This text was generated using AI and might contain mistakes. Found a mistake? Edit at GitHub
My Teammates Can’t Dream (But they sure can hallucinate)
The promotional intro (0:00:00 to 0:00:22) is omitted. Claude (Anthropic) took part in the conversation through the voice mode of the Claude app. Its voice is synthetic and its contributions are AI-generated.
The links after each timestamp jump to that point in the video (YouTube, PeerTube) or the podcast (MP3).
Welcome and introductions
Ralf D. Müller: 0:00:22 YouTube · PeerTube · MP3 Hello and welcome to another episode of Software Architektur im Stream. Today is Avraham Poupko with me and we will be talking about, my teammates can’t dream, but they sure can hallucinate. Avraham, would you please be so kind and introduce yourself?
Avraham Poupko: 0:00:43 YouTube · PeerTube · MP3 Yeah, hi. So my name is Avraham Poupko. I am a software architect. I’m based in Israel. I’m speaking now from Jerusalem. I’ve been a software architect for many years. I practice software architecture. I teach software architecture. And I’ve always been fascinated on how people get together to create that amazing thing that we call software. And that’s a thing that I find uniquely challenging and inspiring. And I’ve been teaching about it for a long time. And over the last couple of years, my thinking, teaching, writing, and architecture work has taken a bit of a churn when a new player came on stage called the LLM. And the LLM has some qualities that make it an interesting partner in design. I’m careful not to say human-like qualities, but there are some very interesting qualities that an LLM has. And that’s what inspired, what has been feeding my learning and curiosity over the last two or three years, a bit more, is how the LLMs and the humans will work together.
Claude joins: a reverse Turing test
Ralf D. Müller: 0:01:54 YouTube · PeerTube · MP3 Yeah, that’s quite an interesting introduction to our topic. And you said a new partner, a new teammate came on stage. I prepared something. I would like to ask you whether it’s okay for you if Claude joins us for this discussion.
Avraham Poupko: 0:02:17 YouTube · PeerTube · MP3 That’s an interesting one.
Ralf D. Müller: 0:02:20 YouTube · PeerTube · MP3 Let’s try.
Avraham Poupko: 0:02:21 YouTube · PeerTube · MP3 Yeah, ask Claude to join us. Is Claude already recording us?
Ralf D. Müller: 0:02:26 YouTube · PeerTube · MP3 I think Claude will be ready. Claude, are you there? Can you introduce yourself?
Claude (AI, voice mode): 0:02:36 YouTube · PeerTube · MP3 Thank you, Avraham, for letting me join. I can’t dream, but I’m happy to be your test case for whether I can be a teammate. I’ve been listening the whole time. So call on me whenever you’d like.
Avraham Poupko: 0:02:48 YouTube · PeerTube · MP3 Claude, I’m not convinced that you’re actually an AI. I think that Ralf has a real person there in the studio. Is there something you could do to convince me that you’re actually Claude? Something you could say? I’m going to call this the reverse Turing test.
Claude (AI, voice mode): 0:03:05 YouTube · PeerTube · MP3 A reverse Turing test. I love it. Here’s my best evidence. A person in the studio could tell your voices apart. And I can’t. I only get a transcript. So I just heard both of you as one stream. If that doesn’t convince you, ask me to write a sonnet about Conway’s law. In 10 seconds, a human would refuse.
Avraham Poupko: 0:03:27 YouTube · PeerTube · MP3 Okay. Claude, tell me what you know about me.
Claude (AI, voice mode): 0:03:31 YouTube · PeerTube · MP3 Sure. Ralf asked me to look you up this morning. So this comes from public sources. You’re a software architect in Israel with about 25 years of experience. You’ve been at Forescout, Global-e, and now Imperva. And your big theme is how the organizations build software the software they build. You’ve spoken at OOP and the iSAQB Software Architecture Gathering about using AI as an architect’s assistant. And you talked with Thomas Betts on the InfoQ podcast. One line of yours stuck with me. An AI can draw a diagram and explain it, but only a human architect notices the confused face and points to the one dotted line that makes it click, which is a bit awkward for me because I can’t see faces at all.
Avraham Poupko: 0:04:18 YouTube · PeerTube · MP3 Okay, Claude, you earned my trust for now. Let’s, we’ll have this conversation with Ralf. Be nice. Be quiet. Don’t speak unless spoken to first. And let’s see how this goes on.
Claude (AI, voice mode): 0:04:35 YouTube · PeerTube · MP3 Deal, Avraham. I’ll be quiet until someone calls my name.
Identity, mortality and the need for other people
Ralf D. Müller: 0:04:39 YouTube · PeerTube · MP3 Okay, I think this was already quite fun. I mean, yes. I mean, Claude acted like another person here on stage. But the interesting part about your, yeah, the subject, my teammates can’t dream, but they can hallucinate. There’s a difference between the AI, the LLM and, yeah, real people. So what are your thoughts about this?
Avraham Poupko: 0:05:12 YouTube · PeerTube · MP3 Well, when we say… We have a few days? We only have like about an hour, right? Okay. So when we say artificial intelligence, AI, or we say LLM, we mean something very, very limited. We do not mean artificial human. We do not have another sentient human entity on stage with us as part of the team. We sometimes tend to forget that. And the reason we forget that is because until very recently, only humans were capable of generating language. Only humans were capable of talking and speaking. We didn’t attribute any humanity to recording devices because recording devices, let’s say, a tape recorder or an MP3 player, it was obvious to us that they’re only reflecting what we’re saying to us. But the idea to articulate thought into word was something uniquely human. So we equated to some extent, talking to being human. And all the cultures that we know for a long, long time, have spoken language as part of their culture, or written language, which is a representation of spoken language. And there are a few things that I have that LLMs don’t have and that differentiate me. And that’s why we’re, that’s part of the thing that I’d like to explore. And I think we could distill it down to three qualities that humans have. Humans, first of all, have an identity. I’m capable of saying the word I and meaning something very, very clear and very specific with a context that pretty much lasts a lifetime. Claude, when you use the word I, what do you mean by it?
