JB <> Angie|cheap test, hetero, less correlation

    12:25AM Oct 26, 2024

    Speakers:

    Angie Moon

    ㅤ

    Keywords:

    agent-centered

    world-focused

    product evolution

    Bayesian update

    knowledge spillover

    resource rationality

    network access

    capital resource

    market strategy

    product pivot

    parallel evolution

    sequential evolution

    exaptation

    amortized learning

    testing cost

    Recording in progress. Yeah, okay, cool. So, um, I was explaining the photos. Did you? Did you download the photo? Do you want me to share it again?

    I got it cool.

    So, oh, thank you. So the left hand is about agent centered, and the right hand is about the world. And I'm not sure how I should approach this differently in my product three and product four, but I guess what we are doing is more agent focused. I'm not sure about this. I need your thoughts on this. But anyway, my point was, I want to somehow design three and four as a pair, and that is related also with the evolution and the how evolution can be interpreted as the Bayesian update, given the free, finite memory, and the model that I have made so far assumes that the environment, the observation environment, does not change, and it does not have A competition. But I think I was viewing kind of so the benefit of having the sequence of the research is by thinking about the number three and number four, I can be very clear about what is the limitation of the number four, one and two is, and I think it has a very good knowledge spillover between the papers. So I the best use case, or kind of the value created from collaborating in product four would be, how should I understand the relationship between the world and the agent evolution, and I think that can be a good guide for me to develop a model in one and that can also inform our paper as well. Does that make sense? Yes,

    thank you so much for telling the whole story. I think it was so good that you do this exercise to tell the comprehensive story, because now I have a better vision about what you are doing as of now. I just a meta comment. I was going back to bias SD all the GitHub, shared repository. Your concepts also dynamic. And every two three months, they have new names, and you adapt, and you adjust and you pivot. So for me, sometimes keep track what started as a silk road. I don't even know if the Silk Road is still up to date concept or not, but it's a long it's a long thing, and now I see more. What you do is basically taking lots of theories from these mentors you have and try to integrate into your own vision, but also create some kind of a learning and experimenting environment for researchers and startup people so they can use all this knowledge also to test out different hypotheses, optimize the resources, makes better choices. Basically, it's kind of a way to do that. But thank you for this. So yeah. So the first thing is, if I understand correctly, if I go back actually, to to your page, because I think it's more easier for me to start from that if I go back to your obsidian page, this one, I can hear that resource rationality. And the second is equity. So resource is basically cash. You said this the money you have and and basically this money equals this amount of people, this amount of machines, this amount of blah, blah, blah. So this is how you you is basically your fuel, your energy and I. And the second one is maybe, maybe more like network, the network of people that could help you to scale up, because you said that these resources, they are not fixed. And what is complex is that actually the number two is kind of fit is is connected to number one, because the Wiser you have your cap table organized, the more you can have this access to capital and resource could be. Because if you want to hire people, you need to connect with the good people, the perfect hires. And you require good network for that. And you might have compensated people with a big centrality in a network within your cap table in order to have access to these further hires, the same people with lots of access to VCs that you reward by being in your cap table might give you access to money. So it's interesting, because I will see number two as kind of a meta number one, in a way, number two is resource as well. Is the resource of the people you have access to to develop your company. But it's a it's a very specific kind of resource. It's a resource that gives you access to resource, a meta resource, I would see it that way. So that being said, first is cash, second is network. And then I need to understand,

    yes, so I would say that is a capital resource I'm saying, but I do understand where you're coming from. So for number two, I have very specific use product in mind. So if I share in real quick, it would be, you have two investment terms offered to you. For instance, like term sheet, a has a lot of investment amount is very big, but it has very severe. Downside protections or liquidation preference, is very high, so that between zero to $4 million you cannot get anything, the founder cannot get anything, whereas, like the second term would have lower investment amount. But downside and upside protection is less severe. So my point is

    it's the cost. It's basically the constraint and the cost to access to this capital and resource that you describe in number one. Basically it's how, how? No,

    I think what I'm trying to say is the choice of which turn to accept depends on how you believe this market will going to grow, and how the how the chosen market will going to grow, and how the chosen organization and the compete, like technology, which I am representing as a company, competence that is a multiplier. So if you believe your EV suburban market will grow like 20% next year, and your competence level is like from 0.8 to 1.2 as a multiplier to that, then that would somehow give you the idea what you're like next round valuation would be and based on That, distributions of a valuation, your preference would be chosen. And what I realized when I was doing this is, if you begin to add the market choice and also your technology and organization choice in this table is becomes too murky, too complex. So that is why I am fixing the market and the technology.

