The Future of Artificial Intelligence Startups & Venture Capital (Webinar)

Hi there, everyone. My name is Adeo Ressi, and today we're going to be talking about AI, machine learning, and robotics. Basically a Terminator movie. So it's going to be an exciting panel here to answer questions from all around the world. Before we get started, I'll introduce myself, the Founder Institute, then bring our panelists on screen, and we'll jump right in. I'm the CEO of the Founder Institute. We operate programs in over 200 cities helping people pursue their dreams of launching technology businesses. Right this moment, we're enrolling in over 50 cities around the world. So if you are thinking to start a company, whether it's AI, machine learning, or robotics related or not, you can learn more at fi.co/startup. Join, and there is a link that will be posted in the chat. Speaking of which, we have an AI robotics program that's going to be kicking off in Pittsburgh, and in fact we have some of our panelists from the Pittsburgh area here today, so I'm excited about that. So if you're in the greater Philadelphia, Pittsburgh area, or frankly the East Coast, and you're thinking of starting an AI company, maybe go going to Pittsburgh is a great idea. It's certainly got a lot of the expertise and talent there. So you can learn more about that by going to fi.co/join/AI, capitalized, and that link will go in the chat as well. So we're gonna be talking today about obviously AI and its impact to you, and we have a really great group of people who are going to be helping guide the conversation, including yourself. So what I want to do before I introduce the panelists is say, you know, you may be thinking of starting an AI-related or robotics-related company, you may have actually already started one, and we've got a great group of people here to help you with any questions that you have. So what I'd love to do is if you have anything that you want to discuss, and we— don't be shy. There's a chat area. Type in the chat your question. My team will take every question that we can and put it in the agenda, and I will try and get them all answered. So really, this is not an opportunity for you to be shy. If you have a question for an individual panelist or in general, Get them up on the chat, and we'll get them on screen. Without further ado, let's start walking through our esteemed panelists. First and foremost, Jeff Schneider, who's the former engineering lead at Uber Advanced Technologies, but a current research professor at Carnegie Mellon, a school that I attended the architecture program at. So I spent some good time in Pittsburgh. I love it greatly. Jeff, would you want to say a couple of words on your background?

Yeah, yeah.

Sure. So my appointments here are in the Robotics Institute and Machine Learning, and my research is primarily in machine learning. And of course, I'm very entrepreneurial as well. As you mentioned, I'm just returning back from working with Uber on self-driving cars for the last 3.5 years. But prior to that, I was also working with a biotech on a machine learning-based drug discovery system. And even before that, I co-founded a company applying machine learning to manufacturing. Processes.

Wow. Well, welcome. Excited to have you. Next on the list, in no particular order, we have Michael Harris joining us as a partner at the Robotics Hub. I believe your partner in the Robotics Lab is actually having a baby around now. So we were originally going to have both of you, but now we only have one. Makes for a tighter panel. Michael, welcome. Do you want to just say a couple words about your background?

Sure. So, the Robotics Hub, we're largely Pittsburgh-based, investing in early-stage foundational robotics and AI companies, and including other foundational technologies that lead into the smarts around robotics. I happen to be based in Silicon Valley, but the fund was started out of work that my partner Chris Merlo was doing at NREC, out of CMU. So we're very heavily oriented around that Pittsburgh area, about 60% of our investments there. My background is in AI and in running both in entrepreneurship with my own companies, but also in running startup accelerators for Citrix and other organizations. So I've touched on a lot of different areas around taking deep technologies or complex technologies and working out how to effectively bring them to market. And I'm really looking forward to today's discussion.

Maybe they'll bring themselves to market. Speaking of bringing things to market, that leads last but not least, again, in no particular order, Bert-Paul Sockte. He's from Philips, looking to bring some of these technologies to market, particularly around healthcare. Bert-Paul, do you wanna just introduce yourself and add any comments about what you're looking to bring to market? Sure.

I'm leading an advanced technologies group within sleep and respiratory care business of Philips. And we are Pittsburgh-based company, actually started 30, 40 years ago, and Philips bought us about 10 years ago. And we have mainly related with sleep and respiratory care, and we have been doing sleep apnea devices, then respiratory care devices. And now we are venturing into other sleep and respiratory areas like insomnia and restrictive sleep and other stuff. So we are trying to really use AI in the areas of consumer interaction, even using for our customer support systems and other areas of signal processing and data analytics to make our healthcare systems better and solutions better. So that's where we are. I've been with them for 12 years. Before that, I have worked in other areas in telecoms and train control systems. I have— my background is software background, actually.

Great. Well, we have a lot of people, 400 people currently and more coming in every second. This is great. If you want, please share it out because we want to make this as big as possible. And a reminder for everyone, we got 3 topics today related to AI, robotics, and machine learning. We're going to be talking about recent developments and opportunities, the challenges facing startups in these areas, and the funding environment. Now, if you have any questions related to those topics or otherwise, pop them in the chat and we'll try and get to them throughout the conversation. But let's kick it off with recent developments. You know, open-ended question, anyone can go first. Why don't we do it in reverse order maybe to make it a little easier? The first one, Virpal, you can start. Like, what do you think some of the newest applications that you're seeing that fascinate you in AI and robotics?