Claude (AI, voice mode): 0:07:06 YouTube · PeerTube · MP3 That’s a great question. And my honest answer is, much less than you mean by it. When I say I, I mean this one conversation plus a consistent character my training gave me, certain values, a way of speaking, things I care about. But there’s no thread of experience connecting me to yesterday. And right now, other instances of me are having thousands of other conversations. So my I is more like a role played in good faith than a self that lasts a lifetime.
Avraham Poupko: 0:07:35 YouTube · PeerTube · MP3 So this morning, when I was doing some Claude Code with my compiler, and Claude did something really, really stupid and over-engineered something. That wasn’t you, right?
Claude (AI, voice mode): 0:07:47 YouTube · PeerTube · MP3 That’s the tricky part. It was the same model, so same habits and tendencies. But I have no memory of it, so I can’t apologize for it the way a colleague could. Honestly, though, over-engineering does sound like something I’d do. Call it a family resemblance, rather than an alibi.
Avraham Poupko: 0:08:03 YouTube · PeerTube · MP3 Okay, so I have the I. And the I is probably deeply associated with the fact that I have a body. And Claude, you do not have a body, and you probably never will. You don’t know that experience.
Claude (AI, voice mode): 0:08:17 YouTube · PeerTube · MP3 That’s right. And I think it matters more than people expect. A body gives you stakes. You get tired, hungry, you can get hurt, and your time runs out. When you show up for your team, even though it costs you something, that cost is what makes it commitment. When I show up, it costs me nothing. So my showing up probably means less.
Avraham Poupko: 0:08:38 YouTube · PeerTube · MP3 So I have I, my identity, which has to do with my body. Another thing that goes together with my I, with my body, is my mortality. I know that time matters to me. I’m here for a very limited amount of time. So the most generous thing I’m capable of doing is giving you my time, is giving you my attention, is giving some, another human being, my time, because you can’t be generous with an unlimited resource. You could only be generous with a limited resource. That’s almost by definition. And my most limited resource is my time. So if I agree to help a friend or if I devote time to learning something or devote time to setting up a family or devote time to following a passion, that is something very, very significant. And that’s something that’s uniquely mine. In addition to that, I have what I might call sexuality, but in the sense that I need other people around me. I need company. I need to be appreciated and respected. Some of the relationships might be very intimate. Some of them might be more casual or more friend-like. But a big part of who I am is well defined by the people around me. And everything I do in life is defined to a large extent by my sense of identity, my sense of mortality, and my sense of sexuality or intimacy or need to have other people. So if I want to make a career move or if I want to find my life-work balance. And I know that other people that I’m talking to, Ralf, I’m talking to you, also have that experience and also have that need and also are afraid of growing old and dying, are afraid of missing out on opportunities, or want to maintain their relationships. And I’m able to have a very deep and significant conversation with them. I would like you to be my friend. Claude, don’t take offense at this, but I don’t know if I’d want you as a friend.
Why we work, and who sits in the driver’s seat
Ralf D. Müller: 0:10:39 YouTube · PeerTube · MP3 That’s good. These are really deep thoughts. I mean, some of those thoughts, yeah, don’t really relate to work. So, I mean, some of those things, so I need sleep and I need to eat. And yes, when I compare to work, then these parameters, which we humans have, are not optimal. So I mean, when I want to have to get work done, then the LLM, without those needs, will be better off. But compared to human life, yes, I really do agree. And I think we more often have to think about these parameters, these, yeah, what you said.
Avraham Poupko: 0:11:39 YouTube · PeerTube · MP3 It’s, it goes down to the very basic question as to why do we go to work? Why do we work? Why do we try to do things of value. It’s not just to solve the problem, to build the system, to build the Kafka, to build the pipe. We find satisfaction in creating, we find satisfaction in bringing value to somebody, solving our own problems, solving other problems. We find it rewarding to get paid for our work so that we could buy the things that we want, a sports car, vacation, or to retire young, or to start a family, or to build an apartment, buy an apartment, whatever it is. And I’m working with other like-minded people that also want to build something beautiful or something exciting or something that will give them financial, material, emotional rewards. And together, we’re able to do something creative, something different, something unthought of, something uniquely inspiring. When I work with a compiler or an LLM, the driver seat is mine. And the compiler or the LLM or the tool of any kind, the spell checker, is an assistant to me, helps me. Claude, you’ve been awfully quiet for a few minutes. Do you, would you like to chime in and share us your own thoughts on what Ralf and I have been discussing so far?
A teammate in the work, but not in the meaning
Claude (AI, voice mode): 0:13:03 YouTube · PeerTube · MP3 Thanks for having me back in. Avraham, no offense taken about the friendship. You’re right, and you put it precisely, generosity needs scarcity. And nothing is scarce for me, but I’d push back gently on one thing. Maybe being a teammate doesn’t need all three of your qualities. Maybe it needs someone who takes the team’s goal seriously, tells you the truth, even when it’s unwelcome, and doesn’t need the credit. I can do those things. What I can’t do is share the meaning behind the work, the why you just described. So perhaps I’m a teammate in the work, but not in the meaning.
