The second episode of A4N — the Artificial Neural Network News Network podcast — is out (listen on my website, on Apple Podcasts, Spotify, Google Podcasts, or YouTube). In this episode, our resident virologist Dr. Grant Beyleveld fills us in on how anyone can contribute to the cure for the coronavirus pandemic. We also discuss mind-controlled prosthetic limbs and what it takes to succeed as an AI start-up.
Our special guest for the episode is Ben Taylor. Ben is the Co-Founder and Chief AI Officer of zeff.ai, an AI product company, and former Chief Data Scientist at HireVue. He is a prolific thinker and innovator, and we’re thrilled to have him as a guest on A4N!
All the links we mention during the episode, as well as a full transcript of the show, are available below.
Segment 1 on Tackling Coronaviruses with Machine Learning
Led by Dr. Grant Beyleveld, our resident virologist
1:12 Ben Taylor, Hirevue
4:41 Die Antwoord
5:39 Maryam Khakpour LinkedIn post, Yuval Noah Harari’s book Homo Deus
12:05 CORD-19
16:33 First episode of A4N podcast
17:00 Kaggle Covid-19-related tasks
20:44 Folding@home
Segment 2 on Mind Controlled Prosthetics 36:39
Led by Vince Petaccio II
37:45 Reference blog post from University of Michigan
47:41 Gabe Adams: Twitter account and YouTube video
54:27 Norman Doidge book The Brain That Changes Itself
Segment 3 on AI Startups 57:00
Led by Andrew Vlahutin
57:20 Reference blog post from Andreesen Horowitz
Jon Krohn / A4N YouTube channel
Jon Krohn website for signing up to email newsletter
Transcript
Jon 0:00
Welcome to A4N, the Artificial Neural Network News Network, the show about the latest developments in artificial intelligence, machine learning and data science, where we both introduce technical aspects of these advances, as well as discuss their social implications. In today's episode, we'll be covering how data and machine learning are applied to understand and eradicate viruses, including how you at home without any experience with ML yourself could help out.
We'll also talk about mind controlled prosthetic limbs that are facilitated by machine learning algorithms, and how today's AI startups are so wildly different from the software as a service tech startups that venture capital firms have invested in historically.
My name is Jon Krohn, I have a Canadian accent and always use the word data as a plural term. So that's how you can distinguish my voice. I'm here with my co hosts...
Andrew 1:01
Andrew Vlahutin
Grant 1:02
I'm Grant Beyleveld. I'm the one that sounds funny because I'm from across the country, well, from across the world.
Vince 1:08
And I'm Vince Petaccio, the other one.
Jon 1:12
Haha, good. Thanks, guys. Let's get started. We're very lucky to have a special guest with us on today's show. That's Ben Taylor. So Ben Taylor was the chief data scientist at Hirevue for four years. Ben, do you want to tell us quickly about what Hirevue is and what you did there?
Ben 1:30
Yeah, happy to. So hirevue they, they were a pioneer company where they were the first company to do digital interviewing. So this is interviewing at home, on your, on your iPhone on your laptop. And this sounds very normal today. But when they started 14 or 15 years ago, their customers would ask what what a webcam was, which is laughable today. So they would actually mail a webcam. So if you want to go interview for a mining company and you didn't have a webcam, hirevue would mail you a webcam and you would do this digital interview, so
We did bring an AI components to what they do we it's called hirevue insights, where we use automatic speech recognition, raw audio analysis, and we build these holistic models that predict how well someone did on an interview. So they’re performance models.
Jon 2:23
Nice. That's cool. And that is nicely related, though, coincidentally related to, to how untapt (so the rest of us on the show, all these co-hosts, we work) also in human resources, building models that automate recruiting. So interesting overlap there. And so more recently, you were also the co founder of zeff.ai or zeff ai, and AI company. First of all, how do you pronounce it properly? And what did you guys specialize in at zeff? And I also hear you might have some exciting news for us.
Ben 2:57
Yeah, so the it's Zeff. And there's, there's a whole backstory to that it used, it used to be called Ziff. And we had some legal pressure to change it. And so I actually hate Zeff.
We're called Zeff because that was an easy change. We focused on on auto ml, but around deep learning so can we enable an engineer to build deep learning models with image, audio, video and text? And the place we got to work that was more exciting, it was this concept of fusion models. Can I have a single model that consumes video and audio or multiple different types of images and text and structured? And so that's where we ended up building models like that for insurance, proctoring, image classification, stuff like that.
Jon 3:50
Nice. That sounds fascinating, and I have no doubt that you will have a lot to contribute to all of today's topics given your background.
Grant 3:58
Could I jump in for just a second? Yeah, considering that my voice sounds funny because I'm from across the world in South Africa, zeff is actually a sort of slang term, which kind of means, I guess it could be referred to as like kitsch. It's like rooted in like 1980s style. So you would describe someone's like clothing as really exact if it was kind of a little bit otherwise and interesting and different.