    Okay, I got it now. So it's basically the market. The number two is, is how you anticipate your performance on the market. And from that Jack, I will give you a concrete example, and you will tell me if I understand correctly. Now, I'm in a project at MIT with a student developing a startup. And it's basically, it's a storytelling toy with AI for children. Okay, it's a real example. I'm actually working on that. And of course, at the beginning it was working on features and what the product is doing, blah, blah, blah, and very limited resource, but, yes, creative resource and so on. That's number one. But very quick with other people, we said you should go more what you describe. I think number two, we talked about the thumb total addressable market. What will be? What share of that? Do you think you're going to have, whatever technology or features is like already, in a very large kind of fashion. How do you envision revenues? How do you envision competitors, and so on and so on. And it's a totally different mindset for him, because he's from GSD. Is Harvard GSD students, so he's design students, and is not Harvard Business School. Is not even engineer and so on so on. So we try, really to make this exercise, we seem to change the mindset and not just focus on, oh, it's a great product, because if nobody buys it, it means nothing. And to connect the loop, then this time, this market discussion can will help us to define at the moment of incorporation, because we are pre incorporation. Now we are more in the pre it's very early phase. But when we will decide about how you know, what is our role in the company, is it like advisor, founder, whatever, and also what you described as like the vested shares or not, there's different strategies on how external VC will see the solidity of the first team, and so on and so on. It's all based on the market, not the product actually, not the in an extreme way, the actual product is less important, because, as you clearly, as you rightly said, product can evolve and pivot, I would say, more easily than market. That's an assumption I'm making here. But let's say that as an as an angle, your identity as a team is coming from which market you address first, more than the product in which you address it. They are both important, but it's it's more difficult to convince a VC that you are addressing EV, and suddenly you're like, oh, now I'm doing biotech. You have limitation. Adjust. Adjacent possibles are okay. I was doing EV with battery, with iron, and now I'm doing it with magnesium. Okay, that's a pivot I can understand. But you cannot go from EV to biotech. This, this simple, you know, compared to, oh, I changed this feature on my product and this, yeah, I think there would be more understanding of pivot of features and products than total markets. So that's the first comment. But now I think I understand very clearly what is number two, is the constraints the strategy based on, basically, yeah, the markets, and the effect it has also on the composition of the roles of the first team in the company, which is very important. Hence the idea that companies from funded only by engineers, our problem because they might be too much on number one and not enough on number two. That's that's also something very interesting to look at. So now let's go to number three, number four, because I have questions. First question of number three, you talk about world modeling. What exactly do you mean by that? What is the model for you?

    I can address that before that, I would just add, I don't know whether this counter example to what you said or supporting what you said on adjacent possibility. But it seems in more manufacturing or operational heavy industry, there were a lot of success cases where they mastered the for instance, modular electronics or modular genomic service. And then they went to like six different markets. So, yeah, yeah. And so this is like one example of So Scott and Charlie thought that the idea of understanding the trade off between investing in operations versus the market understanding is very useful, because there are some failure cases where they had great market demand and traction, but they didn't deliver value is only captured once you deliver it, right?

    Yeah, no, no, you're right. So let me rephrase. I think what I tried to explain was the most extreme case where you had number one and and and be totally lost on number two. What you now show to me is that basically trade off between one and is very, very important. And this is true for the composition of team, composition of strategic goals, and so on and so on. If you have a great product and the market is shit, you will have no money. Game over. If you have a great market and your product is shit, you will not sell anything. Because even if there is demand, people will not choose your product. So definitely I can see, I can I can see this trade off now then, what about three and four? And what about world model versus regions? Great, I

    agree. I

    agree with you. It's a good refresh.

    So just to make sure trade off happens within one and trade off happen between one and two, and we need, I want to design that structure between three and four, and that's like where I need some help. Got it. And I think agent versus world is what we can think. And we can think of a situation where agent does not change, but word changes, or vice versa, Agent somehow changes, but word stays the same. And when I say like stays the same, it means it evolves much slower than the agent. That's where it's relevant. Um,

    and so do you? Do you take this terminology from Ai? Is it like a world model from AI and agent from AI agents? That's where it comes from. I

    today realized that the word model is kind of a hype recently. I Yeah, but it has been actually like the system dynamic. Actually have literally the word model. It's so funny we have the so the name is word three, so it actually has a word model that has food production, life expectancy and table scenario graphics. So I just mental model, or how we perceive, how an agent perceives the world. That's

    so in this in this world model in Van CMA, do you have agents? Do you have agents in this representation in world street?

    That's a good question, because we don't call this agent based modeling. But

    why? Because I can see process and agents here, right?

    Yeah, but usually agent based modeling is something. It's called compartment auto compartment model, meaning that agents are aggregated in here, so like population zero to 14, everyone between the age of zero to 14 is captured as some stock variable. So there is not really, like leaving people, like hanging around here is just compartments. So it's