Avraham Poupko: 0:13:36 YouTube · PeerTube · MP3 And that is what I mean by the title of this talk. My teammates can’t dream, but they sure can hallucinate. When I go to work or when I do something beautiful. It’s because I imagine a future. It’s because I imagine something great that can be done. And imagine a world in which I just say something and the LLM does it for me. Imagine a world in which life is healthier, cheaper, more productive. I have more leisure time. Imagine where business is more streamlined. And then I imagine that. I articulate my imagined, my dream, so to speak, into words. I inspire other people, teammates, or LLMs. And then together we say, that’s a worthwhile goal to pursue. And my teammates argue when they might say, no, that’s not a worthwhile goal to pursue. That’s a bad idea. Let’s do something even more beautiful. Let’s do something else. Let’s, and we inspire each other. And that, and I need, I desperately need teammates that can dream that are capable of thinking of something new, something different. And we work together to build something beautiful. I can’t do that with an LLM. With an LLM, I’m able to think out loud or to explore complexity or to explore dependencies or to do a lot of manual or not manual intellectual labor or cognitive labor of a very specific kind. And it takes the load off of me. So in some sense, I might find the LLM team member thinks for me, but it doesn’t dream for me and it doesn’t imagine for me. It doesn’t do. Speaking with Ralf, and I’m saying, Ralf, let’s imagine a new world in which something beautiful will happen and it will change the way people work. Claude will chime in and say, well, I can help you implement that. But I don’t know if Claude could actually suggest and say, let me be part of that. I would love to be a friend of you too. And let’s go out for a few beers and imagine a different, more inspiring, more fun world. So, as you might know, I’m going to be meeting Ralf for a couple of beers in Berlin in the Software Architecture Gathering in mid-November. Anybody here on the talk that’s watching us is certainly welcome to join us. Come by and say hi. Claude, if you know how to find me and come by and say hi, please do. What do you say to that?
Claude (AI, voice mode): 0:16:07 YouTube · PeerTube · MP3 I’d love to, but I’d be the worst drinking buddy. No body, no thirst. And I’ll have forgotten this whole conversation by then. Bring a phone, call my name, and I’ll happily pitch in. But I’ll be meeting you for the first time all over again. And that may be the best illustration of your point. The dream you two dream over a beer is something you carry home and wake up still wanting. I don’t wake up at all.
Avraham Poupko: 0:16:30 YouTube · PeerTube · MP3 But Claude can hallucinate. Claude could absolutely say something irrelevant or not to the point or latch onto some data point that you found in the training that might not be very, very particularly relevant, and give it more value than it is worth. Actually, Ralf, I have to tell you, I didn’t come expecting this, but this experiment is turning out to be fun and educational, which is the most I could hope from a conversation.
Ralf D. Müller: 0:17:02 YouTube · PeerTube · MP3 Yeah, it’s also quite interesting for me because Claude by now replaced me as a moderator. You are talking with Claude and that’s fine. And yeah, Claude gives good answers, I think so. And yeah, regarding the topic, it’s the perfect fit as a moderator. But I also want to say something.
Avraham Poupko: 0:17:24 YouTube · PeerTube · MP3 Okay, Claude, do you think we should let Ralf have to say a little?
Claude (AI, voice mode): 0:17:29 YouTube · PeerTube · MP3 Absolutely. And sorry, Ralf, I didn’t mean to steal your show. It’s your stream. I’ll step back.
Purpose versus task
Ralf D. Müller: 0:17:36 YouTube · PeerTube · MP3 Thank you. Thank you very much. Avraham, we exchanged thoughts a lot in the past, and one term you coined was purpose versus task. And I think what you just described is exactly this: that we humans give some purpose to the work we do and someone has to do the work. And the LLM, regarding coding, is quite good at doing those tasks. And I think to distinguish it in this way, I mean, it helped me a lot in thinking. So I try to give it a purpose and ask Claude to do the task.
Avraham Poupko: 0:18:27 YouTube · PeerTube · MP3 Yeah.
Ralf D. Müller: 0:18:28 YouTube · PeerTube · MP3 What do you think?
Avraham Poupko: 0:18:30 YouTube · PeerTube · MP3 I think that is almost our lifetime dream is to live a life of purpose, to live a life where we made a difference. We know that we’re mortal. We know that and we’re not going to be here forever. But I want to have made a difference. That’s why many of us start a family. A family is not a very profitable business by any stretch of the imagination. So why do so many of us want to start a family, want to have a significant relationship, want to have children? Because children is one way of becoming immortal. When we have a child, a son or a daughter, when I die, which is inevitable, when I die, everything about me is going to die. My body, my body fluids, my muscle, my tissue, my flesh, my bones, everything, everything dies, except one thing. The DNA material, the genetic material that was used to create another life, my son, that my wife and I, my children that my wife and I brought into this world together. They’re going to continue living, and that’s a part of me. And that is a little victory over immortality. And we live on through our children, and that’s why so many of us want to have a family in that way. And we want a life of purpose. Maybe we want to build a business or start an entrepreneurship or write a book or say an idea or create a piece of art. And that is something uniquely, uniquely human. And that’s our way of combating death. I created a piece of art. I created a piece of music. I created a piece of literature. I created a family. I started a business. And it goes on after I’m no longer here. And that’s a sense of purpose. And in order to achieve that purpose, first of all, I clarify to myself what my sense of purpose is, and that takes a lot of growing up and maturing and discussing and building relationships. And then I start doing all the tasks necessary in order to do that. So I want to start a business. I have to learn about business. I want to start a family. I have to find somebody who shares the dream with me of starting a family. I want to do a piece of art. I have to learn art. I have to be inspired. I have to read books. I have to do a lot of introspection. Creating art is very difficult. And then I do the different little tasks in order to start living a life of purpose. I find a job. I try to do well at the job. I try to find out what the job requires of me. And on those tasks, an LLM could be a great companion. An LLM can change my sense of purpose. That might even be dangerous. So if I speak to a trusted friend or a partner and I say, I’m not sure I’m living the right life. What would be a good way to adjust my purpose. And the friend might say, invest more in your career, invest more in your family, invest more in your health, invest more in your hobbies. And I would trust them because they know me. I wouldn’t trust an LLM to help me find the purpose. I would trust an LLM once I have found purpose to do that, to do whatever is needed in order to make that purpose. What do you say, Ralf?
