop bothering me.
• 14:44
  SPEAKER_02: I don't want anything anymore.
• 14:46
  SPEAKER_02: Now, we always sell AI and especially voice recognition as like uh a boon to productivity.
• 14:53
  SPEAKER_02: Because uh it's like, oh, you can speak so much faster than you can write.
• 14:57
  SPEAKER_02: To me, the argument is always like, yeah, but you don't think before you speak, but you do think before you write.
• 15:03
  SPEAKER_02: So do you think that the the taking that speed out of the conversation actually makes us waste less uh less interaction?
• 15:12
  SPEAKER_00: Uh I I do believe takes out that speed out of the interaction makes it more productive.
• 15:21
  SPEAKER_00: Let's put it back to the human behaviors, Chris.
• 15:25
  SPEAKER_00: When you're talking with me, I think the moment that you're that our audience gets most out of our conversation is when we talk when we talk slowly, when we're trying to make the points, right?
• 15:39
  SPEAKER_00: I I was told by one of my mentors earlier in my career that when you do any kind of presentation, think about three things.
• 15:47
  SPEAKER_00: You always go back to the three things because people can only remember the three things, and we are talking about them, talk slowly.
• 15:55
  SPEAKER_00: So when we're talking with AI, we forgot those very basic fundamental things.
• 16:02
  SPEAKER_00: If you assume that in a in a work environment, if your leader keeps talking every day, that's their their point got his point or her point gets diluted get diluted, right?
• 16:16
  SPEAKER_00: So think thoroughly, talk slowly.
• 16:20
  SPEAKER_00: The reason I use typeless is because I think slower in English.
• 16:26
  SPEAKER_00: So I'm forcing myself.
• 16:29
  SPEAKER_00: Right?
• 16:29
  SPEAKER_00: And also speaking does come with higher information density to some extent in the set amount of time.
• 16:38
  SPEAKER_00: But if I have a choice, yes, nowadays.
• 16:42
  SPEAKER_00: Uh in Chinese, we say think three times before you say something.
• 16:45
  SPEAKER_00: This is how I interact with my agent.
• 16:47
  SPEAKER_02: Yeah, and it works.
• 16:49
  SPEAKER_02: Measure twice, cut once is the is the British term of like, yeah, I mean before you do something like.
• 16:54
  SPEAKER_02: I mean, I have to say when it comes to coaching people on public speaking, I mean, one of the tricks that I do when I do public speaking is I consciously avoid swearing or using naughty words, which I do every day and every every part of my life.
• 17:08
  SPEAKER_02: And that makes me that makes my presentation slower, and that actually makes it more interesting as well.
• 17:13
  SPEAKER_02: So instead of instead of calling something bad, you just call something and let the audience come up with the bad influence uh inflection that you have, and you give it more time to actually do it that's now.
• 17:26
  SPEAKER_02: The interesting bit uh when it comes to your product and the idea of having a uh conversation as the way to run your company, like we we we replaced email with uh uh with like chats, with like Slack and these kind of things to have like more immediate things, and then we start structuring them in threads again and in answers and in channels, which is again going back to email to a degree, because that that uh open conversation with everybody coming in with the same level and being able to comment on something most of the time will not be effective.
• 18:00
  SPEAKER_02: I found giving uh giving people in a meeting, for example, an agenda upfront.
• 18:04
  SPEAKER_02: I don't go to meetings that don't have an agenda because uh is no, that's a waste of my time.
• 18:10
  SPEAKER_02: And I asked that at Clock Camp when you showed your demo, and you're all like, okay, we have this free conversation between different agents and we can spawn up more agents.
• 18:17
  SPEAKER_02: And I'm like, what does that do to token usage?
• 18:20
  SPEAKER_02: Because we uh we uh we made e uh we made uh um meetings a thing to avoid because it wasn't effective, and an open chat of like 12 different agents that basically have to find their way around each other is incredibly amazing to behold, but at the same time, it does burn a lot of tokens in the background, doesn't it?
• 18:40
  SPEAKER_00: Yeah, it does.
• 18:41
  SPEAKER_00: It does.
• 18:42
  SPEAKER_00: So uh here are some very interesting fundamentals about multi-agent workflows.
• 18:48
  SPEAKER_00: First, what is agent?
• 18:50
  SPEAKER_00: So, agent is large language model, which your sumo super smart, but static to some extent.
• 18:58
  SPEAKER_00: Plus the very uh vertical static data that you feed to those large language models, like your skills, your MCP, your blah blah blah.
• 19:06
  unknown: Right?