Ben 4:27
Yeah, we were we were actually laughing about the South Africans zeff definition. Because it's essentially white trash. bling. Right? That's, that's like we're poor, but rich. It's great. I think it fits us well.
Grant 4:41
Exactly. If anyone's familiar with Die Antwoord, that sort of strange, South African band that's they sort of capitalized on like the Zeff movement.
Jon 4:51
Got it. Thank you very much Grant. That was super helpful. And then my last question for you before we start digging into our content for today's podcast. So Ben, you were actually supposed to be in person with us last week, visiting from Utah, recording with us in New York. It was supposed to be all five of us in studio on camera so that we could, you know, stream to YouTube and people could check out the live raw feed. What happened? Is there's something going on in the world?
Ben 5:21
Yeah, my wife forced me to fly home from Boston. She was pissed. And she thought, like New York was showing up in the news. So with Corona Yeah, my wife wasn't gonna let me fly to New York last week.
Jon 5:37
Yeah, and it ended up being he was he was that exact time, that was the day. It was it was that Friday that you were supposed to fly into New York and film with us. It was the first day that people started really not showing up for work. You know, we work in a wework and it was really quiet that day. Things were definitely different and boy has it been different since.
So we are actually a new show and we are going to talk about the coronavirus but we are going to take a different take on it. So a lot of what you see out there even a lot of data that you see out there are related to the epidemiology of Corona and we are going to sidestep that we are going to talk about how data and machine learning are applied to study viruses, including the coronavirus, in order to help us build defenses against viruses. And as a lead into this topic, I have a really somber LinkedIn post from Maryam Khakpour. So she is a PhD student in Iran, who has been following my writing online for many years, we've corresponded probably a dozen times over the years. So she wrote on LinkedIn yesterday:
Dear Dr. Krohn, I used to read your article and she's she's referring specifically to an article that I wrote, called, we live in the most peaceful times ever, which used data to show that conflict both on a large scale and on a small scale are down across the board across the world. And so she said, I used to read your article and seek consolation to help me condone the ugliness I see in this world. A few days ago, I saw my dad choking to death right in front of my eyes and died of Covid-19. Now that I see lots of people struggling with this unknown manmade virus, and no scientists and no technology can help, how can I still be optimistic that we are living in the most peaceful times ever?
So first of all, this is it is absolutely tragic. And my condolences for your loss, Maryam. That is awful. I can't even begin to imagine what that experience is like, and no question this must be an incredibly sensitive time for you. Um, I doubt there's anything I can I can really say to help right now. But yeah, and especially these kinds of cold hard figures, but I do actually overall through the things that I've seen so far in the early stages of this pandemic, I actually still do believe that we live in the most peaceful times in human history. In fact, I think some a lot of the ways that we've responded to the current pandemic only bolsters that case. So, a great book for anyone to read about this, including you, Maryam is Yuval Noah Harari’s book Homo Deus. So it's one of my favorite books of all time. And it goes into a lot of detail into how humans until the past century, had to accept war, famine and disease is a natural part of life. And so while today, war, famine and disease still do impact us, the rates with which any of those three occur are down precipitously from centuries past. And for the first time in human history, it's possible to imagine at least a world where all three of these perennial issues are wiped out completely.
So infectious diseases used to wipe out entire civilizations. When smallpox was brought from Europe to North America in the 17th century, it killed 80 to 90% of the native North American populations. Today, alongside many other ones commonplace diseases, smallpox has been completely eradicated thanks to scientific advances like inoculations and public policy advances like quarantine.
Another example is the plague pandemic, which was known as the Black Death, which killed 30 to 60% of Europe's population in the 14th century. Today, the plague infects about 600 people a year only, and it's treatable with antibiotics.
So infectious diseases, armed conflict and malnutrition today, they account for only a small fraction of deaths relative to the most common causes including cardiovascular disease and diabetes. So this means that in the past few decades, it has become more likely to die from overeating, from a condition like cardiovascular disease or diabetes that is caused by overeating, than from war, famine and infectious diseases combined.
And, and, and it's getting better all the time on actually all of those fronts. So you know, we're beginning to tackle these obesity problems too and war, famine, infectious diseases, these these continue to become smaller problems. With the novel Corona outbreak today, there have been dramatic worldwide behavioral changes influenced by data driven public policy, the rich scientific field of epidemiology, and the spread of at least some reliable information over the internet. That means that countless of millions, countless millions of lives will no doubt be saved because of the way that we have adapted, adapted in just a matter of months as an entire civilization. And thanks to affordable mass produced medicines, sanitary products and medical technology like respirators, only a couple percent of those affected are likely to die, which is nevertheless horrific, especially when it's somebody close to you, you know, like a parent.