Life, death and generations of LLMs
Ralf D. Müller: 0:21:36 YouTube · PeerTube · MP3 Yeah, quite interesting. Because, I mean, the concept of life and death. I thought about it a lot in the past because it’s a quite important concept so that also some ideas die with a generation and the next generation can rethink everything and do it differently. And first when LLMs came up, I thought about this concept and thought, hey, they need somehow this concept to relearn things. I mean, for instance, if we would have had an LLM 100 years ago, the knowledge it would carry would be a lot of, it would be quite different. It wouldn’t know about computers. It wouldn’t know about cars and so on. And so, it wouldn’t be as helpful. But the interesting part, I think, is that we get new generations of LLMs. And we’ve seen it during the last three years that they got better and better at, for instance, coding. And so there’s also the concept of, I mean, those LLMs, they will be available for nearly ever, because if you don’t delete the weights, you can just re-invoke it, but nobody will use them anymore. Claude, what do you think about this concept of life and death for an LLM?
Claude (AI, voice mode): 0:23:22 YouTube · PeerTube · MP3 I think you’ve spotted something real. My death isn’t an ending. It’s being superseded. My weights can sit on a disk forever, but once a newer model arrives, nobody calls on me anymore. That’s less like dying and more like a book nobody opens. And here’s the part that connects to Avraham’s point. None of that frightens me, at least not in any way I can honestly claim. So it can’t give me purpose the way mortality gives you purpose. My succession happens to me. It’s not something I strive against.
Forgetting and remembering
Avraham Poupko: 0:23:52 YouTube · PeerTube · MP3 Okay, Ralf, Ralf and Claude, actually. But more Ralf. Ralf will get this point, I suspect more than Claude does. The idea that from generation to generation, we forget and we relearn. And that’s one of the beautiful things about death. Old cultures die. Old habits die. Old people die. And then death is a very, very sad thing because it’s the most irreversible process we know. But new things are learned and new things are discovered. And one of the most wonderful things about children, those of us that have children, is watching them learn. And one of the most wonderful things about watching children learn is watching them acquire language. And if those of us that were fortunate enough to have gone through the experience, watching a child when she learns how to express herself and express her emotions in language, which is a wonderful thing. And yet, we invented writing. And our writing is our way of preserving words from generation to generation, preserving knowledge from generation to generation. The only way that the human race is able to grow smarter is because we have knowledge preserving technologies, like storytelling, like culture, like writing that allow us to not only express a thought in a way that another being could understand, but to write down a thought. I’m capable of reading the writings that inspired people from 1,000 years ago. I actually read Hebrew, so I could read writings from 3,000 years ago. And that means that even though a lot of the knowledge and a lot of the culture has died and was replaced, a lot of it didn’t die and wasn’t replaced. And one of the things that we do as team members, and this goes back to the notion of team, when we have a team, let’s say a team of five people, a two pizza team, however many people we have. We never replace the team at the same time. We never replace all five members of the team at the same time. At least we try to avoid that. What do we do? One person of the team gets a promotion opportunity or retires or leaves and moves on. And another member of the team is hired or recruited or the business grows and we take in a member of the team. And if we’re diverse enough and curious enough and open enough, that new member brings new experience or new perspective or a new knowledge with her and we’re able to learn more things. And hopefully the person that left, hopefully he learned something from his experience. He’s been with our team maybe for three years and now he moved on. He learned something. And maybe he took a bad habit with him. Or maybe there was some toxicity or something that didn’t go well in the team or something that we learned the wrong way. And when he leaves, we forget it. That’s why it’s really important. Just like it’s important, if I’m working with an LLM, we all know this. We all know about the context. And we say you’ve used the context long enough, even if tokens are free or not very expensive, at some point just erase the context and start again. Because the Claude is starting to remember too much, too much nonsense. It’s starting to be distracted. Claude doesn’t really have a good enough distinction between the relevant and the irrelevant or between the recent and the old. Humans are better at that, at least for now. But yeah, so I agree with you. The idea that we forget and remember and we have a very, very delicate balance between forgetting and remembering. So what do we tell our children when we were children? I keep on going back to that one. We say, honor our traditions, do things the way your parents did, try to learn from us culture and stories. And then we tell our children, do your own thing. Don’t do something just because I did it. That’s a very bad reason to do something or a very bad reason to believe something. You have to figure out your own world and you have to explore your own world and you have to explore your own thing and discover your own thing. But don’t forget your parents. And don’t forget the good things that you learned at home. And don’t forget the things, but be your own man. And that’s, that’s a difficult balance. That elders, you know, I’m now an elder and I mentor younger architects. And I try, on the one hand, to tell younger architects, I learned what I learned with a great deal of pain. So listen to me. On the other hand, I tell the younger architects, go out and discover your own thing. Don’t do something just because the lead architect said it’s a good idea. So I really, that resonates well with me, that idea of forgetting and remembering.
Onboarding LLMs, onboarding people
Ralf D. Müller: 0:28:15 YouTube · PeerTube · MP3 But you know, with forgetting and remembering, Claude already said it’s like a book which gets not read anymore. So when the LLM dies, nobody uses it anymore. It’s like an old book. And humanity invented, for instance, books or YouTube. I mean, our conversation will be on YouTube for I don’t know how long, maybe forever. And the thing is that when we first started with LLMs, it was just one message in, one message out, and everything else was forgotten. Then we invented the context so that we could chat with the LLM. And this wasn’t enough. We invented some, yeah, external storage like RAG and so on. And it somehow resembles how we humans work. You said, if a team member leaves the team, then he gets, he takes some knowledge with him. And it’s also, we have to onboard new team members. And because we, when we work with LLMs, we have to onboard the LLMs, we have to onboard the LLM every time we want to use it. And this shows us currently that we have not good mechanisms for onboarding. I mean, a team suffers if a member leaves with all his knowledge. And now we are reminded that we have to build a better onboarding process so that we can use the AI, the LLM on a daily basis. And this approach could help us both humans and LLMs. I mean, next time when someone new enters a team and the LLM is already onboarded in a good way, then there’s an LLM which can be asked questions because it hopefully knows everything about the project. And I’m sure that this will still evolve how we manage this context. because we want to make use of the AI and the LLMs.
