Speaker 1 0:11 We have three things more. So I Unknown Speaker 0:16 spent a lot of time writing something. Thank Speaker 1 0:19 you. I would spend three times more than that. to process this. I think this is really important Speaker 2 0:32 so I'll work on this in the next couple of days. Thanks. Speaker 1 0:38 I'm building dashboards and I wanted to so before jumping in today's agenda is for the next two quarters, like what like big picture for what I am trying to do. And I want to get your feedback on that. And also, the first quarter all revolves around what are the deliverables for the simulation models are. So yeah, and develop some simulation to discuss with you and Scott and then I will see where it goes and that will also affect what classes I'm going to take next semester. Okay. Yeah. Speaker 2 1:33 I'm not sure what classes you take should depend on what happens to your simulation model next month. Why? Because you take courses to develop core capabilities. The simulation model is one application of applying core capabilities. But Speaker 1 1:54 like as I go through, if I'm having trouble with some certain parts, maybe I can think oh, I should need some more capability on that side. Like for instance, I've been trying this coding with GPT on the side for the last two weeks, and I really improved my capability to code with AI. And I think the clock speed in the data science space is really fastly evolving. So yeah, Speaker 2 2:24 okay. All right. So well, so courses in this fall are still TBD. Speaker 1 2:29 Yes, but I'm thinking in terms of more like coupling to science, computational cognitive science, but I think I need some discussion with you. More on that part. What's your general thoughts about me learning computational cognitive science Unknown Speaker 2:51 What do you want to be when you grow up? Adult. Speaker 2 3:07 So, when we go from operations management to entrepreneurship, I'm already on relatively thin ice in terms of my own knowledge. If you asked me to know something about cognitive computational science, I want even thinner ice. Not even ice layer anymore. Speaker 1 3:30 Are you sure? Shifting coming in, right? I mean, the article ethanolic article you sent me. I'm just I'm just curious, what's your general take on that? And also, I sent you some analysis and I also wanted to get Speaker 2 3:44 my take on it as is that things are moving very quickly. And which makes it even more important to think about what are the core capabilities that one needs because as tools are changing really rapidly. It's not clear when we becoming an expert in any one tool is going to be a very long lasting advantage. Being able to think deeply with the frameworks of other people who've already thought deeply about things seems more long lasting. I mean, the laws of physics aren't going to change, right? I don't think the laws of economics are going to change and the laws of human biology and psychology are probably not going to change. Now, they may be adjusting their makeup, bio humans, I mean, bio robots or something like that. But Unknown Speaker 5:05 maybe that's going to change. I don't know. Speaker 1 5:10 Do you think I should train myself in jumping around in ISIS? Speaker 2 5:18 I feel a lot more comfortable. Being in my 60s and being in my 20s. Speaker 1 5:28 See the answer to your question, what did I want to do and I grow Lancers 60s? Unknown Speaker 5:41 Okay, Unknown Speaker 5:43 agree these are important questions, but I'm not sure I have answers. Speaker 1 5:48 But do you generally agree that the research is being more integrated? Like are we on the slide from? I was reading this book again today. The part was the bike bicycle showing. Like I'm having trouble there are so many words coming in that I'm not very comfortable with like different brands. Different brands have applied. So I find it much much denser than like a math textbook. Speaker 1 6:26 So my question in terms of a clock speed, where are the research currently in? Do you think? Speaker 2 6:35 So the reason I wanted to latch on to genoise Singapore project is because I thought as a group, they were being ambitious. About sort of trying to ride that wave, whatever, wherever it may. Now, I suspect within the campus of MIT, there are many other groups that are also but genoise was within reach to me and also there are some slower ball and I knew I just felt like as an individual or duo, you would be well, we need more. I needed a bigger team to be able to try to comprehend what's going on. This guy Ethan Moloch is very impressive. We seem to spend 100% of his time doing AI stuff and he's seems to stay on top of a lot of Paul Yellen. is sort of also staying on top of a budget. But I feel like you and I would benefit from being a part of a larger set of smart, capable people who are all trying to figure this out. Where they did was group is the best group to be attached to I don't know, I think a woman from EECS to Romanian woman was yeah, she's kind of really is kind of a superstar, I think. Some of the guy from the Media Lab lab is very, very different. But he also seems to have a knack for staying on the cutting edge of things at tombolo even from a different angle, so I just felt like they're interesting people there. I wasn't that excited about sort of the over focus on the Changi Airport. That is the wasn't very clear to me that there. They had completely the best focus but nevertheless, they are very capable people and and I thought bringing the entrepreneurship dimension into their world, which was not in their proposal. Was the value added I think, the same. So, so part of my answer to your question of what's going to happen is, I feel like to see to help find that answer. One needs to be engaged with a set of people who are also seeking that answer. I don't think we can do it by ourselves. Speaker 1 9:51 I agree. You're around remember