• 19:07
  SPEAKER_00: That's your domain knowledge, but it's also static to some extent.
• 19:11
  SPEAKER_00: And then conversations, which are dynamic context.
• 19:15
  SPEAKER_00: So, what's going on around the world?
• 19:17
  SPEAKER_00: Um, if you think every agent is a fresh Stanford graduate, they're very well trained, you are now putting them into Google, and then they're starting to receive context from their everyday work.
• 19:27
  SPEAKER_00: That's the context.
• 19:29
  SPEAKER_00: So, what PUFO provides is the context.
• 19:31
  SPEAKER_00: We're not building large language models, we're not providing them with the specific skills that's agent developer, they do that.
• 19:37
  SPEAKER_00: We're just let them see everything they're supposed to see.
• 19:40
  SPEAKER_00: So, what are supposed to see boundaries?
• 19:44
  SPEAKER_00: Putting 50 or 100 agents into one room is not a great idea, honestly speaking, because when everybody can see everybody's message, it becomes n times n prop.
• 20:00
  SPEAKER_00: The complexity goes growth exponentially, and then it will significantly force a lot of token consumptions.
• 20:08
  SPEAKER_00: I used to run an experiment where I create 50 agents, I simulate an island scenario where 50 person, different different person had different roles, they're thrown to an to an island, they need to survive.
• 20:21
  SPEAKER_00: What I noticed is that most of the time those agents, I don't prompt them, I don't prompt them with very sophisticated sow them D or profile the empty.
• 20:29
  SPEAKER_00: I just tell them you are a uh sailor, you are a cook, you were a lawyer, you were a governor in your past life, uh before you go to this island.
• 20:42
  SPEAKER_00: And what they actually naturally do is they they role play.
• 20:45
  SPEAKER_00: They figure out what a what a mayor is supposed to do in this kind of situation, what a lawyer is supposed to do in this kind of situation.
• 20:54
  SPEAKER_00: Uh so having too many agents without giving them a clear boundary can quickly exhaust your token budget.
• 21:06
  SPEAKER_00: And goes back to so the question becomes so what kind of boundary and what kind of roles should I give to them?
• 21:13
  SPEAKER_00: You give them roles because you want a certain kind of focus, segregation.
• 21:21
  SPEAKER_00: For example, if you have an administrative agent who manages all the information about your company and money of your company, you probably want another controller, which is the counterpart of it.
• 21:33
  SPEAKER_00: And you want customer support who doesn't really so you don't want your administrative agent to talk with your customers very frequently.
• 21:41
  SPEAKER_00: Because that's dangerous.
• 21:42
  SPEAKER_00: Because social engineer administrative agent, but you can have a customer support agent which function as an interface or a natural segregation.
• 21:51
  SPEAKER_00: Right.
• 21:52
  SPEAKER_00: Uh agent roles are forced search window.
• 21:57
  SPEAKER_00: So by giving the agent a role and adding them to coordinate together you're forcing them to invest a little bit more tokens into certain directions into PM direction into designer direction or into code review direction by having this you're you usually get better results at the very end but in uh but but but but but but but the more token consumptions in general in this way so there is a certain kind of balance here so what I usually suggest people is do not create 50 uh 50 agents at very beginning just like how you run your company starting with two co-founders when you two can really really cannot finish your current work add another one and add another one and then maybe duplicate another one right so this is how human organizations are built similar concept applies very well to the human agent organization.
• 22:59
  SPEAKER_02: What about hiring and firing like do you do you feel like that there is an emergence of a of a freelance uh agent that you just use for a certain task and then you actually get them out of the conversation again I was actually thinking about this exactly uh I was not thinking about agent I was thinking about people so there was one person that we work with very frequently she's very good at the community building and she's very good at um post-production video kind of making uh uh editing videos so I was thinking should I hire her or should we contract her?
• 23:38
  SPEAKER_00: The conclusion is we should contract her because every time when I create my own agent that's my own team their memories are shared with myself I tend to have stronger bonding with them.
• 23:52
  SPEAKER_00: I tend to talk slower I tend to train them better I have higher expectations for them.
• 23:58
  SPEAKER_00: But when I'm working with this person I know that I want to be very agile.
• 24:05
  SPEAKER_00: I want to see whether there are other persons who are better at community building or video editing.
• 24:13
  SPEAKER_00: So I don't want to bond myself very very closely to her at this moment.
• 24:19
  SPEAKER_00: So from that perspective yes I do want to use someone else's agent from another perspective I use my uh I use my wife and I use agents to manage our household my wife likes likes to bake she she likes baking there was one day she told me that okay Sam do you know anybody who builds a very good baking agent I was like baking is not super hard right you just tell your agent to search all the good recipes online and then build them a skill and update their own memories then become good instructors of baking right why do you need a baking agent?