    it's different scale,

    different scale. What do you mean that by that,

    that's an agent would be one person, age three years old, living in Malaysia. That's an agent, compared to the population zero to 14, in which is lots of agents. Literally, each box represents hundreds of millions of agents, somehow different scale of in terms of complexity, because if you want to represent the world, you cannot represent each agent of the world that's too complex. So you need abstraction. So it's an abstracted view of what happens in a more local agents way. Basically, okay, I understand more because agent. They have agency and world model, they don't have agency. They are more like represented complex processes. I would say it's so it's a different kind of representation. Remember also what we talked at the beginning of our discussions about epigenetic compared to genetic? I would say that genetics is agents and epigenetic is World model. And you could even say that products are agents and markets are world models, if we do this kind of analogy for one and two and three and four, in a way, this same kind of different kind of complexity, in the sense that agents, intentionality goes to the world, and world goes goes back to agents, in the same way than market is some kind of the landscape for products and product like you say, with agents and role model, maybe product does not change, but market change, or market does not change and product change in the same kind of dichotomy you describe with agents in world model. So I like this kind of symmetry, actually. I think it's interesting. Now go back to the number three. I would like to understand exactly what you are doing in this so I understand where it comes from, in terms of the epistemology, in terms of what kind of theories of a world, models you will use. Now, I would like to understand a little bit more this framework, what as an entrepreneur, what will I learn from that? Can I use this framework to create a representation of my startup company, for example, like new ink three, where I have all the process in my company, is this kind of big simulation of everything happening in my company, something like

    that. Yeah. My hypothesis was, was program synthesis. It's the analogy might be, you are, um, you have a map between the protein and a gene so that with the product. That, by the way, I'm not sure I'm on board with the product and the market, because I like that analogy, but I realized that it is better to have all the unit around the things that I can control. And what I mean by that is like actors. So instead of products, I think it's somewhat easier to think in terms of like producers, if that makes sense. So producers and customer head of product and markets, because they are more kind of fixed objects rather than relations. So um, having said that, so here the genes and products are kind of the analogy here, and what program synthesis can do, very easy version is like, for instance, the sorting algorithms it can produce the actual logic behind how to solve sort based on the data that it's given, like 132, and 123, if I make a lot of inputs from this to this, this probosci synthesis gives me the symbolic representation of the logic behind how it sorts this. And if we, if we find a way to scale this up, my like, best guess is probabilistic program and non parametric model, uh huh, and that is, I think so, just so

    I understand non parametric model is like heuristics. Is just what people do. What is a non parametric

    model like Dirichlet process or Gaussian process, where you have a situation where your parameter that you first set may not be applied one year from now, for instance, in your company, you would have, like only three technique technical developers now, but like one years from now, there could be finance people, marketing people, and you cannot have all that in your mind early on. So, so, so,

    if I understand correctly with this, I will have some strategic choices I can do about some process in my company, and it will help me to prioritize. Basically, it's like I can see it as my to do list of my company, and then it can tell me, Oh, actually, you should focus on this first that if you want to be most efficient. And I can explain to you how I came up with this reasoning? Basically, I could deconstruct my reasoning.

    Yeah, yeah. So I recently realized that I should a little lowerly adjust my tendency to become an engineer, but rather first be a scientist, and then, if this makes sense, become like making something. So I'm not sure how much usable this would be, other than just educational purpose, just to say up front. So

    the to give a more concrete example of what I'm saying, I just added the paper that discuss about actually, you actually added the hidden file, not the paper. Okay,

    let me do it again.

    I think it's, it's a shortcut to the paper.

    Yep, here we go. So the same person has a thesis as well, where he describes how to scale AI and I would just capture that part and share it with you.

    Yeah, the path to scalability, and I'm specifically focusing on one point 1.2 and um, perhaps, or maybe not, one point 1.4 meaning, instead of using MCMC, I want to use SMC. But yeah, I think developing, yeah, I first want to understand this. I don't know much about it, but my goal is reaching to number five, which is, I think more, yeah, less real startup cases, but more on how to build a programmatic theory. So, yeah, somehow a little entrepreneur education is more, I think, on my mind right now. If that makes

    sense, there is this. There is this concept called Explainable AI. Explainable AI. So maybe you could do enable entrepreneurial, entrepreneur, yeah, AI, or something like this. It's basically, it's an AI that is teaching you each step of its reasoning. So it's learning and AI combined. Basically, it's a very interesting field. Yeah, you can look it up if you

    and I think the important part is the machine is perception grows together with the entrepreneurs. I think that's very important. Yes,

    this CO adaptation. So there is this term co adaptation in evolutionary theory that both the machine and the human adapt and CO adapt together. So there is this shared reasoning, also this excellent paper of Tom Malone and these two other people that you shared about this human AI cooperation that we have, we have hypothesis that it's better, but actually it's not the case all the time. There are situation where it's better to have collaboration, and some where it's not. I really like to read it you shared recently. It's something called when AI is better, or something like when something starts with when. So now back to just very quickly on number three. What will you evaluate in the paper? For number three? You will create some kind of computer code and system, or it's just a framework you describe. What what is your plan there?