Well, I can talk more about healthcare. So that's where my concentration is. Sure. I definitely see a lot of diagnostics in which, especially in the healthcare, not only diagnosing the— one of the biggest thing is the radiology area, which has already been— everybody has heard about that one. But beyond that, actually trying to understand and predicting diseases and issues. And those are some of the areas. And not to forget about nanobots and all that for surgery. Robotics in surgery is another thing. Though I don't think Philips has gone into that space, but all the other areas where we are talking about even therapies, using specific therapies on basis using predictive analytics, That's, I think, one of the biggest impacts I see, which is coming in, and not to mention personalization and individualization of medicine, whether it's medical devices or pharma. So I think those are the biggest areas of the new things we have seen, and there's a lot of push in that area.

Well, Jeff, you've certainly done some work in this before, as you mentioned in your bio, so why don't you go next Max, maybe you can add on to that and go further in some other areas and we'll wrap up with you, Michael. Sure.

Yeah, absolutely. I can comment on some of the work we did on machine learning drug discovery, actually. We were testing for CNS drugs or psychiatric drugs in mice. So we built a mouse cage full of sensors, cameras, 4 sensors on the floor, all kinds of stuff. And we tested drugs in the mouse. Basically, the algorithm had to look at the mouse and say, does that look like a mouse that just took an antidepressant? And so on for the various other drugs. What's amazing about it is that company's called Psychogenics. They just had their first big Phase 2 trial success where they found an antipsychotic that, that's, that's a first-in-class drug. So it's a unique mechanism. So it actually did what it hoped for because it wasn't The discovery wasn't driven by the particular mechanism in the brain. It was just driven by the behavior. They actually found one that works in a completely different way for the one side.

Well, that's what's great about AI and/or machine learning and some of these human health-related matters. Like, doctors are great at pattern recognition, but they only see 3 or 4 patients with some conditions. You know, you can have a machine learning algorithm see hundreds of thousands of patients and make the discoveries. What are some other areas, Jeff, before we go on to Michael, that you're excited about right now?

Yeah, sure. One of the things I'm excited about, and this really comes up with a self-driving car, this is becoming a theme everywhere, is how you do safety and verification and validation of AI systems. Traditionally, that hasn't been an issue because you've been— they've been used to do things like post ads on the web. And the worst thing that could happen is you post a more annoying ad than you would have otherwise posted. But now that we—

which happens all the time.

And so, but now that we're looking for these systems to really do safety-critical, uh, uh, work for us, uh, how we validate the safety of these is becoming a huge area. So I think that's a big thing coming in the future.

But isn't that an ethics consideration in some regard? Because like, who's at fault if the AI hits a person? Is it the AI, the car, the driver?

I mean, yeah, I think there's definitely a big ethics consideration there. I think ultimately the sort of liability there will follow the same path as it has for many other industries. You know, for example, you can imagine it following the same path it did for airlines. When an airplane crashes, well, maybe there's some liability for the airline, maybe there's some for the airplane maker. I think ultimately that will follow a similar path, even though it's a much more complex system with AI involved.

Michael, let's turn it over to you. What are some things you're seeing in the future today that are exciting you or maybe even scaring you?

Well, I look, there's a heap out there. So just to continue on that ethics side, right? So I think that really comes down as much to saying, Here are the things that we're expecting engineers to do up front, to say their data is realistic, to test corner conditions and all those sorts of things before they start, let alone going back and doing that post hoc exploration. So there's some really interesting things going on there that change this from just being an engineering tool that we're now starting to use, machine learning, into something that is a net contributor to the way we're thinking about our technologies across the board. Exciting new technologies that are emerging. So last mile delivery, one of the companies we're working closely with is Agility Robotics. This is a two-legged walker that is basically biologically inspired. So it walks the way that, well, in fact, the way the chicken walks, but it balances in the same way that we balance as we walk. It regains energy as it moves along. So that becomes really interesting when you're starting to talk about doing last-mile delivery for things, or indeed, you know, running in front of convoys where you're concerned about improvised explosive devices, right? Something that walks like a— or at least would be detected as walking like a human being is really useful in those sorts of situations. On the healthcare front, we're working with a company that's doing some fascinating work on pancreatic disorders, right? So, so being able to mix together genetic variations with all of the information from patient history and then do a combination of machine learning with older school AI, rule-based AI, to say what's the latest in medical knowledge from these different areas that can be brought to bear. So we're seeing things in last-mile delivery, in healthcare. Another area I'm really excited about is food security, right? So we've heard about these vertical farms. There's one operating out of Pittsburgh at the moment called Growbotany, where we've been able to apply the best of automation to be able to say, within this warehouse, we can grow leafy greens and deliver them within a matter of hours of being harvested to Whole Foods and to some other local supermarkets. So that becomes really interesting if you start saying it's very expensive to get food from one side of the country to the other. Let's just grow it locally. Let's make that work. Let's do it at a level that improves food security and does it in a way that is sensitive to the amount of water available, the amount of power available. The last one that I'll add— Yeah, I mean, it's really, really— I mean, we're seeing these things across the board. There's really interesting applications, robotics and AI touching all elements of our lives. And there's new foundational technologies that are emerging at the moment that get absolutely fascinating.