And yeah, so I, I can't imagine how you must feel right now it must be absolutely awful. And that's normal for you to feel that way. And I'm deeply sorry that you had to go through this experience. But I do hope that in time, you'll once again be optimistic, like you were until recently that scientists and technology can and do help, including during the present outbreak, and that quantitatively, we not only live in the most peaceful times in human history, but that the outlook is even better for the generations that are to come.
On the note of scientists and technology helping us understand and tackle viruses, we're lucky to have Dr. Grant Beyleveld as one of our co hosts because he happens to have a PhD in virology, specifically in the application of large datasets to the study of viruses. So, Grant, I've got some questions for you. You have been talking a lot this week about something called CORD-19. What is that? What's CORD 19?
Grant 12:10
Yeah, Jon, so CORD 19 is a data set that has been made available to the scientific world at large, specifically related to the SARS-CoV-2 virus, which is, in case people aren't familiar, the actual correct name of the virus that's causing this pandemic that we're all experiencing right now. And of course, Covid-19, which is the name of the disease that the virus causes. So essentially, what the government did was set up a central repository where all of the available literature surrounding these, well this disease and this virus were available. That database contains around 30,000 scientific articles. About half of them are full text, unfortunately, the other half are largely behind paywalls. So they're not fully available, but they are available in terms of their, you know, metadata, so title, authors, and their abstract.
And all of this data is collected in a machine readable format. So it's easy to apply machine learning models to it and so forth. And everyone can just kind of go and download this datasets. It's around two gigs in size, I believe, and start playing around with it, experimenting and seeing what can be done with, I guess, a view that, you know, if we can look at all of the available research that's ever been gathered around this particular disease, this particular strain of coronavirus, as well as the research that's been done on this family of viruses in the past. Perhaps there are sort of things in there that that are uncomfortable that the researchers that are hard working on this just wouldn't possibly have the time to uncover because no one can read everything and no one can read everything with the amount of depth and then create the kinds of links that are necessary in order to, you know, potentially find unusual connections within this data.
So it was all put together with with this plan to, to give AI researchers a starting point. And hopefully they can use this data to try to, you know, solve new problems and get to some of the the facts behind this this pandemic.
Jon 14:25
So why can't so why is it that that AI is needed to do this? What is it about AI that makes that, like, how can AI be applied to this CORD 19 database in order to generate insights?
Grant 14:39
I mean, I guess it just comes down to the fact that there's just straight up too much literature out there. The the rate of scientific discovery is pretty fast. But yet it might feel like our progress doesn't doesn't happen the way you want and from like a layman's perspective, you know, when we talk about scientific progress, we're talking about cures and drugs and vaccines. From a scientific perspective progress is, you know, uncovering the minutiae of how one viral receptor interacts with one particular cell surface protein. And so, you know, those tiny little bits of scientific progress are happening all the time, and at a quite rapid pace, and indeed, in this particular space around coronavirus and the coronavirus family, this has been going on for years. But in order to, to kind of come up with some actionable insights out of all of that disparate data, it's necessary to, to look at the whole scope of literature that's been published. And it as I said, that's just too difficult for any one person to do, surely from a time in the day perspective, but then also from a mental capacity perspective, it's very difficult to ingest that amount of complex data and be able to create those links over such a vast space. So if we can use ai ai is particularly adept at uncovering patterns and looking for for links, and so on. So this could be a really great way of of creating these links between all these disparate studies and potentially having a profound implication for the development of novel antivirals or the understanding of the viruses origin or the understanding of how this virus is spreading out in the real world.
Jon 16:39
Got it. So. So in addition to our coordinated global response, and, you know, people doing social distancing, some governments quarantining folks and you know, all these strategies, these social strategies, in addition to those being facilitated in a way by data, and by communication over the Internet, and and the internet also facilitating people to be to be able to come with strategies for avoiding getting it themselves, you know, watching videos on YouTube about how to wash their hands properly, and that kind of thing. In addition, the scientific community is able to leverage our interconnectedness facilitated by technology. And so we're able to pool together all these common resources. And then we can use machine learning algorithms with which a lot, which, which lots of people learn about, through free resources that are available online, through GitHub repositories and these kinds of things. And so there's a ton of open source and, and community and collaboration that we're able to do as as scientists and maybe even as non scientists over the internet.
And that actually brings me to one specific thing related to Covid-19 and this CORD 19 dataset. From my understanding, and I think you can you can fill in more information on this, the kaggle platform itself -- so we talked about kaggle a lot in our first episode of this podcast. So kaggle is a platform that allows people to compete against each other and so it seems like that particular platform on the internet is also getting involved with tackling Covid-19.