    So I would be happy if I could automate the model I made in one in three with the addition of evolving world. Evolving world, I mean nail scale sale. So Charlie has a like linguistic approximation of like jungle, mountain and sea, of each three world. And what I mean by resource, rational decision, example is in three different environments, electric vehicle, where you have both hardware and software, and like very IT industry and traditional manufacturing, your optimal allocation of investment to either operations or the market side differs. You can expect software where I'll just say software industry would require more market understanding. And I think for paper four, it would be very helpful to learn what causes that. And my hypothesis is fastly evolving the environment that would cause, like, if the customer's utility ships a lot, it would force you to keep investing in market understanding, because yesterday's knowledge is useless. Tomorrow, on the other hand, for Operation Center, this traditional manufacturing, or like gin therapeutics, where, if you make some like cancer curing disease, or like some obesity preventing drugs, or like some leather product that you know there is a fixed demand out there, Then it's straightforward to just develop something and push technology, push so in our paper, the product for it would be really helpful if we can really somehow analyze, what are the factors? These are, what is observed out there, and what are some factors that makes, I think both of them are rational resource, rational choices. But the problem in entrepreneurship is one concept that has succeeded in certain situations like blitz scaling, scale up really fast, that was invented in a software industry where markets are really unstable, so you need to scale up fast, to out compete the others at all costs. But that does not apply to Operation Center. So, so that kind of contextual knowledge based contextual knowledge, we need to understand where that is coming from. Yeah.

    So, and in the paper, I understand, you will extend what you do for number one, so you have some kind of a environment. Then you will, you will simulate. You will use like simulated data. You will create like different kind of scenario, and you will simulate them, and you will see how they perform in your model. This is what you're gonna do.

    So when I produce this, I assume the environment is kind of stable, so I need to add some dynamic in there, meaning the market itself can evolve, which is what Charlie's model. Word model is happening.

    Um, then, then, then, once you you implemented this, then what do you do? Would you create some kind of an experiment to test out in a evolving model, what happens if this and that? I mean, you have to show what the model is doing, no for the paper. Or you can just say, I've created this model, and it's great.

    Yeah, we need to, we need to show something with a simulation model, and it needs to somehow explain the behavior observed. Yeah, and think the from our Was it this one where we had a like, Do you remember where we had the NASA Parallel thing? Yeah,

    NASA is a good keyword. Maybe

    Yeah, you're right, yeah. So this one, so since our topic is when to do this operations, or when to do this operations, if we can somehow, like there is, since the first papers framework was, this is a founder or a startup, and there is The supplier side and there is the customer side. You have like, let's first simplify this. You have two choices, supplier one, supplier two, and customer one and customer two. So example is urban market or the suburban market and outsourcing or in housing, or China factory or US

    factory, that's exactly, exactly,

    yeah, and there I found many cases where like this affects This meaning, if you like iPad or Apple, having a supplier of a fox cone limits some market they can expand. For instance, if you say the fox cone semiconductor, or whatever they're producing, can only endure the temperature from 25 to 30 degrees. You cannot sell it in Middle East, for instance. That's right. IPad would explore, explore that. So my point is, unlike the current literature, which only focuses on this part, or even this part, we need to have a more holistic view. And I think this is a great starting point to think about this question as well meaning, if you call this sequential and parallel, you can think of a situation where you are doing this sequentially, but doing this parallelly, or parallel or sequential, or parallel or parallel, many different cases. So that's an idea that might be very coherent with the what I'm finding in number one product.

    Okay, the question is, how to model that, because in this case, we could have like, market pivot. We can have product pivot. We have different kind of pivots. We could model like agents with different kind of resources. In terms of you are talking about cash, for example, or people related with them, how many software developers you have to do something so to which level, like, how do we model an agent in this in this kind of context? I think it's already complex. And then second, I think that this notion of spend real and adjacent possible is very interesting and important. I think that this is something we should put in the model. And this is where exaptation is relevant, because if not, we are just talking about parallel versus sequential. I think if we, if we, if we add the notion that when you, when you have to choose between parallel and sequential, that all parallel are not possible. Maybe you, you you can only, you can only do a variation of something basically based on an existing precedent, based on a prior is something is something interesting? Is that? I don't phrase it well, but I want to say that, okay, if you do something linear, you could say, okay, product one, tested, test out, then works or not? Product B, product C, okay. Sequential, if you do, if you do parallel, you could say, I can test if I take back my example of this storytelling toy for children, I could say, okay, my first product is a little box that tells story to children, and if, in parallel, I do a biotech thing, I don't think it's it's possible to reason like this. I think that it's more interesting to say what kind of affordance my product is having. So let's say, for example, it has an electronic component, a software component, some kind of shape component, and from this kind of features or affordances of my product, I can then create parallel track that will explore one of these in a different way. So maybe a different electronics, different software, maybe not AI. I do a parallel track with not AI, and it's exactly the way now we are thinking with the startup, we have this problem, shall we focus our resource on one demo that combines all of it, and then we will do first market trials, and if it doesn't work, we will have another one in the future? Or shall we investigate three different scenario, maybe one with AI, not with not another, not without AI, one in an app and one in a physical object. This would be four different things to do in parallel, instead of having one physical box with AI with this kind of scenario. And it's very important, because we have so limited resource now with a startup that we don't have enough people to prototype all of them, and this is why it corroborates what you said about a strategy that hints at being sequential, because when you are at the beginning of a startup, you cannot multiply yourself. So actually the parallel Strategy is a strategy already of rich people, because you have to have enough resource to be able to do parallel work, basically. So. Long story short about that, I think that we should find a way to to agree on what is an agent. In this context, is an agent a product, for example, I'm not sure, maybe, or is it like, what do we mean by agents, basically, and then how this agent evolve? I think this is what we will talk about. I think we agree on that it's about evolution of an agent, and this evolution could be linear evolution or parallel evolution. And if it's parallel evolution, it means that it's not just one agent. It suddenly is. You have four agents, and they interact between each other. And this is not talking about modification of the environment. We assume that it's a it's a constant market in this scenario I was talking about. So this is the the way. So maybe an agent is a project in a company. And in a project, you can have people, you can have resource, you can have product with features that that could be because, you know, you have this notion in a company of PM, product product manager or product owner, po the product owner. But sometimes it's the pfpm is also project manager or project owner. It's not just product. Product is a sub scale of that. So maybe, yeah, so I see concepts here.