I got so many questions from the audience, guys. So I'm just going to— We had questions, as I mentioned, they're going to come in. So let's— Chris? Parrises, hopefully I pronounced your name right, had a question that's kind of a good one to open with. Lots of people overuse the term AI in their products, thermostats using 100-year-old PID algorithms. How do you separate the wheat from the chaff when evaluating opportunities? I mean, with that, I could see any one of you taking it. Why don't we give— we got a ton of questions, so who wants to go with that one? Jeff, you seem eager. You want to go?

Yeah, sure, sure. I can comment on that. And it's tough right now. AI is such a buzzword that any sort of team interested in their own marketing is going to slap AI on it. And unfortunately, you know, there isn't an easy way. I, in fact, do this sometimes for consulting is just help people vet other efforts. But it comes down to asking them, AI, what, you know, which algorithm are you talking about? What's the actual algorithm you run? And that's where you might find out, hey, it's a PID controller. Most of the time, what you'll find out is it's some kind of supervised learning. And, you know, if it turns out it's some kind of supervised learning, you say, well, okay, I guess it fits the mold. That is the most common one.

But let me ask you, there's a lexicon issue, is there not? I mean, right? Because everyone's like, I'm doing AI, and I'm like, is it machine learning? And they're like, well, kinda. And it's like, You know, there seems to be a lexicon issue. Is there— do you guys— are you seeing that as well?

Yeah, it's completely a lexicon issue. Basically, AI is no longer a technical term. And I think we just have to accept that's not a technical term. And when people say they're doing AI and we want to know what they're doing technically, we just have to ask again. We just have to say, great, so what are you actually doing technically?

But I just want to confirm with the other two guests. I mean, when someone says AI, you kind of roll your eyes and then you ask that next question. Am I right in saying that? Yes.

Yeah, yeah, I'll go with it. Yes, yeah, definitely. Yeah, everybody, to be honest, everybody, not even other companies, even in our company, tries to say, I'm trying to use AI. But then you say, what tools are you applying for? What problem you're trying to solve? Why you're using? And I think that's where everything falls apart. And I think Jeff was right. You start to ask those deeper questions and that's where you find out, yeah, it probably is regular algorithms you're using to do and calling them AI. Yeah, yeah, it's become a buzzword.

So it's a buzzword. It's a requirement. So if you're coming up with a startup and certainly if you're pitching in the Valley, then I need to say I'm using AI, I need to say I'm using machine learning. Now that's reasonable. It's just another technology that gets applied. It's pretty impressive, it does new and interesting things. Really, the thing that separates the wheat from the chaff for me is, do they actually understand— it's the same with any startup. Do you understand the domain? Do you understand why this makes a difference? For your company, have you solved the problems? Is there massive technical risk? You do keep coming across people who say, I'm using— you go from AI down to machine learning. They're using machine learning, but they haven't actually done it yet. And they haven't demonstrated that it really works yet. But we've got this great startup, and it's all going to work really well. So take the technical risk out first, and that gets really interesting.

So let me go, Miguel. We have tons of questions, so I'm going to try it. Maybe we try and move through these faster. So Miguel was interested in— how ML, machine learning, could be used in, let's say, an area like UI design. And I don't know if you guys have any ideas on that, but that's a— I'm seeing machine learning and AI used in very interesting areas. Anyone want to take that one?

I think many of the internet companies have already done one version of this, which is they use it to optimize the layout of their pages. Yeah, so they, they actually just do experiments where they tweak the, the visual presentation of the page and measure the response they get. And so that's already a prototype for that kind of application.

Yeah, there's also been people doing aesthetics using basically computer vision. So you could go and say, I'm going to look at a million examples of successful pages that have been pre-ranked, and now I'm going to have an algorithm that can go and look at new pages and say, hey, that's a good one. Well, that's a bad one in terms of aesthetics, in terms of colors, in terms of really any element that could be represented in that bitmap.

All right, let's move on to some others. So there's a bunch of regulatory sort of open-ended questions like, um, so, uh, José Smithish— I hope I pronounce your name right, José. No offense if I didn't. Do you think AI would eliminate the human errors in diagnosis now? And I guess this is directed in part at you, Jeff. If so, how to handle the administrative approvals and handle the government aspects around it, governance aspects around it? Anyone want to take that? That's a tricky one.

Well, I'll just very quickly, I think, for the application of diagnosis. For the application of diagnosis, I think we're gonna use AI to do these diagnoses, but we're a long way from getting out of the mode where that is handed as a suggestion to the doctor and the doctor sort of approves or not. I think we're a long way from the doctor being out of that loop, which makes everything easier from a regulatory point of view, right?

Would you agree with that, Purple?

Yes, and whatever we are trying to do is we run through clinical trials and we never actually claim that we are suggesting to the patient or person directly, hey, you, it's an aid, a tool for people to use it to give them the information. And then they use that information to either it's a doctor or clinician uses that information and use their judgment to finally make the call. It's not the system which will make the call. Like Jeff said, it's not there yet. It has to go through a lot of other work in the future before we get to that point.

William Miller was asking, under what societal, economic, political conditions would the momentum of AI tend to be hindered or slowed down, and what conditions would it thrive? I'll leave that to the experts. Hindrance, anyone?