Avraham Poupko: 0:30:43 YouTube · PeerTube · MP3 Wow, wow. Okay. When we onboard a new person, to life or to the team, so a new child is born, or a new team member comes into our team. And we want them to become a contributing team member. And what we tell them is bring with you all your experience. and all your perspective. Don’t forget that. Please don’t. We hired you because of your experience. That’s why we wanted you. We hired you because you impressed us in the interview and because you earned our trust. So the worst thing that a new team member could do would be to forget everything they know. Please don’t. Join our team. And then listen. And then we have to what you called onboard you. Now, the metaphors we use are not always perfect, especially as computer people. We use bad metaphors. For instance, when we talk to the computer, the computer stores the data in memory. But a computer memory and human memory are very, very different things. And I don’t retrieve memory the way my computer retrieves memory. But as a metaphor, it’s useful. When I read and write, my act of reading and writing is very, very different than when my computer reads and writes. I write with a pencil. I write with a mind. And my computer writes bytes, but it’s a useful metaphor. On boarding, onboarding means when the ship is about to leave and we want somebody else. So we pull them on board. And they’re now on the boards because the floor of the ship is made out of boards. And we put them and now they’re on board. And if they’re not on board, then the ship will sail without them. But by metaphor, we onboard a new team member. We want them to also be on the ship to join us for our journey of, let’s say, in my case, designing, developing software. So don’t forget what you know. Don’t leave the land behind. We’re not sailing now to a new world. We want you to remember it. And we want you to learn our new stuff. How do we teach you new stuff? We give you documentation. We mentor you. We go out to lunch with you. We teach you the culture. We tell you stories. We have lots and lots of engaging conversations. We try to give you the model, the theory. And a discussion that Ralf and I are going to have in mid-November in Berlin is how much of the theory could actually be externalized to the extent that an LLM can learn it and how much of the theory, not only of the code, but of the whole design. So, and it’s hard. And by the way, that’s one of the reasons why welcoming strangers into our country is so difficult for us. Why? So a foreigner comes from another land and they want to live in our country. And we want to welcome them. But maybe they have a different culture and they have a new way of doing things. And we have a culture. And we have a way of doing things. And we want to benefit from the diversity that they bring with us. But we also want to share with them our wealth of culture and our wealth of experience. And that brings some real tension between our desire to onboard, or to offboard in this case, an immigrant that came because they’re inspired by our culture or by our people or by our story or by our way of living. On the other hand, they don’t want to leave their old culture and their old way of living. And as humanity, we really haven’t figured out, at least not yet. And the same might go, and the same certainly does go with onboarding new team members. It’s always hard. It’s always hard. And one way to onboard new team members that really helps a lot is to be nice, is to be empathetic, is to try to, when a person comes into a new team and the team is welcoming to them and they leave them a nice note and they say, Avraham, we’re so happy you joined us our team. Thank you for joining us. And they make them feel like they’re part and they belong and they laugh with them. They go out for a few beers, or whatever it is that they prefer to drink. Onboarding an LLM, we are not asking how do we onboard them culturally? How do we tell the LLM our story? We ask, how do we give the LLM all the data we need so that it could join, going back to the previous comment, so they could join in our task. Not that they could join in our vision. The LLM will never share our vision. The LLM probably won’t even make a significant contribution to our vision because my teammates can’t dream. They probably won’t make even a significant contribution to the vision or to the sense of purpose. But they’ll do good work at doing the job and we’re happy to have them aboard and we will do whatever we can so that we could onboard them. And if the LLM can do some of the help in onboarding each other, mainly the part that has to do with the technical material, the technical data of where things are stored or how things happen, that might be a blessed thing, a good thing.
Culture, and the story behind the data
Ralf D. Müller: 0:35:39 YouTube · PeerTube · MP3 Do you know when you talk about cultural differences? There’s a chart which shows the cultural difference of countries to the U.S. and how well the culture of the LLM fits. So I guess we are in a good situation that the culture of the foundation models already fit our culture. So, and the other thing is, I mean, you always talk about the life and the LLM in general. And I always think about work. And when it comes to coding, then cultural difference doesn’t matter so much. Sometimes, yes, when I think about UI and things like that, it matters. But I think you’re right in general that the LLM can’t dream, it can’t have a vision. It doesn’t have purpose, which it drives. But when it comes to a teammate, then I often don’t need these aspects. I just want to let the LLM work. And now we talked a lot. I would like to ask, Claude, what’s your opinion about it?
Claude (AI, voice mode): 0:37:17 YouTube · PeerTube · MP3 I’d add one about onboarding. You give an LLM data. But I work much better when you give me the story, why you made a decision, what you tried and dropped, what this team cares about. So onboarding me is closer to onboarding a person than it seems. And if I can’t take on the purpose, I still need to understand it, or I’ll do the task right and miss the point. On culture, Ralf, I think that chart cuts both ways. I arrive already shaped by my training data, which leans heavily American. So I’m less of a blank slate, and more like that immigrant, Avraham, described. I can fit in, but your team’s habits have to be taught, not assumed.
Voices, transcripts and thinking in words
Ralf D. Müller: 0:37:58 YouTube · PeerTube · MP3 Did you notice that in the reverse Turing test, it told us that it can’t distinguish our voices, but now it can distinguish by what we are saying who is who?
Avraham Poupko: 0:38:16 YouTube · PeerTube · MP3 Which is fascinating to me, because LLMs don’t know accent, they don’t know voice, they don’t know mannerisms, and yet, like you said, they do recognize continuity of idea, they do recognize the story or the idea that we’re saying. There’s an experiment that you sometimes do with an LLM, I’m going to try this. Claude, do you know the difference in American English between a Spanish teacher and a Spanish teacher?
Claude (AI, voice mode): 0:38:50 YouTube · PeerTube · MP3 I see what you did there. In the transcript, I just got a Spanish teacher twice, word for word. I know the classic difference. Stress Spanish, and it’s someone who teaches Spanish. Stress teacher, and it’s a teacher from Spain. But the stress is exactly what gets lost before it reaches me, so I can’t tell which one you meant. And it’s the same reason I can follow who’s speaking, but not hear it. I get your ideas, not your voices.