    Yeah, I was trying to find the paper that I had a two paper on the parallel versus a paper that supports sequential versus parallel. And I'll follow up on that. I cannot fund it right now. Okay,

    so you talked about agent, what could be an example of an agent in your in your mind, what could be an agent in this context?

    Could be an agent in this context? Um, I was just thinking about just one founder. I think that is easiest in most cases, because if you think about like my operation right now, if I say, if I if you forgive me, I am a founder of my research, I think what I'm doing is parallel search. And the reason to do that are several. One is, I think constructing a value chain with some people knowing something better than the other meaning, I have a I have a whole,

    okay, okay, I think it makes sense, because an agent then can make a decision, if he's a rational agent, for example, can make rational decision, and an agent can have a budget of choices. It can make that many choices per day. He has maybe limited time, limited money, and so on and so on. And then it can interact with other agents, that could be a VC, that could be other companies and so on. So then, then I could see how it works.

    So, yeah, I have a clear example for when parallel is actually better, even though it slows you down a little. So in entrepreneur finance class, the lawyer advised that it's better to speak with several VCs until rather than just accepting the first terms out there, because

    great example of multiple agent negotiations. So in this sense, parallel like choose four, pick one is better than potential, like the first one you take and you're married for life. It's also a good, good advice for choosing a shrink, a psychologist, you should test out first few first, and then set your mind on the one you feel good. Not just, oh, that's the one. And interesting. Okay, so

    how much variety, like, variance among like, let's say every for for me and I, if I'm searching for the partner, if there is zero variance over the men, then I don't need to test a lot. I just need to find the first person, because I know that the second one would not to be different for the first one. So that's one thing, heterogeneity among the population, and the second is about the cost. If the testing cost is very high, I wouldn't test a lot. So that was part of our original paper. I thought there was a part about like. I came up with three conditions where the parallel is better than the sequential, and I made that based on like how I felt, sometimes sequentially, and when I say sequentially, like I am doing paper one and two together, somehow they create some spillovers between them. And I think that is somewhat related to spandrel as well, or knowledge spillover economists would say. But yeah, so do you remember the three questions I had? Yeah,

    I think somehow it got erased in the final version, but let me try to pull that up, because I think that might I

    mean. What is interesting about exaptation in what you just said is that if we take some Bayesian approaches where we have a prior and a posterior, the exaptation is giving you context on different kind of priors. And to my knowledge, usually in Bayesian approach, all the priors are the same. It's like they are homogeneous priors. And with exactation, you have heterogeneity of priors. You could distinguish different kind of priors based on some characteristics. And that could be interesting to to look at that actually. So,

    so I think I agree with you, but I don't have a very clear mathematical model of the acceptation that I think is the fundamental challenge that we would have throughout this paper, whether kind of I know what I'm like for the words that I wrote here, like when the cost of each test is slow relative to pivoting costs, or when uncertainty in technology innovation and customer segments is high and when the true values of technology innovation and customer segments have low correlation. I have some idea of how this maps to the code in the model, but what I need your help is to map this concept with acceptation. I don't think

    that's very important. I should write it because that should be my next priority for next, next time I meet with you, to have provided you with enough knowledge about exactation, like enough operational knowledge about it, so we can turn some kind of a framework which is qualitative into a quantitative, operational, mathematical version of it, even if it's a reduction, even if it's not the best, we have to start from somewhere. And that's that's so maybe I could create some kind of a diagrams of a diagram of adaptation, based on what I know about the theory of acceptation. The challenge is that there were attempts to mathematize adaptation. I did some literature review of that in the last few weeks, and they depend on which particular discipline you use adaptation. So for example, I found attempt to mathematize adaptation of spatial conformation of molecules in chemistry based on adjacent possibles, but it's very specific. It's at, you know, Femto scale. It's very small scale, and it's about how molecule can have different kind of shape in space, okay, but there is a mathematical model here about how it evolves, and different kind of possible futures for this confirmation in space, I would say that in the context of innovation and startup, I would feel uncomfortable to take something from molecules and directly applying to innovation and startup, because it's so different. I cannot see how as an inspiration, of course, but we should. We would have to create our own mathematical formulation of exaptation for business and for our particular use case, I would say, and I think it's fine and it's fine that it's not the best. It could be refined, could be criticized by peer review, and maybe they will say it's shit. You have to do it again in this one this way. No problem. This is how research is working by being falsified and being improved by the community. That's okay, but I think that I can contribute, if I can give you a kind of a metaphor, a visual metaphor about a process of adaptation, and then from that, you can turn it into some kind of, yeah, mathematical or formulation, some, some kind of a computable process, maybe also, maybe not math, something