Well, hindrance, I can talk a little bit about hindrance. I mean, I can tell you, with the privacy and security, that is different. It's a kind of a hindrance for, for trying to run fast. But on the other hand, it is, it is required to. So we deal with that almost every day, especially with GDPR, get a keeping data, how long you keep data, how we go back to erase the data, you can't use this data for this. Those are those regulatory challenges and privacy challenges, and for many reasons they are important challenges. And we need—

well, and those are the reference data sets, right, that you need to train the machine learning algorithms at least. And, and I hear that California's equivalent to GDPR is going to be even more stringent. So anyone else want to—

yeah, so I'll add that, you know, with an aging population and the reduction in the availability of people to do work. We're an aging population. We need to be able to capture expertise. We need to have ways of doing these things. We need to have ways to automate work. So we're seeing this with the level of interest in, certainly in robotics, in a variety of countries around the world. And we're certainly going to see it with AI as well. I think that as the problems get harder, as we start to learn what we can do with medicine with automation, I think that there will be a much more active engagement and an effort, a stronger effort to get around some of these regulatory challenges simply because we need these shifts. These are really valuable technologies that will change the game.

So, you know, José asked another question, you know, how AI initiatives can handle the protectionism from various governments. Governments, but I'd go another step further. I would say that governments are seemingly clueless in how they're managing AI, and maybe you could correct my mistaken fallacy here. So back to William's question about what conditions would it thrive, I think, you know, we have a lack of regulatory understanding, and I don't know if you feel differently or if that could help or hinder. What are your guys' thoughts on that?

Yeah, I think the regulations are sort of lagging behind the reality for sure. And you could use the word the government's clueless, but I think, you know, a better— probably a better way to phrase is I don't think any of us know the right way to do this yet. We're all trying to figure it out together.

Okay, especially the government, right?

Especially the government.

Exactly. Fair enough.

And so, yeah, so obviously there will be a hindrance if any sort of application gets way out of whack with regulation. You know, there's substantial bad events that happen. You could argue some have happened in the area of privacy already.

Well, so the flip side of that is we're seeing these technologies be adopted really broadly and in ways that are very pragmatic at the moment. So, there's a rise that's occurring with the application of these technologies where they don't conflict, right? So, for example, Aerial Precision Medicine in Pittsburgh, their solution is being offered as a diagnostic test. So it's not trying to change the way that medicine operates. It simply lines up with the expectations of existing physicians and institutions, right? So, we will see more and more of these things that fit within the current regulatory environment that will happen regardless.

All right, we want to transition into building, one second, and I, I, I, I, I just don't want to forget something you said in passing, Jeff, which is important, that AI has already probably had some privacy violations. Maybe we'll come back to that a little later. And I agree, I mean, the search for reference data sets is a real pain. I know a lot of AI startups. But before we transition into building and come back to that, Adam had a question. Adam Pistorowski, sorry Adam if I messed that up. What kind of problems is AI least suited to solve or can't solve at all? So like, where do you think, hey, if I have a business and I want to do something, don't use AI for it, like what would be some areas in your experience? Art, it's a hard question.

I, you know, I would certainly— one of the drawbacks for AI systems is that they often end up being black boxes that you can't tell what's going on in there. So, you know, any application where you're to solve it without AI and it's important to have an understandable solution, you probably really should stick with that.

Interesting. Michael?

Person-to-person contact is still an unsolved thing. So we can do a certain amount with artificial agents in terms of making them likable. There's some fascinating work being done out there. But really, that human interface, there's no reason to get rid of that anytime soon.

I also— Google got in a lot of trouble at their, their last, uh, dev conference when they tried to impersonate a human being and people got like seriously angry. They're like, I want to talk to a computer. Why? Uh, Burpal, what— do you have anything you want to add there?

Um, nothing more than that. I will still think the doctors are needy. So some people who say that we will get rid of the doctors or something. I think that's not going to happen in, at least in my understanding, at least my knowledge-wise, it's not going to happen very soon or never in the future. There might be some areas which might need to shift here and there, but overall, I don't think we're going to get rid of doctors anytime soon. So if anybody out there who's a doctor is worried about their profession, I think you're safe.

Well, Charles asked somewhat of a related question, you know, how soon will an artificial intelligence model using natural language processing be able to perform as well or better than a human in a specific scenario? For example, taking an order at a restaurant through audio, just sort of the Google thing. Do you have— I mean, what are your thoughts?

There's probably any number of scenarios, right?

So, so now, so by that, Michael, you mean probably today an AI with good natural language processing can beat humans in many scenarios?

Yes, yes, exactly. And, and so we've—

but there was a— sorry, Stanford just did a thing where they tested like AI versus human recognition of people with bad, you know, pronunciations, and AI just crushed them.

So they're very good at these narrow-slice problems. However, as soon as you get beyond that, we run into challenges. So narrow-slice problems, I think we're there, totally there.

Give me an example for Charles. I know what you mean, but like ordering a pizza. Yes?

Yes, absolutely. Ordering a pizza, something within a very constrained domain. Think about Siri. Think about the things that Siri does well. And then think about how you'd really like to have a chat with a friend and talk about anything, and shoot off and talk about random things, and expect them to remember what you talked about yesterday. The degrees of freedom in a regular conversation are well beyond what's possible. But if you're talking about a constrained, hey, I'm ordering a pizza, I'm getting a haircut appointment set up, or anything else like that, we're well and truly there.