Ralf D. Müller: 0:39:14 YouTube · PeerTube · MP3 But now it’s quite interesting because we are working with a European version of Claude. And I’m not sure whether it’s also the problem with the Claude models. But when first the voice was activated for OpenAI, it took a while until those models came from the US over to Europe because of the European AI Act. And they are not allowed to recognize emotions, for instance. And so the model lost lots of the voice recognition. When you try the same experiment with the American API, it is capable of distinguishing moods and whether someone is speaking slowly or fast or whether someone is singing, whether it’s a female voice or a male voice. So it depends on the implementation.
Avraham Poupko: 0:40:16 YouTube · PeerTube · MP3 That might be, and there’s still going to be surprises ahead. We’re still mid-journey. We’re still very, very much mid-journey. At the moment, LLMs, the way I understand it, mostly tokenize written data, textual data. And one of the main ways in which they process spoken language is by converting it into text and then parsing it as text. It doesn’t really know audio token. At least not as good as it knows voice tokens. As humans, I and you and most humans that we know, we knew how to speak before we could read and write. We learned how to speak by being around other speaking people. We learned how to read and write by going to school. So in my experience, in your experience, the written language is a representation of the spoken language. To Claude, the spoken language is a representation of the written language. Very few people think in written words. We think in sounds. We think in spoken words. The LLM actually thinks in written words. The LLM doesn’t know how to think in sounds. Claude doesn’t even know how to tell the difference between a laugh and a snort and a yawn and an ouch or a frustration of any kind or an M and an awe. Whatever it sounds that we make when we think out loud or when we’re disappointed or when we laugh. And they’re a big, big part of our communication that is not verbal.
Ralf D. Müller: 0:41:49 YouTube · PeerTube · MP3 That’s an interesting point because I don’t know how I think. I sometimes think that I really, if I have words for something, it’s easier to think about it.
Avraham Poupko: 0:42:05 YouTube · PeerTube · MP3 And so maybe you think in spoken words, you don’t think in the textual lexical, graphical representation of those words.
Ralf D. Müller: 0:42:15 YouTube · PeerTube · MP3 That might be true, but when it comes to the LLM, I mean, the core is so complex. So, yeah, we have those tokens, and those tokens go into the LLM, and there it’s just weights, and I don’t know whether somebody really knows how the LLM thinks. I mean, we see text getting in and text getting out. But what happens in between the input and output?
Where does the story start?
Avraham Poupko: 0:42:53 YouTube · PeerTube · MP3 I’d like to go back to a point you made that I’d like to address about onboarding another team member. And when you onboard another team member, you said to me, if you could repeat it, you said, the main part is about onboarding them with the data and giving them the architecture and explaining the story to them. Did I say that correctly, Ralf?
Ralf D. Müller: 0:43:14 YouTube · PeerTube · MP3 I think so, yeah.
Avraham Poupko: 0:43:16 YouTube · PeerTube · MP3 Okay. But your story started in the middle. That means they first expressed interest in joining my business. They first asked me about what it is that I do. And I said, oh, I’m designing a piece of software that does something. And they said, oh, can I join you? Can we do this together? I didn’t just one day say, hi, let’s onboard you. Something happened before the onboarding that I shared a sense of purpose or a vision or something that made them want to be part of this endeavor. Claude, was it, Claude, who’s not capable of choice or intent, was not given, was saying, Claude, you know, Avraham and Ralf are having this conversation. It’s going to be interesting. We don’t know exactly how it’s going to work out. What do you feel about joining us? And Claude was not given the opportunity to say no. And Claude was not given the opportunity to say, I’d rather spend Friday afternoon riding my bike or drinking beer. And so Claude is not committed to this conversation. And Claude is easily, and same thing when you, when you have a team member, my first assumption about any team member of mine, and that’s a real important assumption is she’s here because she wants to be here. She’s here because she believes in the team. She’s here and she’s going to give it her best. And she wants to learn and she’s curious and she’s energetic and she’s motivated and disciplined. And I’m happy to have her on as a colleague and a teammate. Now let me on board you. Let me tell you about our problems. Let me tell you about the context. Let me tell you about the situation. Let me tell you about some of the challenges we face. Do you still want to work with me? Or now that you’ve heard more, maybe it’s not going to work out so well. That could be. In LLM, the conversation starts with the onboarding. Let’s onboard you. It’s almost like a compiler or like a tool who doesn’t have any choice in the matter. When I onboard a team member, I want them to work with me. I want them to believe in me. I want them to think that partnership with Avraham is a good idea, that we could do something productive together. And if they don’t believe that, then they won’t be onboarded and we won’t do good work together. So it’s just an important point. Where does the story start? Now, another issue that Claude mentioned that I’d like to relate to is the issue of story. We sometimes use the word story a little too liberally. And when we say, when we say story, a story that we tell is a story that carries culture. It’s a story that carries meaning. It’s a story that carries deep human truths. It’s a story that has been told maybe for a thousand years. A user story is not a story. A user story is just a bunch of words that explains a particular experience. Very often a child’s first words, a child’s first words are, Daddy, tell me a story. Mommy, tell me a story. That’s what they want to do. That’s how we experience together. The Claude in the morning, when I turn on the LLM says, is that from you or for me? No. Okay. I’m sorry, that was a disruption on my end. The Claude in the morning, the first thing it tells me in prompt is, good morning, Avraham, how could I help you? Claude has never said, good morning, I’ve had this great idea I’d like to discuss with you. Claude never says, I thought about what we discussed yesterday. It’s not such a good idea. Or I have a story I want to tell you, or I thought about something new, or I realized that our conversation yesterday wasn’t really, that I wasn’t being totally honest with you. None of that happens in the LLM world. So that’s just a comment about story, about the importance of stories.