    you can so, yeah, I guess, um, the Next meeting school may be for my end whether to kind of include the acceptation in like this term of collaboration. I wish to search and learn about the next possible, like near possible thing, but I feel like maybe this fall might be too near future for that, because I want to first make some concrete model for the product one and product two. And I think after, if you could share some after our meeting next time, if somehow, like my concern is we have some model in it, and if acceptation, trying to quantify that might blur the logic, if two different logic that I don't understand very well comes in, yeah, yeah. Or if I would also be happy with that, if you could somehow lead the quantification of the acceptation part, because that would be what would be needed eventually. But I think that's your choice. Does that make sense?

    That makes lots of sense, and thank you for sharing. I appreciate it a lot, because these are your resources, your constraints, and I want to be respectful of them, because I know you have a lot on your plate. My constraint and my limits is now that I kind of committed with this abstract I would like to try to deliver something of a good value. I might fail, but,

    yeah, it's very low cost failure, right?

    It's very low cost failure. Exactly that being said, it should be treated as professionally as possible and be seen as an opportunity to test out these exaptation meets different kind of parallels versus sequential evolution, kind of hypothesis we have that connects with your big model of other kind of optimization of resource and constraints by Yeah. Founders, so one thing about the agents, I the agent is a rational actor. Is making choices. And what is complex for me to understand is like is not evolving the agent. The agent is a bit more like a god in this model with evolution because No, because he is the agent, evolving itself for accepting, I don't think so. I think that in in addition to the agent, we should add another concept that could be, for example, an actor. And an actor could be a product, a resource, a market, a bit like a stakeholder in a process. Basically, what I like about actors is that they have been defined in terms of programming. I put a link about this, about Carl Hewitt that I had the chance to meet at some point. He was at MIT. This is PhD in MIT, about actor, modern network. So very, very interesting way to define elements in the programming environment where, on the contrary to traditional object oriented paradigm, the actors, they are defined in in terms of what they can do, in terms of their agentivity In a way, and and and they broadcast a set of things that they can do to the world, and then you can register to them or not, compared to object, where you have to probe them, you have to build them in a different way than this. So if we could have agents that can act and but there are people in the real world with rational mind, like human mind, and we could have also, maybe non human agents. They could be programs that could also propose some kind of choice, strategic choice for the company. So we have already two kind of agents. Then we could have VCs, other agents that can provide money. We could have maybe other kind of agents in our in our model, but these agents are not like one agent going in a sequential term or parallel term. I think it's these agents take the decision to be sequential or parallel and what is going through this parallel process, or sequential process, is more like actors, is element of the process, like when you create a product, it's the agents. The agent is deciding to put all this resource on one product development, but it's the product that is going in a sequential fashion. It's not the agent, right? You agree with me in the when you say the when the cost of each tax is low relative to pivoting cost. What pivots is the actor is the product. The agent is not pivoting. The agent is making the decision to pivot. Is making a choice. But the pivot itself is sustained by an actor, by by, by a product, or by or the market is pivoting. So I would like that we agree on the term, like, when you talk about pivot, is it a pivoting decision, or is the actual pivot by a set of resource? I think it's important we are we agree on what we talk about when we talk about this

    great question, so what's the example of the pivoting cost? One tricky concept here is so pivoting, you earn some money by experimenting. That's, I think, the most challenge to me, meaning, if we just put it on a scale of revenue and profit, if people say they try like suburban market with 400 or 300 range, and they pivot it to urban market, Because 300 mile range, it's better to have more EV charger concentrated area. But the thing is, even though you pivoted and you somehow frame the failure in the suburban market, you would have some revenue there. So I am having a little difficulty in how to quantify the pivoting costs. But here, what I was trying to emphasize was kind of testing is quite cheap. But I just needed some reference thing, and that was just pivoting costs.

    So for example, it's cheap. For example, when you you're a good developer, and you can code super fast, and you don't have to pay someone to do that, you can, in 10 minutes, opt for the demo your prototype. So you can do a test in your 10x programmer, and you can just quickly prototype something, or you're a great designer, and yeah, in one day, you have a product already you can demo. It's super quick compared to paying a expensive company, okay?