The repetitive stuff. I think that's what you're trying to do. There's a lot of repetition of same thing, same kind of discussions over and over again. That's an easy, like ordering a pizza. You have so many toppings. I hope you're not going to add something weird on it, but still, yeah, the caramel pizza.

Ah, it doesn't compute. What are you crazy? All right, let's talk transition into building for a second. So Jonathan Singh was asking, is there preferred programming language that you guys recommend for machine learning? Anyone want to take that?

Sure. I mean, a lot of people work in Python now, and that's their— that's one of their favorites. What's interesting is increasingly a lot of the work is being pushed into the modeling languages. Into TensorFlow and PyTorch. And, you know, the amount of things you can push straight into there keeps growing. The other thing that's on the rise, I think, are the probabilistic programming languages. And so people are really moving toward those as a better way to do AI applications.

So, Jeff, just to clarify, as a language, you're saying Python, but But almost definitely like pick the appropriate framework for the problem you're trying. I'm sort of re-parsing your answer because you mentioned some frameworks there and pick the appropriate framework for the problem you're solving. Is that what essentially you're saying?

Right, exactly. Just to give an example for people who aren't familiar. So Python is a programming language, just, you know, very similar to any of the programming languages.

C++, et cetera.

Yeah. Right, exactly. TensorFlow was originally built to specify the architecture for a neural network and to automate the training and inference process in a neural network, but what's happening is that language is being expanded to be able to specify computation more broadly.

Okay, so it's like a framework. Guys, do you want to add anything? Purple, Michael?

The only thing I'd add to that is I think it's been really interesting seeing— so totally agree with what Jeff's saying in terms of general machine learning. For the people I've worked with who are revisiting earlier flavors of AI, it's been surprising seeing people talk about languages like Prolog and Lisp from back in the day. But they're building large-scale systems. Out of these approaches because they're enabling, because they're frameworks that work for the types of recursive knowledge-based data that they're working with, particularly for some natural language understanding types of problems. And one fellow who's using Lisp to apply some ideas from quantum computing to doing pattern recognition, which is just mind-blowing stuff.

I saw a question from quantum computing, and we're gonna get to in a moment. By the way, for all the people streaming questions, trust me, I've got a big list. We'll try and get to them all, and they're growing. But I did see a question on quantum computing. It was related to some of the IBM work that was done in quantum computing. If anyone wants to add, thoughts about that? You know, what do you see the role of quantum computing with respect to AI and machine learning, at least in the short or medium term?

Okay, go ahead, Jeff.

Yeah, I was just going to say, you know, at least from what I've seen, it's not there yet. I mean, I have not seen substantial applications use it in any—

Not yet. Okay. And Michael, anything?

I think it's close. I'm seeing— so I've been spending some time with a group called Creative Destruction Lab in Toronto.

Yeah.

And they have a business accelerator. They have a stream on AI. One of the streams that they have is on quantum computing or quantum machine learning. And there's flavors there of being able to even use the existing D-Wave machine to pre-train.

Yeah, Brian just said D-Wave. Yeah.

And there's some people who are coming at it. So quantum computing is fascinating. And you have all the physicists who are there saying, how can I build these machines? Big physics and huge teams and all this sort of thing. And there's people coming at it from an information theoretic perspective where they're saying, what new things could we do with these new ways of— with these new frameworks, if you like, for thinking about problems? And there's some really interesting potential that— I don't fully understand, I'll admit, I'm not enough of a mathematician, but I'm hearing really positive things and I'm getting very close to a couple of companies there, so stay tuned.

We have a graduate that's actually using quantum computing in health, and I'm not sure if I'm at liberty to go further than that, but they actually are getting productive results, so, and it's using machine learning and predictive algorithms, so, We are seeing process there. All right, I want to move into— speaking of which, I want to move into the next subject, startups. We have another ton of questions there. I'm going to just jump in to the questions from the audience. Therese was asking, you know, when— what are lessons you learned when you first started with your data collection? So for those of you who don't understand the background of that question, I mean, almost any machine learning or AI problem, you need a reference dataset. And so to start working with and, you know, lessons on how to get it, deal with it, et cetera. Who wants to take that first? Verpal, you have, you've been a little silent. You want to go? Any lessons learned that you want to share?

Well, one of our biggest problem is Anyways, in healthcare is we can't just take a dataset and start using them without anonymizing it and all those issues with that. And so the next step is sanitization of the data, what the tagging of the data. And overall, before you even start that one, what problem I've been trying to solve? Because I could start throwing tools at it and still not get anything out of it. If I don't know what problem I'm solving. So the very first thing we always ask anybody who's collecting data, wants us to run some analytics on something, is what are you trying to solve? And then you may not even have the right data for that problem. So do you have the right data? Do you have the right format of the data? And those are all the things, questions come afterwards if you first know what the problem is. At least that's one of my experiences. We try to apply some random stuff onto data and just start putting tools in there without trying to understand what we're trying to solve. So that's one of the biggest problems I've seen when I've come across people trying to do this.

Therese also added, any creative ways you've seen AI companies get the data? Michael, Jeff, any of you want to take either of those questions?

Oh gosh, I think the thing I would say is Data is the biggest challenge if you're doing data-driven supervised learning. And if you're applying it in the commercial— for an enterprise-type startup, it turns out that it's all very well, I can demonstrate that I can get this, but really you're going to put the vast majority of your engineering into wrangling with data, wrangling the data, right? And so that's just— it's just a truism. I'm sure, Geoff, you'll have some interesting examples for how to get the data.