that like early this I think. Last winter alone we had Jim Chad was General ume and I asked is Gene Ron cod on the like, between the supplier to the customer side, how can we put them like place them in and you mentioned generates more customer side for for our product? UI product and Scott is more on the supplier side. Speaker 2 10:24 Remember saying that but I can understand the logic of why Scott's providing theory that helps to contribute to a tool use for two purposes. Speaker 1 10:38 So my take on Gina's project is I think it's kind of I was expecting to find some supplier from like CO supplier for my theory, but it was very hard to find someone like Daniel who is a hardware side and it's very broad, but that somehow, I think makes it less focus for me to find someone who has the exact same interest. So it's an amazing market. And but it's hard to find some suppliers who share interests with like probabilistic programming. So I admit that I have a very like narrow core rigidity. But I I made a bad thought on it. So I'm kind of biased towards him I think. And I was why as like, given this cool rigidity, are you able to kind of forgive because of his potential benefits? That's a male that I sent you during the weekends. So especially in terms of LLM, I think there isn't any core supplier in June has like the whole group are able to do that. Large language model unlike like poverty programming, who are really making this part with language model and do you remember the Bayes DB that I showed you last time Bayes DB where they condition so they have the technology to get in certain words or sentences and turns it into structure output. And they express every operations into three things conditional and sample and not probably distribution. So when we offer too generous group about combining Jen AI with intrapreneur theory, I don't have the specialty in dren AI, but I think if I could learn from that I can get the capability. But that cannot come from like genoise group. So that was why I've been interested in their research and one of the PhD fears project this was his thesis is preparing it for it and he his brand. He's branding himself as probabilistic programming for data science. So he came up with probabilistic programs that are marked to task learner for data science and also uncertainty over parsing of natural language into domain specific language. So when they say domain specific language, imagine this as Scott's entrepreneur compass. He has certain words that associates certain parameters with it, and they provide some connection to it and the last one is above our old data types or Tabular. But in entrepreneurship, for instance, it's a stakeholder between investment and market customer and different co founders. So database is much more efficient if it is stored in relational. So we need another set of logic in order to do conditional distributions on the relational database. So I think this technology would kind of go along with what we are trying to do like so that's why I want to learn this but it's like yet another task. So Speaker 2 14:08 what do you have to learn? That was what does it take to become proficient in probabilistic programming? Speaker 1 14:18 So this area itself is expanding, so I need to kind of choose between Stan and Jen. Jen is their language that they're betting on? But I think too, so because she is providing seminar and Josh Tenenbaum is providing computational cognitive science next seminar next semester. So that can be one starting point. Speaker 1 14:49 When you ask that, like Where were you expecting some out like numbers of hours or like what will Speaker 2 15:00 I No, no, I don't have a picture in my mind of what it looks like to master probabilistic programming. And what about what is the probabilistic program master look like? And what does it take to become a probabilistic program master and maybe Master is not the right word, but someone there's a certain level of competence in probabilistic programming. What do you look like when you've got that and what does it take to get there? Speaker 1 15:35 I see what you're saying. We I asked the same question in economics. Speaker 2 15:40 So economics, I think, economists have a lens that they can use to look at any problem. Show me any problem and I'll tell you it's a labor economics problem is that an industrial organization problem has an incentive problem. They have a set of models and a set of frameworks such that they can put their analytical lens on a situation a certain circumstance or situation and give you some analysis of that situation. That might be predictive, that my, that often is predictive, it's theoretically predicted that I predict given this, this this, this is going to happen, and here's the logical reasoning. Why it's gonna so that's what I think at a high level, The Economist worldview this that useful, does that align with? Speaker 1 16:39 Is it Congress itself like expanding the field itself? So Speaker 2 16:48 there are multiple trends in economics, I'm not on the forefront of economics, but that right there are multiple trends. One is, I think, more data analysis, and other one is more incorporation of other social sciences. So economics, sociology, economics, psychology, that is, the neoclassical economics assumes that individuals are hyper rational. We know individuals are not hyper rational. They are psychological beings. And they interact in sociological frameworks in ways that are not consistent with hyper rationality. So economics is try and political, political science I think, as well. So economics is growing in the sense that it's trying to incorporate ideas from these other places at the same time, it's trying to be more quantitative, the more intelligent about data, analytics to answer questions empirically, that they might be able to hypothesize about theoretical. Speaker 1 17:56 So I guess there