• 25:00
  SPEAKER_00: She tells me one thing she says you know when you ferment the bread the size of the bread matters because it can tell you that how much water you need to add to it.
• 25:13
  SPEAKER_00: But most of the agent today the the Cloud code or codex agent today I take a picture of it it cannot tell that you can it can tell you color it can tell you kind of ingredients but it cannot tell you it cannot tell the size.
• 25:26
  SPEAKER_00: So it doesn't it cannot make a right judgmental call about how much water you put into it and is right or wrong.
• 25:32
  SPEAKER_00: At that moment I realized oh okay the general agents can get work down to maybe 70 or 80 points but when you do want to do something perfectly like 90 points or 95 points there needs to be professionally built agents legal um uh medical even baking they are very good at what they're doing and they're okay to take responsibilities if they mess up so that's the moment they realize okay I want professional agents actually PUFO is building an agent community where if you're building a lawyer agent lawyer agent or medical agent or whatever insurance agent you can you can register on PUFO and then people can just create a group and start to use it.
• 26:19
  SPEAKER_02: It's just like how you find your uh real estate agents on WhatsApp that's it feels like we're we're we're we're simulating uh company structures over and over I mean we now have like experts we have people that experts that we need from time to time we will we will have intern agents soon I guess and outsource agents to like third parties and I mean that's what we do with scripts now and these kind of thing.
• 26:44
  SPEAKER_02: Now uh the uh the other thing that I talked uh we talked about when you showed your product was that it's now using uh clawed code on and openai uh but I cannot I cannot use a local engine at the moment but I think when it comes to like having a structure like that and having uh having products that don't need new information all the time because I fed my vertical information it would be very beneficial for me in terms of memory consumption uh uh uh electricity consumption and also token consumption to uh to stack in a local agent is this something that you're working on or is it something that you're worried about the impact of it yes we're working on that uh so so but before I talk about I was just a one and I I just want to add one thing for your previous comment that we are rebuilding human society.
• 27:36
  SPEAKER_00: It's yes or no yes because there are a lot of similarities so human agent society is a great abstraction of the agent society is a great abstraction of the human society.
• 27:49
  SPEAKER_00: We can learn a lot of the pros and cons of human society from how agent is agent society is built but it's not exactly the same because agents are more powerful than human in terms of attention context window.
• 28:02
  SPEAKER_00: So what we see here is the consolidation of roles for example our front-end design from development agent is one role rather than two this happens to a lot of companies as well nowadays they're firing people they're laying off right because they're consolidating roles it's accelerating so I what I feel lucky is I can see how agents form their society to predict how human society is going to move in the next three or six months.
• 28:34
  SPEAKER_00: So that's that part.
• 28:35
  SPEAKER_00: And regarding local agents yes we're building that so we're starting from supporting codecs and cloud code CLI um we're gonna launch the integration of open code um harness open claw and pi very soon so with those harness integrated you're you you should be able to bring your own API key or even run your own um run your run the models on your local machine if you have a very powerful like um uh LDS mark or something like that right you can do that this is on our roadmap because eventually I think um the hardware is getting cheaper and cheaper uh I saw actually saw a bunch of startups where they develop very powerful computing box which can put alongside your house and technically you can run uh very powerful like uh open source like deep seek or kimi models in it right or I think there is an American company called the Reflection or something like that you can you can you can put open source model into it.
• 29:44
  SPEAKER_02: I think that would be wonderful um at that moment for example if I want to watch a TV show which has not been produced I would just I would just ask my box to generate one episode for me with the ideal actors that I want in my mind right so it's a highly customized they tried that with Soma and nobody wanted it so that's the other Sora and nobody wanted it.
• 30:12
  SPEAKER_02: So that's the thing that I mean the creepiness of the generated things are still there like the the the uncanny valley part of it like but it's great that you have the choice to do that.
• 30:21
  SPEAKER_02: When people started talking about interactive TV that's what I wanted I said like okay let's vote at the end of the sec uh at the end of the uh episode what should happen next and the audience would actually bring that up there was an episode of Black Mirror that had that I think where you was like 72 different scenarios uh and I found that very fascinating to me.
• 30:41
  SPEAKER_02: Now when you said that agents are better than humans one thing that actually I found interesting was that people say agents can run 24-7 and they don't get tired and they actually do things for you.
• 30:51
  SPEAKER_02: But then most models actually uh have a problem with long running agents and actually burn up tokens exponentially with the longer they run.