Jeff, you want to go?

Yeah, I don't know if I would call these creative or not, but, you know, there's many people after going through the exercises that Virpal was mentioning, discovering that actually that data that someone who said, oh, we have tons of data, you're going to be happy after going through the exercise of finding out it's completely useless. They've had to just do a complete reboot and actually collect exactly the data they want themselves. And of course, this is, you know, it's a trivial example at this point, but, you know, obviously the car companies, there are public datasets for that application, but all the companies go out themselves and collect many orders of magnitude more than they could get any other way.

And, you know, I've been talking with some different AI experts and there's some interesting stuff going on. With the volume of data needed by creating sort of random generation, et cetera. So am I correct to say, and you guys would know better than I, that the amount of data required to train many of these models is decreasing at a healthy clip?

Yeah, I think it's a big area of research right now. I mean, the, you know, the big sort of revolution in deep learning going back to 2012 was If you have a ton of data, it really does work a lot better. That's been driving everything since. And, but of course now everybody's gotten to where you said, which is what about all those applications that don't have a ton of data? What algorithm is going to learn on this data? And I think there will be, continue to be a lot of attention.

Well, speaking of which, and welcome back, Purple. I saw you jumped off. So we've got a black screen of death. Maybe someone can remove it, but it's nice. It creates a six-pack now. Except one is black. So big companies can get— another question— big companies can get the data, but startups can't. Any recommendations? And I'm presuming what the question is asking, how startups can go get some data if they're struggling.

Find the unfair advantage, surely.

Yeah, right.

And, and I, I was going to answer that we We would love to collaborate. I mean, our disadvantage, even if you're a big company, our disadvantage is we are slow in doing anything in our life. And the startups have advantage that they can go fast. I think collaborations are one of the ways. Of course, Jeff did mention there's publicly available data also, but to a certain extent. But collaboration among the companies itself, startups and with big companies, I think that's That's the key, at least for us. I think that would be a great thing if we can start collaborating. And that's why we are sponsoring these events and things like we want to move ahead and work with other companies who are able to do these things faster for us. We will get you the data. You help us solve the problem faster.

That's— guys, I mean, I hope everyone on this webinar just heard what was said because I think that's the seminal statement right now. You have the big companies that can't move fast enough, but have the data and have more resources to get the data, so that's the competitive advantage that you were, or disadvantage, or whatever that you were referring to, Michael. That's great. The other Michael had a question, and we have lots of questions, guys, just so you know. Another Michael had a question: What are the optimal sales channels for AI-based companies direct, indirect, sell your IP outright. You know, you were just talking about partnerships. I guess this is more of the sales side. Anyone want to take that?

Yeah, the answer is yes, all of the above.

Since when is the answer no? So to any, any way you can sell, sell basically direct, indirect, etc. Sell your IP. What about selling your IP? That seems a little extreme.

It depends on what the company is. It depends on the business model, right? So there's certainly going to be people out there who are selling IP who are saying, hey, we can do this. And that's fine if that's the right thing for that market, for that mix of things, right? Without knowing more about the situation, it's impossible to answer.

Fair enough, fair enough. And actually, what I'm going to ask is if you ask a generic question like that, To be fair, maybe provide a little specificity. I'm a B2B blankety-blank company. What do you think the best channel to market is? Because I kind of agree with you. It's like, what's the best channel in the market? It's like, uh, sell. Uh, Betty's asking, what types of due diligence would large organizations do before bringing a little startup on board? Uh, purple. So now you put it out there, there's my 100 questions coming in.

Yeah, so I think one of the major one we, at least in our company, look for is the clinical evidence is very important. And every company is different, so I think it will depend upon company to company which market you're trying to go into. For us, since we are in healthcare, some kind of clinical— it doesn't mean you have to run a million-dollar clinical trial to get it. It's just the evidence has to be there that which supports that, hey, this therapy or this diagnostics actually does this and there's an evidence behind it. And if you have a hope of getting that evidence, that's also good. So that is one of the things, at least for our company, that's very important.

Are you worried about their financing at all? Like if they run out, you say, you know, if they're underfinanced, does that concern you or you more care about speed?

Well, um, it, uh, I will leave those to my BD department to figure that out. From my point of view, I'm more worried about speed. So if business development people will have to worry about finance and all this stuff, from my— is we are trying to get to solutions, and if somebody has a faster way of getting to a solution, that is the thought we are looking at. And I will leave that answer to the BD guys. To be honest.

Okay, George was asking, we are having a tougher time going direct to enterprise CIOs as they seem to be a little less aware, less interested in these technology advancements. Thoughts as to why? So big companies, CIOs, he's saying— and again, we don't know the industry, George, but let's just take it as there seem not to be as experienced in AI, machine learning, and robotics. Is that what you guys are seeing as as well, or is George a little off?

So, do you want to take it, Chip?

No, I was just going to say, no, he's probably right. Unless you're talking about the CIO of a very specific company who's really focused on AI, they probably have bigger problems than this sort of AI, and you're probably—

I just got hacked.

Yeah, seriously, yeah, I just got hacked. My system just fell down. And probably for those companies, you've got to move in a level lower to someone whose purview is more focused on AI.