are different ways to learn this. And like one thing I'm trying to do this summer is read through a Josh Tenenbaum textbook and they have the code that I can run but I'm mostly want to learn about like conditional inferences and language of thought like this language, large language model and how to represent in a probabilistic programming is new to me like the basic structure of like a demand and supply curve. I know that amount of portion in probabilistic programming. So the new part I think, I should learn for instance they have some model of generating Speaker 1 18:50 or Smethwick function and also inferring the function based on the output Speaker 1 19:02 Yeah, but other than that, I think learning by doing meaning, applying the probabilistic programming framework to my my problem, and my data said would be the best way to learn. Like for instance, I had the meeting with the person who was doing problems the program for data science, and we were able to plot the his framework into our pivot market and pivot product situation. Like for instance, we have a joint distribution of market product and like profit, which is a continuous variable and the objective our goal and what it does is it computes based on this market I'm in and product I'm in what is the expected outcome or the profit that we can get. So that is that can be translated into several lines of probabilistic programming codes, and how to connect this with the two library they have built which is cross cat. And Jen SQL is I think, the next step for me to learn about as I interact with them. It's very like new field. But I think they they're very promising. They're working with Google and DARPA and many other institutions. So what's a Unknown Speaker 20:35 wave directing women Ruth does she think she left something in the office here? Sorry. Unknown Speaker 20:46 To take all the gifts Speaker 3 20:51 I don't know. You can see like a rolled up piece of paper. She's a massive no no. I'll just let her know Unknown Speaker 21:01 if I find Yeah, Speaker 3 21:02 can you see something like rolled up I don't know. I can't really hear the connections documented too well, but yeah, thank you. Sorry about that from from Unknown Speaker 21:11 are trying to ask me Speaker 2 21:14 what is what's a question that they want to answer the probabilistic programming people that they say probabilistic programming is a really powerful way to answer this kind of question or do this kind of analysis what? So what is DARPA want them to answer what does Google want? Speaker 1 21:35 Uncertainty modeling? What I mean, Speaker 2 21:42 what, what's what's a real world question Speaker 1 21:50 like for instance, for intrapreneurs, the signal they're getting is very noisy, and when they get a signal, it may be the case that Mars strategy is amazing and good, but it may be the case I got the kind of unluckily bad the signal bad signal. Whereas like, so by making this as a hierarchical model, it somehow informs the distributions of each decisions that I've made is right or wrong so that I can pivot well. It might make sense to Speaker 2 22:31 take it away from entrepreneurs. They have they only they weren't talking about entrepreneurship until you showed up. So what were they talking about? What kind of problems they want to address or where their tool sensible tool to use? Speaker 1 22:49 So like, like insurance company. If if I see this driver speed up more than like 50 Something per hour, what is its probability that it will going to crash? Speaker 2 23:06 Okay. That's very concrete. Okay, so, so, given driving data, I want to predict crash probability. Yes, yes. People have been doing that for a long time. Unknown Speaker 23:18 But we need that very fast. Speaker 2 23:22 Because new data is coming in very quickly or you want real time you just sped up. Speaker 1 23:28 Yeah, data is increasing. And we want to simulate different cases. Like for instance, if this car I spent more than 100 Or what happens more to add, so that we want to simulate many different cases instead of cost test to choose one we are testing like 200 and plotted so that we can have more informed decision towards Speaker 2 23:49 a simulation. Yeah. Okay, so potentially, we can simulate lots of different driving situations and understand and see what's the frequency of crash given certain driving situations given some probabilistic environment Speaker 1 24:13 and the way they do that is based on like, this has read and certain like driver's behavior that is discrete variable and this is continuous variable from like one to 10. So it has like this and let's say this, this driving is really like bad behavior. And they have like different rules for it. Either like good and they have this result, whether they crashed or not. And when we quarry something in meaning, I have this driver with a very good behavior, and it's driving more than like three. So what is the probability of being crushed? And what it does is it goes in and finds everything that is larger than like three and compute the probability of results. So there are a lot of conditioning problems coming in that is very symbolic. Unknown Speaker 25:21 It's novel that's a new way of analyzing these kinds of situations. Speaker 1 25:31 I think probabilistic programming language itself, Amis that and if you ask me, whether it's novel meaning other field band, probabilistic programming there, people are able to do that or not. I don't think they are able to do that. Speaker 2 25:52 Because it doesn't seem that different from what I would call classical statistical decision theory, Speaker 1 25:58 but it's, I think, very small clock speed. Yeah. It's very important decisions and like once in a year, that it's being done. But for those like once every seconds we need to do that like computer vision. This robot is like traveling and perceiving what's the obstacle? In the