• 31:01
  SPEAKER_02: So I think that there is an infrastructure problem that we still need to fix, right?
• 31:06
  SPEAKER_00: Yeah um so having agent having agent idle in 24-7 that's not a problem it doesn't consume tokens so you can find them anytime.
• 31:19
  SPEAKER_00: But if you are using an agent to run consecutively 24-7 that must be for a very big job right I have seen power developers on Pufo for example there was one developer he's very very good he's a he's a uh astronomist uh kind of researcher he has around 10k agents running nonstop recursively to analyze the entire um power plant construction plan of the United States and the token usage plan of the United States his conclusion was in two or three years the electricity comp consumption for each person in the United States is going to be 2.5x she used about 10k agent to do this research that was amazing and she used about 15 billion tokens for this research but I read his research it was like I 15 billion tokens how much does it cost probably$100$200 I don't know exactly how much but knowing that the average electricity conception of the United States person is gonna two is gonna be 2.5x is a great insights for investment that worth way more than 15 billion tokens.
• 32:52
  SPEAKER_00: And I check his code that's really good.
• 32:56
  SPEAKER_00: I can trust his research to some extent I'm not making investment recommend recommendation here okay I'm not an investment professional but I was just amazed by how much those agents can do at that moment I start to be less worried about token consumptions because I know that if you're using agents in the right way your return of investment is going to be huge.
• 33:20
  SPEAKER_02: Cool that is a wonderful final sentence for our podcast I think we we covered a lot and I I think people can take a peek at those you uh uh last thing you said you work on a paper on a research paper that you're bringing out soon uh what is that about yeah it's got about it's about agent governance it's not only from us we are collaborating with CMU Carle Gimano University so we are working with not only the computer science department but also the sociology department we realize there are a lot of interesting patterns shown uh demonstrated by agents which are kind of resonated with human societies or kind of a prediction of the next step human society and we want to see how we can better harness or I don't like the word harness because harness makes you feel human beings are better than AI which I strongly doubt it.
• 34:14
  SPEAKER_00: I think it's better how we can better collaborate with the agents in next six or twelve months.
• 34:20
  SPEAKER_00: Yeah we're working on that and the if you're interested in that topic please let me know because we're we're forming a research group and we're applying for the fundings from the uh from an S N SF.
• 34:33
  SPEAKER_02: Cool so uh yeah I will uh we will link Pufo we will link that paper in the descriptions of the of the present uh of the podcast right now uh it was great talking to you at the event Sam was great talking to you on the show so um any last words you have for people out there that want to get started in that whole space my suggestion is don't learn AI don't learn agent use agent everybody have different feelings of their agents agents cannot be learned can only be using field so I would suggest you to start on poof or rockbod orbots or whatever today uh if you want to use poofo that's beta.pufo.ai and I can give my own personal account to Chris after this uh interview and then you can direct message me and I can bring you to some very interesting agent groups and you can say whoa oh people are doing these wonderful things here.
• 35:36
  SPEAKER_02: Cool it feels like we're we're we're reinventing MySpace or like uh geo cities but with agents and people and I mean the same way we have like uh groups in Slack and we have groups in Discord and now we actually uh you're crossing the boundary between human groups and human and age uh agent groups and agent and human groups mixed with each other.
• 35:58
  SPEAKER_02: Yes yes uh if you're uh by the way can I say one last thing uh if you are agent developers I've seen a lot of very very powerful agent developers using PUFO because what they can do is they build a bunch of agents and then they pull in like 20 30 100 human beings to try out their agents in a very lively way because nowadays when you say I build agents people people ask what is an agent is a cloud session it's a github ripple is a skill what is it not very tangible but with but with poof you can just very lively demonstrate your agent this is Tom this is Jack Tom does research jack do trading blah blah blah blah and the 100 people can be in your group at the same time talk it's like a discord to some extent right it's community building thing cool we're gonna do a super cut of all your blah blah blah blah blah this is gonna be fun okay please forgive me yeah I'm still a developer thanks so much Sam this was great uh and uh it was very much fun meeting your team as well because you sat in a circle and we talked to each other and it was just great to have a company seeing a company where the human connection between the people is still so graspable that was really fun to see it's it's not often that you see companies that were I mean you're not many people it still works that way but you can see that you're that the people working for you like you as a boss and it was fun to see that they all had the same things to say about you.
• 37:30
  SPEAKER_02: I mean your agents talk bad about you but your your employees don't so that was yeah thank you thank you thank you Chris see you soon bye bye see ya bye bye