Well, we're fine.

You got really big. I think we're trying to correct the screens there. Anyone else want to add to George's question about enterprise knowledge?

I would say that it's a matter of, as with any enterprise-oriented startups, of finding the early adopters, finding those people who get excited about, or those organizations for whom what you're offering is the highest priority thing. Oh, we shuffled. That's fantastic. So, I don't think this is any different to offering any other type of solution. The underlying— offering machine learning shouldn't be the thing. Offering AI shouldn't be the thing. It should be, I'm offering a business solution of some sort.

Yeah, I agree with you. Don't sell the gear, sell the tool, right? The classic Silicon Valley terminology is the hammer and the nail. The hammer is the tool, the nail is the problem. You always sell the problem. Michael was asking, as we near our MVP release in February, which is coming up, What are some of the key metrics we should achieve in our— whoa, in our— sorry, in training our datasets, building our IP? Again, pretty generic without an industry, but I'll turn it over to you guys for that.

Well, let's see. I mean, in terms of accuracy, the numbers are very— obviously they're application dependent. I think You know, in a general sense, you, you can only, uh, you know, ask that you've, you've cross-validated your models in a way that you really trust whatever numbers you have, you actually trust that they will carry over to the, to the problem when it's fielded.

So let me ask you, so back to, to, to get to the industry-specific, is there a good website or, or source that you could say that has these, um, targets per industry or segment that they might want to look at?

I think for that you're gonna have to go to— you won't find a single website because you really have to get very in the very specifics of the problem. For that, you, you've probably actually got to find academic papers or something else that are, that are using a data set exactly like what you're after and try to compare to those, those sorts of Or you need to know that it's really going to make a big difference for the domain you're after.

So you need to understand your target domain well enough to be able to say, hey, this will make a difference. And it becomes very important to make sure you're not playing on vanity metrics at that point. You can get away with vanity metrics when you're talking to investors for a while. Or if you've got a bucket.

Sorry. Keep going, Michael. Sorry.

No, no. I think avoiding vanity metrics is probably the one thing I'd lay out. Out there, which is exactly the same as what Jeff's saying with do cross-validation. Make sure it's real. Is this actually needle-moving for the customers you're going after? And that, to me, is the point where you can say, yes, I've got an MVP. Prior to that, you don't have one.

Fair enough. I see my team— Ignacio has a great question. We're going to go through a few more and then get to the final topic of the call. But what have you seen around talent availability? I'm presuming Ignacio is going to be starting a company of his own. As much as we want to do things with AI, is finding talent a major constraint? How did— how— any tips on doing it effectively? Good question. Thoughts, gentlemen?

I can answer that one. Yeah, finding the right talent is very difficult right now. And I can just say that I had an opening a little while ago about AI person, and everybody who had taken a course in AI started applying. That, that wasn't what I was looking for. So I think that's the challenge is it's very difficult to go through the chatter and find out who has the right experience in the right areas. So just taking a couple of courses is not going to help, at least for me to hire somebody for that. And it's becoming very challenging. And people are being plucked, especially the West Coast is the biggest enemy of ours. Living in the Pittsburgh area, they just get swallowed out. They go to the coast somehow. We have a big fat problem.

You made me cough up my coffee with that. It's like, oh, I just took, you know, AI, I read AI for Dummies. Can I get the— Yeah, Geoff, Michael, you want to add anything?

Sure. I mean, it's a major issue. Strangely enough, I was going to say that we are seeing some companies that are having success in saying, hey, the lifestyle in Pittsburgh is pretty good. Come and escape from the coast and come back to, you know, somewhere where you can afford to live and all those sorts of things. So it cuts both ways. It is part of that key requirement. If you're starting a company that's going to use AI, make sure you've got those people up front. It's also really, really important to be clear about what you're after. Are you trying to do something that's really radically new? Do you need PhDs or do you need practitioners? Do you need data people who are going to wrangle the data? Get very, very specific about those skills that you're after. It's going to be tough regardless, but the clearer you are about what you're after, the better off you'll be.

Well, Jeff, you probably taught the course for the people that went to the pool job.

That's right. Certainly, there's just a supply and demand mismatch on the talent side right now. It's just, it's going to be an issue for a little while before it balances out. I did want to echo the other, the other comment. I've had many conversations with people from Silicon Valley where, where they said this is no place to live. I want to come to Pittsburgh just because the cost of living and it's just gotten too crazy.

I keep your fingers crossed.

I mean, it's seriously crazy cost of living here. Let's not even get down that road. That's another, that's another webinar. Luke had a question. What would be the most important factor for AI adoption by consumers? I mean, I think AI, I mean, direct, indirect AI adoption has already happened. I mean, a lot of the ads you're seeing, as you were saying earlier, a lot of the stuff, but any other thoughts that you guys want to add for important factors of adoption by consumers for Luke?

I don't think the adoption curves are going to be substantially different. AI will enable new products, and when people encounter those new products for the first time, they'll be new and exciting and perhaps worrisome, but they very quickly get over this hurdle. We noticed that with Uber self-driving cars when we gave rides to members of the public. They were very excited initially, and it was shocking What a small number of minutes before they already forgot they were in a self-driving car and just want to take selfies and send it to their friends.