room? Speaker 2 26:24 I don't think the car accident example is as good an accident I think example. I think a better example might be the the Ukrainians monitoring the airspace, above their border with Russia. Because then you've got a huge amount of signals on a constantly flowing basis and you have to make decisions on a second by second basis based on what you're reading what you're seeing second by second. Is that a missile company? Is that a drug company? Yeah, yeah. Yeah. Does that make more sense? There's a data intensive real time heavy data intensity feature to that situation. Am I on the right track? Yes. Speaker 1 27:23 And that's, I think, like, clock speed being faster. Like Speaker 2 27:30 it's the clock speed. Of the observation decision making loop that we're talking about. Second by second. We have to monitor data and make decisions. Yeah, yeah. That will explain why the Department of Defense or DARPA is interested. In a Google same thing perhaps. Although for profit making purposes, they're collecting data on a billion users simultaneous Slee and trying to price their ads, which are priced on a real time basis. So there they have this massive decision making space. Massive flow of continuous data, and they're trying to optimize it that driver right? Yeah, Speaker 1 28:25 I thought like this when you think of like a UAV control center that has to control like ordinate 1000 cars out in the street, Speaker 2 28:35 if you start to if you're controlling a large number of autonomous vehicles, heavy traffic environment Yeah, that I understand. It's not just about one driver that driver's habits Unknown Speaker 28:54 Am I making sense in? Unknown Speaker 28:55 Yeah, well, so Unknown Speaker 28:59 if the eye becoming thicker Speaker 2 29:06 Do I understand that right, that that's the kind of problems they will address and add there and what's Is there a key set of insights about how you model or solve this problem that's that enables you to handle this volume of data with the speed of decision making or is it just just use tons of computation and storage? That is that is there a revolutionary different architecture for how you structure this dynamic decision? Unknown Speaker 29:47 That's a good question. I think. Speaker 1 29:55 The ad V, this article and there we go. Can I put a Speaker 1 30:13 differentiation? So this is the revolutionary methodology for computing gradient descent? Wondering when, like, last summer, I invited 10 core developers and to like have a conversation with Huzoor and has yours conclusion after the meeting was this automatic differentiation is the key thing that Benson does not have 10 hats. So it just computes the whatever whatever you have it computes does, Speaker 2 30:50 is differentiation. As in calculus differentiation. Yeah. Speaker 1 30:55 Not in a mathematical but computational differentiation is what's it doing building a tree Speaker 1 31:06 and they have recently come up with a way to do this like the expectation of like P following something more concretely, we would need this, for instance, in our cases, because we don't know what the expected probability would be like the profit would be at this stage, but we need to decide where to go without knowing like in precisely what the expectation is. So this gives us certain gradients. Speaker 2 31:36 So the entrepreneurs problem is feels like it's slower clocks breathe in Ukrainian Air Force is a problem doesn't have as much data coming in as fast and need to make split second decisions the same. of Ukrainian Air Force. Speaker 1 32:04 I think that depends on which level we want to intervene or guides. Unknown Speaker 32:13 Is this your taxonomy or somebody Unknown Speaker 32:15 else? I'm doing a literature review Speaker 1 32:23 with with help of each but yeah, so if we go down into atomic level, it's like writes an email to a supplier and reads a report and I think that also affects the glacier level. So yeah, I think it's whatever data you get is easy to find a clock on fast clock speed. data and use cases. Unknown Speaker 33:01 Like this big screen thank you. Speaker 2 33:27 Want to take another crack at applying computational probabilistic programming to the entrepreneurs problem? Speaker 2 33:43 So let me just go back to your view caution. Is there an application that they've made a massive breakthrough on sort of their iconic accomplishment that looked at the great thing we're able to do? Like solving the Ukrainian Air Force's problem or solving Google's problem? So do they have an iconic This is our claim to fame, but look at the problem we're able to address? Unknown Speaker 34:18 Me I think about in combat. Speaker 2 34:21 Yeah, I mean, yeah, I don't need an answer immediately, but I'm just trying to the scale and scope of data, speed and intensity for the Google problem in the Ukrainian problem just feels qualitatively different for me than the entrepreneurs problem. I agree the entrepreneurs problem is data rich, but the Google of Ukrainians that they can literally collect millions of data points and have those data points in a data set that they can then try to process using these tools. I'm not sure that the entrepreneur can collect data in the same way and have that volume and speed of data. That's relevant to them. I mean, obviously, they can collect data on what's the stock prices every second of every stock in the world or something. But that's doesn't feel like it's first order to what the what the entrepreneur has to try to figure out. Speaker 1 35:22 Okay, we think about like, absence of the tool makes it more harder. Like if maybe there should be Speaker 2 35:33 there should be a tool. I'm just I want to be convinced this is the right tool for the entrepreneur. I'm convinced Tuesday to test to choose one is a little too coarse. But the Google problem of a billion data points feels a little too fine. Speaker 2 