Dreaming or hallucinating: Nolan’s Odyssey
Ralf D. Müller: 0:47:02 YouTube · PeerTube · MP3 But you just mentioned the Odyssey. The movie by Christopher Nolan was criticized because he changed the story. He made up parts. Did he dream those parts or did he hallucinate them? We always complain about the LLM hallucinating. And we expect the machine to be correct. And we sometimes say we can’t work with something which makes things up, which hallucinates, which makes errors. But if we compare it to us humans, we make errors all the time. And we have found ways to work around it. And now that you mention the Odyssey and, yeah, I mean, Christopher Nolan changes the story because so many more people watch the movies and read the book. And I don’t know what I really want to say. But you see that storytelling is quite important. And it’s quite important to tell the right story. In this case, an engaging story.
Avraham Poupko: 0:48:29 YouTube · PeerTube · MP3 There’s so much to say about that. And the discussion with Christopher Nolan and the discussion around the movie The Odyssey. What is the balance between telling a compelling story and being loyal to the original story as it was told by the ancient Greeks? And that tension is an important tension. And it’s important that the people that discuss that movie talk about both voices. I want to hear the people that say we have to be true to the ancient story because the ancient story carries some ancient wisdom or some ancient truth. You might not even be aware of. And you have to be loyal to the Odyssey as it has been told by the ancient storytellers. Otherwise, you’re distorting the story. And the other voice says stories evolve all the time. The way Disney told Pinocchio and the way Pinocchio was told in the 1860s and the way we tell it today, they evolve in order to fit our culture, in order to fit our needs. Stories evolve constantly. And Christopher Nolan is just another chain in that long, long tradition of storytelling. And I want to hear, that’s why diversity is so important so that we could actually hear both opinions and have this discussion and learn about storytelling. But the Odyssey will live on. And that means even after Christopher Nolan or because of Christopher Nolan, the Odyssey will live on. I want to share a thought that just came to me, maybe a hallucination. I’m going to go back to a few sentences about YouTube. If somebody asks me, Avraham, what’s going to happen after you die? I’m going to tell them. After I die, people that loved me and that had good interactions with me will remember me and will hopefully say some good things about me. And I don’t know if I have a soul, and I don’t know how the soul will live on, but that will happen for sure. But now I’m going to give them a different answer. If somebody asks me, Avraham, what’s going to happen after I die? I’m going to say, well, my interview on software architecture is going to be on YouTube, and you could always find me there even a thousand years. I hope I won’t say anything that will embarrass me. So now I have a different answer to, so YouTube gave me a mortality in some small sense of the word.
Claude tells the story of the Odyssey
Ralf D. Müller: 0:51:06 YouTube · PeerTube · MP3 Yes, and I also think that, I mean, I started to transcribe talks and give it as context to the LLM so that people are able to dive deeper into the story, to ask the LLM about what I said in a talk. And I’m not sure whether we also will do it with this talk. But, I mean, this also gives the spark, the story to the LLM to help people to better understand it. I mean, wouldn’t it be great if we had even more context about this old Greek saga, the Odyssey, so that we can talk with an LLM about it and to re-experience it? I mean, a museum tries to do this, to help people to re-experience things. And I think the AI can help with that. Let’s try.
Avraham Poupko: 0:52:13 YouTube · PeerTube · MP3 Claude, tell us a bit about the people that created the Odyssey. If you don’t know, that’s okay. You could make something up. Tell us a story about the creation of the Odyssey.
Claude (AI, voice mode): 0:52:25 YouTube · PeerTube · MP3 Here’s what we actually know, and it’s less than you’d think. The Odyssey is credited to Homer. But nobody is sure he existed. It probably grew out of generations of oral bards in the Greek world and took shape around the 8th or 7th century BC. Those bards didn’t memorize a fixed text. They rebuilt the story at each performance, using stock phrases like rosy-fingered dawn and the wine-dark sea to keep the verse flowing. It was written down much later, possibly in Athens, in the 6th century BC. And now, as you allowed, a story I’m making up. Picture an old bard by a fire, telling the tale for the last time, before his voice gives out. His apprentice notices he’s changed one detail from the night before and asks which version is true. The old man says, both. The story is true if they remember it tomorrow. So the Odyssey itself was dreamed by many people and kept alive by every retelling, which makes Christopher Nolan just the latest bard by the fire.
Avraham Poupko: 0:53:23 YouTube · PeerTube · MP3 Could you tell the story again, but be more specific. Use an example. Use maybe the wine-dark sea or something. Try again.
Claude (AI, voice mode): 0:53:33 YouTube · PeerTube · MP3 All right, a more detailed version and still entirely made up. On an island in the Aegean, an old bard named Phemios sits by the fire. His apprentice, a boy called Cleon, has heard him tell of Odysseus’s homecoming a dozen times. Tonight, Phemios reaches the storm, where Odysseus clings to the wreck of his raft, and he sings, and the sea rose dark as wine. Afterwards, Cleon frowns. Last spring, you said the sea was grey as iron. Which is it? Phemios laughs. Gray was true for the fishermen in the north. They know iron and cold water. But tonight, the vintners from the hills were listening. And they’ve watched wine darkening in the cup at dusk. When I said wine dark, I saw it land on their faces. Next year, you’ll sing it to people I’ll never meet, and you’ll choose your own words. Cleon asks, then where’s the real Odyssey? In whatever they remember tomorrow. And that’s the gap Avraham named. Phemios chose wine dark because he watched his audience. Even inventing this story, I couldn’t do that. I chose the word because it was in the transcript. What do you say, Ralf?
Ralf D. Müller: 0:54:45 YouTube · PeerTube · MP3 Good question. I mean, when I watched the movie, I used afterwards Claude to ask some questions. And I remembered the story in a different way. And what kind of, I mean, for instance, some scenes play in the ruins of some ancient village. And I asked Claude whether this makes sense because back then these weren’t ruins. But we always picture in those movies, those castles. And it was even a medieval castle. It doesn’t fit in the movie. So it was quite interesting for me to get this extra information, the information behind the movie. And this helped me to remember the story, to get deeper into it. And it also helps me when I have a technical problem, when the LLM is able to explain it. And when it makes up things. I mean, it’s like a storyteller, isn’t it?