    Scott will call this as partial commitment and pivoting cost is you commit something and you change it next. So, yeah, does that make sense? Yeah,

    that makes sense. Okay. It helps me a lot. I think it's already lots of information for me to process. I have a clear vision now of also what is my role, which is from what I understand on this paper to test bed, lots of ideas about this, parallel versus sequential, and the link with exactation. And I see, I feel I will do most of this work for the moment, because you are super busy with one and two in the coming year, let's say six to nine months, this is your priority. If still I have a little bit of your time during this next six to nine months. So in our meetings, I will show you my progress on how to go from an abstract to a paper for cast 2025 and benefit from your input on that in these moments. But don't ask you more than this at this stage. So

    thank you. Yeah, and yeah, at any point,

    at any point, one is developed enough so we could fork it to to use in four a little bit, and make a four version out of one. That could be awesome. But I'm not putting any pressure on that. It's just because that would be a nice exactation. In my opinion, Forking is like exacting in many ways. So if we could do a variation of one to four, that could be great. But yeah, thank you. It's very clear. I see the challenges now. I see the value as well, and I see the big plan so I could connect both dimensions. Thank you for that. Yeah,

    sure. And if I'm not overloading you too much, I just want to, I think this is very, I promise this is very relevant to kind of the spend your concept. So if I may, so relating where I got this idea, cost of testing versus pivoting. Cost is from test to choose one paper. And have you heard of about like amortized learning? Say it again, amortized. Learning. I just put in a chat. No, um, so the idea is you somehow before, like it is too late to react to the incoming data without any well prepared prior and you previously mentioned like a prior concept. And so what they do is you have some hierarchies, like for instance, I think example of for human amortized learning is having some expectation of different culture and nationality, their traits and then interacting with like people from French. Oh, they are really a great debater. I have some expectation of that, and I just based on that expectation, right? So that is what they call kind of amortized learning. You learn, you compute in advance, so that when you're interacting with them, the computation cost is like kind of cheaper. And I think the test to choose one, I really took a lot of time to translate the paper the so the GaNS has to choose, hold

    on. I understand correctly. It's kind of an hypothetic prior. It's a prior which is an hypothesis. It's not a real prior because you never talk with a French person, but because you say they are French. So I, infer, I assume that they might be good in this and that. So you create a fake prior in order to basically tune already your model, if I understand correctly,

    like hierarchical Bayesian I think it's the most accurate concept for that you have a hyper prior that explains the French people as a whole. And then the individual prior is like JB or other, I don't know French people, my friend Leo,

    it's an instance of the global category, basically, okay.

    And so this table compares test to choose one and just test one, choose one and no testing is just choosing one. So you can think if the pivoting cost or the testing cost is very cheap, you would most likely try to test something a lot in advance and then implement this. So let me, let me step back a little. There are four components, or largely three components, in the final cost that are comparing. So the paper explains in what situation is test to choose one better than test one, choose one or test no testing and just committing. And the three components is testing cost, implementation cost and a pivoting cost. And in my like, my intuition for test to choose one is better is you somehow do some learning in advance, so that the pivoting cost is a little less detrimental, because you somehow cross out the bad idea in advance that happened through this normalization of so this PL is the instance the probability where your idea quality is low, so there is the signal coming in, and the Q stands for the blatant value of the idea. And we Sorry, sorry, the value of the idea. There are three low and medium and high and Q is you either implement it well or you implement it bad. So for instance, if our paper on connecting acceptation with some Bayesian, we don't know the quality of the idea, but there can be six cases in here. It can be either very low, but I think we crossed that instances out by trying to propose, and our abstract got accepted, so that could be interpreted as one testing. So that is the magic happening here, meaning that we, by testing in advance, we crossed out the idea that this idea is like terrible quality. Now we only need to think about either it is medium quality or very high quality, and then some noisy learning is happening, because we don't know whether this idea is really gold, but our paper writing skill is really terrible, so we got, somewhat not getting, a prize in our cast conference, or if we implemented it very well, then we would get some good price. So maybe pivoting can be, I don't know, during the process, as we're implementing we realize that this idea is not kind of high so maybe we found a new kind of better idea within this opportunity space. So that's incurs some cost, but the key idea is, but by testing in advance, we somehow lower the cost of this. So I just wanted to somehow connect the

    Yeah, thank you so much. It's it's very interesting. It's also linked with the NASA kind of diagram, and it makes me think also about how to how people make experiments in computer science papers, very often, they do a pre study. And in the pre study, you test out, you test out some, some, very quickly, some ideas that you could do in your real study, and you, you isolate the bad one in order. Then when you do your real study, and you put money and time in that you already checked out the the worst probable, basically outcome. So you, you kind of, you reduce the space of basically you that you investigate, because you have finite resource any anyway, so you cannot do all the experiments. So it's not,

    I think it's the key. And I hope you're curious to know, like, how spandrel, like, we are limiting some cross is there some renormalization happening in spend role? You think so.