35:59 Just can we scope the entrepreneurs problem in terms of where does it fall in terms of data intensity data click intensity and clock speed data clock speed, decision intent, like how many decisions per second do you need to make as Google 1000s maybe billions of decisions per second? If you if you've got billions of people. I mean, how many ads is Google sell per day? Probably billions because they offer these, right? So they're making billions of decisions per day. That's feels like it's qualitatively different than the entrepreneurs problem. Speaker 1 36:42 I don't think this is only a matter of speed or the size of the decisions we're making. Computationally. How do we think about conditioning right, Speaker 2 36:53 but I always say, I said earlier. So how does their approach differ from traditional statistical decision theory, you said, intensity of speed and intensity of data. And I'm saying the intensity of speed and intensity of data of the global problem. Feels like it's several orders of magnitude bigger than the entrepreneurs problem. That's trying to be convinced that this is the right tool for the entrepreneurs problem. Unknown Speaker 37:26 So I think, like Speaker 2 37:28 they said, you're going to spend your whole summer on this. I want to make sure it's the right tool. Yeah. Speaker 1 37:34 I appreciate that. So I think this term data is very relative. do you classify simulated data? As data? Speaker 2 37:48 Could it be relevant of the kind of simulation that we were talking about offer or pivoting thing? Yeah. Speaker 1 38:00 So imagine like three by three and we can simulate like different path that we can take. And that would give us massive amount of what happened if I start Speaker 2 38:10 calling out a second by second, right. We can do it one time, we can do one time analysis. We're gonna do a lot of simulations. But that's not the same as waves of users, dirt entering search terms of billions of per day. That's fresh data that Google has the process of real time. Speaker 1 38:35 I think the reason our data is not like inflowing is because we assume market is stable, which isn't. Speaker 2 38:44 So, so key, but as I said, I don't think we could get data on every stock price at every, every second, but I'm not sure that's going to be the most important data for the entrepreneur. There's a much smaller set of data I think that's probably relevant. Speaker 2 39:13 If let's say we're a pharmaceutical company pharmaceuticals. And there are many scientific breakthroughs coming through on a regular basis, right. So theoretically, we could have an LLM reading every medical paper and every biological science paper there are probably 1000s published per week, per day. So that's a lot of data, right? And so we're constantly monitoring the medical pharmaceutical, biotech literature. And that's data that's coming in that may want to inform our entrepreneur entrepreneurs about tech entrepreneur. as something that's you thinking is, is this a real problem that I'm describing? Speaker 1 40:07 My friend who was doing PhD in Political Department, just graduated after he just got the Masters he quit PhD. And that's the business's story like in the lobby with the lava data. Every day, new information is coming in and he's using GPT to process it and if they're servicing SK, they collect relevant data for SK and process. Oh, SK is Korean company conglomerate, as Kia and Hyundai are like Samsung. And they write a customized report for that. And I think and also, like dad is also interested in this business. If Phil's gonna retire soon and since he has some, because the work I was trying to do during this time was logistics. And that's very slow clock speed industry and he also mentioned people in the clock slow clock speed people business are very wants to stay feel insecure in not following up the AI. So if certain person with a reputation and trust somehow provides that technology to them, it's very fast, it can be very fastly diffused. But my point is, I think there are a lot of data that slow even though slow, crosspiece people wants to get access to that fast, evolving data. It's just there is a tool to do that. Unknown Speaker 41:46 Just go back to my biotech. So if we were to say for a biotech startup we think there's a benefit from Speaker 4 42:01 reading every paper that comes out every day, synthesizing those papers Speaker 2 42:10 and interpreting them in the context of tech startup that were interested in me that to me that has some appeal, as I could, I can justify in my mind that this is high volume of data. A high intensity high rate of data and complexity of data. Famously, biotech, or pharmaceutical companies are slow clocks. They take many years to develop a drug it's because of regulatory things. But Speaker 2 42:56 and I think this will be a tool of wonders will be until more interesting to a big pharmaceutical company that a little, a little so what are they going to do with all that information? Right? That is you're gonna tell me to crunch all the data and my ideas not worth anything anymore. But all I gotta do I Speaker 1 43:21 can learn about which companies are willing to m&a there. This is true technology Speaker 2 43:28 that we could we could paint a similar picture in the semiconductor, maybe. Think about biotech. The variety of solutions, the variety of technological strategies, seems to be much, much bigger. In semiconductors, semiconductors, you're making chips is different chip architectures. You want to get smaller and smaller and faster and faster. But biotech there's there are 1000 human health issues that could potentially be addressed with a million different strategies insurance. And biotech just seems like a bigger space, bigger information space. And more complex you're dealing with in semiconductors you just dealing with physics. Yeah, biotech, you do. Like physics and Unknown