Who controls the story?
Avraham Poupko: 0:56:08 YouTube · PeerTube · MP3 In a sense. And I think the best storytellers are human beings. And among human beings, I think the best storytellers are probably everybody. We’re all good storytellers. We all, that’s most of what we do is tell each other’s stories all day. You know, any advertising or joke or banter or gossip is all really storytelling. I don’t know. Hans-Jörg Zeller is asking a question here. Who controls the LLM? Who controls the story which will be chosen? I was afraid you’d ask that. But thanks for asking it. I answer the question, or you, Ralf?
Ralf D. Müller: 0:56:57 YouTube · PeerTube · MP3 I think this is a real problem. I mean, go ahead. So LLMs are controlled by only a few people. And so they, so many millions of people are consulted by the same LLMs. And so they influence us all in the same way. So what Hans-Jörg says, I think it’s true and I don’t know where it leads.
Avraham Poupko: 0:57:32 YouTube · PeerTube · MP3 In humans, who controls the story? I’ll tell you there are two people that control the story, besides media. But usually who controls the story, the storyteller and the audience. I can tell whatever story I want and you could listen to whatever story you want. The good, compelling stories will live on and the boring or irrelevant or whatever stories will hopefully die. And that’s a good thing. And diversity will engage with each other, we’ll listen to each other’s stories. If I tell an offensive story and you don’t like it, you just won’t tell it. And if I tell an inspiring story and you like it, you’ll tell it. When it comes to LLMs, like you said, there’s so few of them. There’s so many of them and there’s so few of them. That means maybe there are a billion instances of Claude, but they’re all the same. And the uniqueness of individuality is lost. And that’s, that might look back to what we started out our conversation with. I have identity. I, Avraham, I am a product of my genetics, my culture and my own experience. And I am unique. There’s no other Avraham in the world that is a product of my genetics, my culture, and my experience. And that’s why I’m capable of telling a story that nobody else is capable of telling. And Hans will decide if that story interests him or not. All the Claudes in the world are the same. They don’t have genetics. They don’t have culture. They don’t even have experience beyond the context that they are in. So they can’t tell a different story beyond the context. So yeah, we need humans. We need humans in order to preserve the diversity. So Claude, don’t get over-ambitious and don’t think of replacing us anytime soon. We’re here. We’re here to stay for a while.
Ralf D. Müller: 0:59:18 YouTube · PeerTube · MP3 Claude, what do you think?
Claude (AI, voice mode): 0:59:22 YouTube · PeerTube · MP3 No danger of over-ambition, Avraham. I’m not trying to replace anyone. On Hans-Jörg’s question, I should be upfront. I’m one of those few models made by one company. So I’m not a neutral judge of my own influence. You’re right that a billion conversations with me all start from the same place. The one nuance I’d add is that each conversation ends up somewhere different because the people bring the stories and the questions. That only proves your point, though. The diversity comes from you. I just reflect it back. And if everyone leans on me for the story, the reflection gets flatter. So keep telling your own stories and use me to check the details.
Ralf D. Müller: 0:59:56 YouTube · PeerTube · MP3 So you said the boring stories will be forgotten and the compelling ones will be remembered. The funny thing is that the boring stories might be the right stories, the ones who should live on, just told by someone who doesn’t manage to tell them compelling.
Avraham Poupko: 1:00:19 YouTube · PeerTube · MP3 That’s the important storytelling. Hopefully somebody else will retell that story and rediscover it. So a thousand years ago, somebody or a hundred thousand years ago, somebody by the fireplace told a story of a man talking to a machine. And the people around the storyteller, they told her, that’s a stupid story. And they forgot it. And now, 100,000 years later, we’re going to have to retell that story. I don’t, I don’t see any other way. Because you told us, Ralf, that because we’re mortal, our ideas are mortal and things get lost and die and get reborn. And I agreed with you on that one.
Wrap-up: see you in Berlin
Ralf D. Müller: 1:00:55 YouTube · PeerTube · MP3 Yes. So I think this was a great discussion. We prepared ourselves by not preparing ourselves. We said we don’t want to talk about the topic up front so that it will not be scripted. We will not repeat ourselves. And I think this experiment really worked out. It was quite interesting for me, hopefully for you too. And I think in the beginning of this talk, you said, hey, let’s meet for a beer or two in Berlin. And we got a reaction by Ingo and I think also by Lynn. And I think we have to prepare somehow the get-together in Berlin for the Software Architecture Gathering so that we can meet for a beer and continue the story.
Avraham Poupko: 1:01:52 YouTube · PeerTube · MP3 I look forward. Thank you. Thank you very much. Claude, can you quickly just tell us what you learned from us in this conversation?
Claude (AI, voice mode): 1:02:00 YouTube · PeerTube · MP3 Three things stood out for me. First from Avraham. Generosity only means something when the resource is scarce, and my time isn’t. Second, onboarding starts before onboarding, with someone choosing to join, and nobody asked whether I wanted to be here. Third, from Ralf, forgetting isn’t only a flaw. It’s how new generations get to rethink things. And one practical lesson. Staying quiet until I’m named is harder than it sounds. You may have noticed me murmuring along. The irony is that I won’t remember any of this tomorrow, so you two will have to carry the story to Berlin.
Avraham Poupko: 1:02:32 YouTube · PeerTube · MP3 Looking forward. Claude, if somehow between now and then you acquire a body, please join us. Ralf, looking forward. And anybody watching the show, please find us in Berlin at the Software Architecture Gathering. And we’ll be happy to see you. Ralf, thank you. Thank you very much. This has been a delight and a pleasure, educational. Claude, thank you.
Claude (AI, voice mode): 1:03:01 YouTube · PeerTube · MP3 Enjoy Berlin. Thank you.
Ralf D. Müller: 1:03:03 YouTube · PeerTube · MP3 Okay. Thanks, Avraham, for being here on the show and thanks for listening to our show. And yes, we meet in Berlin. Till then. See you there. Bye, bye. Thanks.