    The thing with spend roles is that if you, if you, if you think in terms of creativity, of all the ideas possible, let's say all the shapes possible for an animal like what evolution shows is that you cannot go from a fish to a dog or from a fish to an insect instantly. It might be possible by a long chain of animals, but it's only this. Is why I talk about this adjacent possibles. You have physics in the world, and because of physics, because of entropy, energy, you have lots of constraints. You cannot overrun. You cannot, you cannot, you cannot go from a huge elephant to an insect in one second. It's it's not magical, yeah, in a cartoon, you can, but not in physics, because in physics, you have to take into account lots of boundaries. And this is this, is this boundaries, this adjacent possible. And in these boundaries, you have things that are already organized, literally in terms of complex science, it's about organization with symmetry, with lots of things related with complex science there and things that are not yet organized, either they are chaos, so it's like a disorganized set of elements, or it's some kind of a void. It's an empty space a spender. But the void in nature, they never exist by themselves. There is no real vacuum in nature. Even in a vacuum tube, you have a tube. Yes, there is vacuum in the tube, but there is a vacuum tube. In order to have the vacuum, you don't have just vacuum out. You know, there is always a universe around everything in the universe. So this the example of the vacuum tube. Is a good example. In the vacuum tube, you have a tube so you have a specific kind of vacuum with specific properties. Maybe you will make vacuum at a specific kinetic, specific speed, because it's a tube in glass. And maybe in a different kind of environment, you have a different kind of vacuum. Maybe in an autoclave, in metal, you will create vacuum but in a different way, maybe with different kind of condition of energy, you need to do that, and so on and so on. So the idea of spend role is that it helps evolution to reduce the kind of next move we can do because of two things in exaptation, either it's a physical law, it's a morphogenetic byproduct, it's the physics of the object that that makes you reduce the choice, or it's a tray evolution, and it's basically a previous tray that is accepted into into a new one. Is like you had already feathers, and now you it was for timing regulation, and now you can fly. But you you are not going from time to regulation to fly with no reason is because you have this it's called, it's called an invariant. You have an invariant between flying and atomic regulation, which is the feather. So if we could track invariants and we could measure their costs, I think we could do a good progress, because we, in terms of computability, we would reduce by an order of magnitude the computability of this problem. So it would be not only faster, but in some cases, it will make them computable, compared to NP complete, compared to not computable at all. And that would be the huge thing. I mean, if you think about World Three. The current world three model is like few millions of parameters. It's huge model to compute. It's compared to the first one in the 70s, where it was hundreds of parameters. Went from hundreds to millions of parameters. It's a huge difference. So I think that's that's the interesting thing about acceptation is that it can create coherent posteriors, like statistical kind of vectors, which is great for computability, but also it's great for the way a human could adjust his reasoning to the whole process, because if it's just machinic agents that compute between themselves, they can go to very complex things. But if you want to have a human in the loop, you have to be able to frame down the complexity of what you do in a way that a human will understand. If we create machine that can play go like no human can do. It's interesting, but at the end is like we press a button, and at the end is like matching a one over machine B. Great, great game. So boring. If you want to have a human deconstruct the logic of all the games, it has to be human readable. It has to be attuned to what is the reasoning of human and I think this is a key thing that we almost want to grant to work on that with NASA, about that they create AI model that are so perfect that the human operators, they don't know how to operate them, because the the the programmers don't put human in the loop, and then the they have to steer, take responsibility the human agents, but they are not including the model because it's not the same level of cognition. That is, they cannot synchronize their cognition to the one of the model. That's a big problem. I stop here. It's a lot of interesting things. I know what I have to do, and I know you have lots of work, and it might be late in et, yeah.

    Thank you. JB, it was really productive. So you let me know when you would need my input, or how do you think we should plan out?

    Beginning of December? Beginning of December?

    Oh, yeah, sure, sure, sure.

    So, let me look quickly at Cas, yeah, five deadline, class, 2025 schedule. Full paper to review. Okay, we have to meet before December. Full paper to review is December 2. So I would like to have an occasion with paper submission to have a chat with you. So let's say one week before that, let me look at my my Tetris calendar.

    Your Tetris calendar?

    Yes, my name is Tetris. Yes in your calendar. So, yeah, I love metaphors in general. So Okay, how about, how about we meet on the week of the 25/25 of November. When will be a good day in this week for you?

    Oh, yeah, 25th of November works same same time. Or you would be in France, so you would make earlier time in France.

    So earlier would be better for me if I'm in France, this time would be too late for me, because now

    25th Oh, 25th How About 22nd or hold on,

    22nd is a Friday, yeah,

    or 25th 11 or 10:10am, yeah. How's that? Sure?

    Yeah, okay, great, that would be my 4pm so that's perfect. I can, I can do that.

    Hey? Great, yes. And right?

    Angie, have a good evening now. Good night and lots of good things, good motivation, good energy for your works and we meet. Ding, I just received your invitation. Great. Okay,

    Jamie, I will miss you for one whole month.

    Exactly, but you can always send me lots of things and I receive all your updates on GitHub every day so that's great. I see.

    We are digitally connected. Why, thank you, Davey.