Speaker 44:23 biology and human behavior. Speaker 1 44:30 So one of the Korean VC in Boston for doing biotech says the reason Korea is not able to do a lot of biotech it's not very, as active as Boston is because there is no way to do m&a. Most of them just go to the end to end commercialized IPO. But us Boston has a lot of m&a in between. My point is the more segment the market is meaning the more valuation is happening in the middle, which includes a lot of data processing, quantification, I think, yeah, that's the key. This is clearly a nice answer on where real time analysis will help. So prioritizing which drug can be used to advance for the next stage of testing, identifying new promising drug targets to pursue based on emerging research, like they need to make a choice between which like patient segments or rare disease or just and also they need to do research about what are the existing drugs are and how my drug is better than that they have to make a point, adjusting doses or delivery methods of drugs based on new data about efficacy and safety, deciding when to pivot or terminate research direction, they're no longer promising based on new external data. So I think all of this are conditional like distributions. Speaker 2 45:57 So what do you think about the projects for the summer of an addition to our little simulation model, building like an LLM for reading, medical, biotech literature and synthesizing. I'm trying to think if you want to be able to prove to the entrepreneurship world that this is an approach that has value for entrepreneurs. I think you need a proof of concept to show look built, look at how this now and then be able to say maybe the entrepreneurs don't want it but maybe the venture capitalists that's just as good as far as Yeah. Speaker 1 46:50 I would need the capability thing at some point of my life and I'm all for it. Speaker 2 46:55 Well, you said you want to devote your summer to probabilistic programming. I just proposed problem too. Speaker 1 47:02 Yeah, yeah. So my question is between semiconductor and biotech, do you think by type as well, Speaker 2 47:09 so based on what analysis and gave you my my condition is it's a bigger space now. Are you gonna go the semiconductor thing? Speaker 1 47:24 So 10th or something? Yeah. The last session was very interesting. No, it's June 1. But the simulation thing was also there for education. So that somehow was interesting. Are you going Unknown Speaker 47:46 to five to seven. Unknown Speaker 47:48 So Speaker 4 47:51 it's complicated. I was gonna sit below let's Class A whole week, week long class. So I may do some times timeshare. Speaker 2 48:04 What's the first day? No, it's versus the fifth last one. Unknown Speaker 48:10 So Bill's thing is the third Speaker 2 48:17 so it's just like, if we let's suppose let's look at this with the with the context of if we wanted to understand if we wanted to think about building a probabilistic programming tool to analyze semiconductor industry data and technology data, to give real time, power to a company in this space, which of these sessions will be relevant? Unknown Speaker 48:47 So let's just do the exercise. So Unknown Speaker 48:52 this first one is about environmental regulation, Speaker 1 48:54 and also tech new technology. They have some new purple thing like PFA S is a new technology. Speaker 4 49:04 Yeah. This is this nasty little pieces of plastic that are in our bodies. Unknown Speaker 49:11 They want to get rid of them, it seems. Speaker 2 49:20 This weekend, it's pretty scary. Pardon me. I said I read an article about it this weekend. It's pretty scary every mammal on earth has these things in their body already. In Antarctica, the Arctic was everywhere. Unknown Speaker 49:40 We want more because if they can filter them out Speaker 2 49:46 Okay, so this is about environment here. So this is more about manufacturing and product. It's first Speaker 2 50:00 combination of product and process here Unknown Speaker 50:07 I suppose sustainability Yeah. Unknown Speaker 50:11 Yeah, emerging new technical challenges Speaker 4 50:23 this is about physics. Your 2d animators can we have optical io chiplets 2d animated. Or whatever? Speaker 2 51:00 Boy, thanks good tech, technical and deep and complicated. Unknown Speaker 51:05 Yeah, only thing that I was able to understand. Unknown Speaker 51:29 know Randy and also for a long time Speaker 4 51:37 first knew her when she was a grad student. Unknown Speaker 51:45 was new. I didn't already do cheese in material sauce. So do you recommend me going there? Speaker 2 52:07 I'm not sure I there. The question is, what fraction of what they say is going to be comprehensible to you or to me? I mean, I have no more exposure to this technology than you but Speaker 4 52:23 it's my technological knowledge is at least 10 years old stuff. Speaker 2 52:33 So to answer that question, Is it crazy to try to do what I'm suggesting. That is for the summer to build you're going to learn probabilistic programming to have a an entrepreneurship problem in my application that that requires the computational power that we're talking about. And my perception of what you've told me about probabilistic programming, is it's bringing massive amounts of computational power to a problem that's data intensive and speed intensive. So if you're going to justify that kind of a tool, you know, why are we bringing a bazooka to a birthday party? Well, we need a bigger splash than just the candles. We're bringing this incredibly powerful tool to the entrepreneurship literature that says, Test to choose what right and you're talking about test and test billions, and choose from those billions, right? That's what Google's problems so you're gonna have to be convincing. Right? And I think with an example of a look how much more powerful this is that test to choose one or something? Speaker 1 54:09 