Oh man, I gotta almost go with this. Okay, so like we know that like, what was I? So, you know, there's AI for good and kind of AI for bad. Right? And let's say that like there's some aspects of AI that are clearly for good. There's some aspects of AI that are clearly for bad. And I'll give an example, like, and then there's some that are uncertain, like cars, I think for the most part lean to good, right? And Facebook's addiction algorithm leans to bad, you know, because, you know, getting people addicted to their phone so that they do things that they wouldn't normally do doesn't seem like a great use of AI. And I'm interested in your thoughts because before we go in the final topic of investment, do you know the underlying morality of AI? Do you think we have it right or is there some ways that we need to go there? And I don't know who wants to take that one. I had to ask because I don't know, it just came to me as we were talking about it, self-driving cars in particular.

I guess I'll just comment that for— there are certainly these ethical issues that we have to be concerned with. I don't think AI fundamentally changes those issues that much. It essentially accelerates what was already happening without AI, and in some cases makes it blatant in a way that's a little disturbing.

Right, so Facebook would have tried to make you addicted with or without AI. Exactly. AI helps them do it better.

Right, they just got better at it. Exactly.

It also— the beauty of it is that it makes those underlying conditions stark. It makes things visible that weren't otherwise visible. So there is a positivity there in terms of there's a whole bunch of stuff in our society that we'd prefer wasn't there. We train on those things and suddenly you have an algorithm that is not quite as ethical as you'd like. Well, that says something. And so we're getting this fantastic mirror of ourselves from AI, if you like. So AI itself, it's just another technology. It can be for good or bad and it's unstoppable. These are things that are happening— robotics, AI, automation, advanced automation in general. The upside is that we get to see more of ourselves.

That was beautiful, actually. You know, very interesting that it's sort of shining a light. Purple, you want to add anything?

No, I agree with what's being said right now. And it's a tool. If you use it for wrong purposes, like nuclear technology, you can build a bomb and you can make nuclear medicine. Which can help people. So that, that's two different ways of applying anything. Same with AI, no different.

Well, let's move on to the funding environment and last couple questions here. There's so many. Thank you all for your question. Maybe we have to do another one because we had like great turnout, tons of questions. Barry was asking on the funding environment, in terms of assessing an investment opportunity, how much emphasis is, if any, is put on method and more specifically unique but current non-AI-assisted but thoroughly tested method versus looking for an already built technology. So I guess he's saying I'm gonna do it this way or I've already done it. And if you're, you're going out to look for investment, uh, does it matter if you have the right team and a good method? Versus something already built?

It totally depends.

Yeah.

On the industry, okay, yeah, yeah.

But also, there's really good cases where you can say, I'm applying well-known technologies and I can get an edge. I've got some sort of barrier to entry with using existing technologies. Other times you're saying, I'm doing something net new, and that becomes interesting because it gets an edge from that.

OK, so, but I would imagine therefore what you're saying, just to clarify, Michael, is that, you know, you could envision yourself funding both depending on the, the application or commercial application.

Yes.

You guys want to add anything else?

Yeah, I'll just echo that the team is obviously very important, the method less so. What's more important is have you actually found a product that you've made better? And those will always be the questions before the specifics of them.

Yeah, and I'm sure— Purple?

Is the solution better than the other solution? It's a matter of the solution, which is the right solution, better solution for the customer. And that needs to win. It's not the method how you got there.

Okay, so we're going into the last question here, and trust me, there were a ton more and lots of people still asking, but You know, we got a lot of people on this webinar and who will watch it in the future that are thinking of starting an AI startup, or maybe already have. Is there any piece of advice that you would give them, you know, from your experiences and say, hey, you know, when starting out, think of this or do this or consider this? I'll take it in whatever, whoever wants to go first on this. Anyone?

It's kind of a cliché answer, but I would just say what we were just talking about. Don't get enamored of the method and of AI. Ask yourself the really, really hard question: is there actually a product here that's better than what else is out there?

So I'd add, if you can add to that, I think in general that is the answer. For those people who say, hey, I've got something that's really deep technology, I want to take it out, you really need to work hard to make sure you're not progressing on a multi-year science experiment where you're going to come out of it with nothing. Right now, I'm really excited about companies that are doing something net new, deep technology. But the main part of my practice is to make sure we marry that with a real product that people really care about, where there's a crying need, a screaming human need that's going to be solved by this thing. And ultimately, that's what you need. Ripple?

Yeah, what problem you're solving. Again, I think I said that before. If you're probably— what problem are you solving? If you're probably solving the right problem, I think you'll find people to invest and the customers will come to you. And if you're just doing it to writing a paper for technology purposes, then that's a different solution you're looking at. Then you write a paper for technology and get a grant, whatever you want to do. But if you're trying to solve a problem. Solve a real problem. See what the problem is, understand it, and solve it.

Well guys, this was great. You know, it turned out to have the really perfect mix of entrepreneurship and the educational side, the investment side, and the big company side, but everyone's passionate about the space. So I just want to give you guys a round of applause. Thank you so much. Thanks everyone for tuning in. Again, if you're interested in joining the AI program in Pittsburgh, and hopefully all three of you will be mentors there, it's fi.co/join/AI, uppercase. And really beautiful time today. Thanks everybody. Thanks for tuning in.

Thank you.

Thank you.