Yeah, so I guess we need to decide whether we're doing sequentially or like parallely meaning the simulation model and I introduced to this my PhD from chemical engineering, he said, If I were you, I would set the low bar as the interest rate and high bar as visit the previous database of like 10 years before Woodward companies in your segment, what they're like kind of the benchmark was and take the average of them like top 100. That will be the high bar and I think that requires some access to external database. So So yeah, there's two parts the basic logic of the simulation model and also the data in and out model part. So whether we move on to this parallel or sequential making sense yeah. Speaker 2 55:21 We have a big agenda. Okay. So let's go to your dashboard. Okay. Dashboards, Unknown Speaker 55:31 why dashboard is here Speaker 1 55:45 so this part is finding like three beta entrepreneurship evaluators that if I bring up Andrew colic they'll gonna say whether this is useful or not, so Aaron's caught and Amir perhaps Unknown Speaker 56:01 you're gonna bring what to them? Speaker 1 56:03 What's the birthday party in the pros program? Bazooka bazooka okay. I wouldn't surprise then I have divergent. Speaker 2 56:18 Okay. All right. So we need some customers to test the bazooka. Yes. Okay. That's all right. Technology and PC that's learning probabilistic asset and industry diversity. Speaker 1 56:35 So I have like two or three biotech related feces and also entrepreneurs who can give feedback about what information they need. Yeah. And I think setting a frequency of like between this customer technology and practitioner would be very helpful. resource allocation. And by the so my first quarter is until June 24. And I, I hope to somehow finalize, kind of, like what research proposal team would be by then but do you think due to problems due to you, oh, this is just other than this myself. Unknown Speaker 57:27 What's TF. Speaker 1 57:33 O says yeah, this is fun thing. So I was trying in Singapore, I was trying to simulate different combinations of how far ahead I should look into the market and how far I should look ahead into the technology. And the conclusion I made was, it's better to look ahead in terms of technology rather than a market. I had a lot of time so I tried to mathematically so like, what are the intersecting if the rays are coming in different angles, or the insert intersecting and what are the like uncertainties? Yeah. So that's mot F. And CMO is CEO, T, CTO and F CFO. Unknown Speaker 58:21 Alright, so Speaker 2 58:27 so you've got June 21. as sort of a cutoff point. Leaving for Asia to 22nd Speaker 1 58:33 I tell you that I'm to Singapore, or I'm Speaker 2 58:37 gonna go to Malaysia first for to visit with friends to go to Singapore just for that week. The rules about how long you're in Singapore are complicated. I can only be in Singapore for six days. Or else I need employment paths. They don't have enough time. So I've just said the conference is the eighth to the 12th or something so I'm going to only show up. I'm gonna go to probably a blazer for two weeks. Anyway, so leaving the 21st So the question is for me. Is there a better talk I can give on our simulation model on the 21st on the talk that I gave a couple of weeks ago. And if what do we need to do on her give me a better talk? Because I really want to I'd like to change his mind about funding I want to show that we have something that's useful and valuable. And so I want to I would like to be able to push this ahead Speaker 1 1:00:00 so that's I think, when I say the parallel part, the brain part, I hope by the June 21 There will be some meaningful advances and I need your help in like what would be some interesting thing that you would be proud to present. So Unknown Speaker 1:00:21 I think we should meet again, maybe that's Friday, Friday, Unknown Speaker 1:00:26 Friday to have to finish the sport. Yeah, Speaker 1 1:00:28 it's fine. Like we have two or three things. I don't think Friday is the first Speaker 2 1:00:37 okay all right, so let's meet Friday. I assignments are right down the Optimize for this and then think about deliverables for a second. And to fill out the form. Speaker 1 1:00:58 And like around sometime in June, do you think you I don't know like I hope you and me and Akash might want to have some chat about because we're putting a lot in stake in the promise of computing. And the question you asked to me last time, like, what's the vision of the promised computation lab for the next like three or four years and I think that would affect my dissertation a lot. So, but I think it may be better if the question or query is coming from you rather than me. Speaker 2 1:01:34 Okay, thanks. Yeah, you said to me Speaker 1 1:01:42 do you have three or four minutes? Yeah. So let's say Malaysia model I'm building was kind of so I'm fixing a little but it gets certain parameter combinations of how many opportunities it has and what are the perception? So I think this is very important. So during the chat was caught, so we had kind of discussion and I agree with Scotts saying that people usually think for the startups the best signal is from customers. But it's usually from a co founders. It's the best signal when they're starting. It's because different abstractions is coming in. Because customer don't see me they see the product, but co founders see me and my production visibility and all that. So it's like I'm here and the first concentric circle is within the co founder and the next is investor. Investor don't see me they see this business see me let's say we are do Unknown Speaker 1:02:59 we as the entrepreneur, yeah. Speaker 1 1:03:05 We have the entrepreneur and we have different circles coming in. And customer is in here and investors in here, customer sees this whole thing as a product. They don't see the person or the producer inside who made it. And investor see this as a startup business, and only the co founder can see like this as a person. So my point is, when I decomposed them are we on the same page? So I just for an example, like I think the reason that you are more generous in my deliverables benching pie is because you see my kind of how, for instance, hardworking I am and like I've tried many different things. But Jeonhwa is kind of busy and for him, I like only the interface between me and him as a product. So I think there is some gap in whether the perceived quality of the team is and I'll feel free to I Speaker 2 1:04:13 was gonna say the cold hard reality for an entrepreneur is that the customer is either going to accept or reject the product because they might say, Oh, that entrepreneur is a really nice person, but their product is lousy. So I don't want I'm not going to buy their product. Right. Now, they might say, well, because I think the entrepreneur is very capable. Even though today's product is lousy. I'll try the next one because I think they're capable. But ultimately, they do judge the product, not the entrepreneur, right. Speaker 1 1:04:45 Yeah, but it's I don't know. Like because we are in the early stage. Like this is from Ilya straw, lo. He's a Stanford professor who recently wrote venture mindset. And, like team performance. Yeah. One feedback I got from my model is from UAE, China, she was early stage they don't really care about the profit. They care about how well this can be. Get some investment, and investment needs, like Team factor is very important. So my point is, it's not only about like how the prof product itself is competitive in the market outside but it's like how compatible or mean competent that team has. So that is what I was trying to capture with the new A the red part, like in our asset part. So based on that like this is the story I came up with for the AI company. And I think the like example of where the perceived quality can drop is when talented CTO resigned for instance for like open AI and the reason I brought this up is unlike the product and the market kind of gap between AI and non AI market, sorry product and B to see b2b and b2c market. It's a gap between the two so like the assumption is we want to get here and we're here and if mu B does mu B, this is the real value. This is your and mu C C, B hat and mu B zero. So this is ground truth. And this is something that's only the entrepreneurs know. And my like last line is, I think for this perception of the team quality, what enterpreneurs believe themselves, I think that's the ground truth on like the market and the product gap. So this is this part, how much B to B market is profitable then be to see market is what they learn as they get the signal and how much B to AI product is better than non AI product is that what they learn as they get the signal, but new a itself is if they believe in I think that becomes a truth. So I wanted to somehow have some conversation with you on this part. Speaker 2 1:07:35 While you're talking to us a little thinking a little bit about Elon Musk. Okay, so you might say Elon Musk has a very strong team. He's as some people say, no matter what Elon Musk has as a product, I'm gonna invest in that company because he's a strong team. But he's got some wins and some losses like he's got the boring company that drills tunnels that didn't really go anywhere. He's got a Tesla, which is sort of fading right now a little bit. He's got his rocker company, which is doing really well. He's got his AI company, which is he's hoping will do well. But I think ultimately, yeah, he's a talented team. If you will. But people are going to judge the products individually. They're not going to buy it. At the beginning. They're buying Tesla's just because they thought Elon Musk was smart. That phase seems to be over now. And people only buy Tesla's if they think the product is good, is it desirable? product and they buy the rockets because the rockets are really good out there. So, so I agree that the team is irrelevant. But I think the market ultimately may be enchanted with the team for a while but if the product doesn't deliver, then the team fades and importance. Speaker 1 1:09:05 I agree with that. But wouldn't there be a causal loop between making people believe in our team would increase the talent coming in and that would affect the final product? Speaker 2 1:09:17 Yeah, that's certainly true. If you if you're successful, you attract more people who are high capability people Unknown Speaker 1:09:35 but what's that got to do? Speaker 1 1:09:39 Oh, so I was just making an example that Jim was on the outer circle. So you using me and Jane having me or a little abstraction Speaker 2 1:09:51 is different that's very symmetric information. Sorry, if Speaker 1 1:09:55 I confused like, yeah, my point was new a I somehow revert to and like change this a lot. But unlike mu C, and mu B where I had the estimation of entrepreneur versus real true ground truth, I don't think new A has some ground truth. So this meaning the whole model for now is like five parameter model, or two to one. That's the plate parameters paste I'm searching in here. So like this is the example of how mu A changes and this dotted line is mu zero. Sorry, no price. Speaker 1 1:10:48 So this 0.2 is a ground zero ground truth as your point one is ground truth and depending on like, where it's torched, from, they somehow based on the same signal in certain cells, they update their belief on their perceive capability wherever they are, and this gap and this gap Speaker 2 1:11:17 Okay, I'm gonna have to quit. Thanks. I want to get back to my desk. I'll send you a calendar for Friday. Unknown Speaker 1:11:26 Thank you. Do you agree productivity is twice with the mix can you just sign I wouldn't submit before Friday but lifting all right. I like your shirt as me