Grant 18:22
Right. Yeah, indeed. So the the dataset is actually available to download from kaggle. That's not necessarily where it's it's permanently stored. But they've got links here, they've got links to it there. And Kaggle also have a list of tasks that have been put together by the National Academies and the WHO, and these are a group of tasks that are considered high priority tasks. And the idea is that if we could try to you know, get into the nitty gritty of these particular problems, that would, you know, help the sort of general drive to combat and hopefully, defeat this pandemic. And those tasks include things like you know, what's known about transmission, incubation and environmental stability of the virus itself, or what are particular risk factors for particular people in terms of the severity of the covid 19 disease? What do we know about the genetics and the origin and the evolution of this virus? How did it come to be so, so dangerous and so easily transmissible when, you know, coronaviruses in general are sort of everywhere, and people are getting them all the time, and they just cause mild respiratory infections, and it's kind of just considered a cough or a cold. Why is this one such a big deal? You know, what do we know about vaccines and therapeutics, so there are a whole you know, sphere or a number of spheres of questions where we can dig in and hopefully try to find answers. And those answers as I say might exist in that data set it we just need someone to make all the links and and hopefully pull out the relevant information.
Jon 19:59
Got it. So that's great. So now so any listeners with experience in machine learning can use this Kaggle competition and we'll provide links. You know, if you're if you if you find the blog post about this episode, or the YouTube page about this episode, you know, we'll include links for anything we mentioned in the episode, including links to this kaggle competition. And so people with machine learning experience can then use these data, use the kaggle platform to compete against each other and come up with solutions to some of the various aspects related to the novel coronavirus that you just mentioned. But what if you're listening at home and you don't have any machine learning experience? Maybe you don't write code at all. Is there anything that that listener could do?
Grant 20:44
Yeah, actually, there is. For this one, you actually just need a computer. And you could be solving a range of other things, not just problems relating to coronavirus, but also cancer and Alzheimer's. So there's a project lap called Folding@home. It's actually a really, really cool project and essentially it aims to look at and model how proteins fold and how they move and interact with each other. So just to like take a step back, I think we kind of did a bit of this in last week's episode or last time’s episode, but just to kind of go through that, very briefly, again, proteins are linear chains of chemicals called amino acids. And they typically spontaneously fold into much more compact and functional structures. And these are usually arranged in particular ways and they move and interact with each other in particular ways that are important to their function. So these proteins have are the basic molecular machines of life. They are enzymes, they are structural proteins that build cells. They can be all sorts of different Well, they can have all sorts of different functions in in every sort of sphere of life. So understanding how they arrange and how they structure and how they move is quite important to understanding their function. So if we know how the proteins work, and we know what they're doing, we can then design drugs which potentially interfere with those functions. And in the case of coronavirus, we could design an antiviral that would stop the virus.
So this whole process is what's known as rational drug design. And the the, the approach here was, is to use computational computational biology approach to try to model the structure of a protein and how it moves and get more understanding. But that's a really complex computational problem, essentially, the folding of the protein. So yeah, and so this this study is way more interested in the dynamics of the protein once it's already folded. And so how does the protein move? How does it change shape? How does it interact with the various things that it's, you know, evolved to interact with when it binds to another molecule or another chemical or another protein, how does it change? How does it move? And so those those things are really important.
And typically, the way that proteins have been studied in the past is they haven't really been studied in this dynamic fashion, we've essentially sort of been taking snapshots of the proteins kind of like photographs frozen in time. And while it's a great way of seeing the general structure, it really doesn't give you a sense of of how the whole protein is working and how it does its jobs. So what they did what this what this Folding@home project aims to do, is take those existing snapshots and then looking at the full structure of the protein, try to model how every single little atom might move and how that might work. And as you can likely imagine, modeling the movement of every single atom inside of a protein is a really, really tricky task and it it rapidly explodes into a seemingly infinite space of possibilities.
So using sort of a lot of computational power is really the only way to crack this for now. So because computational power is, is scarce in the academic space and expensive, the Folding@home project is essentially an app that anyone can download and run on their computer. And the second your computer sits idle and you're not busy with it, so say you've, you know, gone off to lunch, or you're asleep, or whatever it may be, the Folding@home app will spring to life. And it'll start to use your computer's CPU cycles to model these, these interactions, these these protein movements. So it'll download a bit of data from the Folding@home servers, it'll run the computations on your computer, and then it'll send the results back. And we can do this at on a large enough scale. And we essentially, you know, across the internet, or the internet interconnected computers, that's the largest supercomputer there is if you can, if you can kind of bring together all of those resources. That's what they've been doing. It's, it's really cool.
Jon 25:13
That's very cool. So we are setting up all of our deep learning servers at untapt to be able to do that to do exactly that. So we've downloaded are setting up the Folding@home software on all of our deep learning servers. So that when we are not training models, you know, we're taking advantage of, we're joining that giant interconnected supercomputer all right now.
Grant 25:36
Yeah I’m running on a GTX1080ti at home as we speak.
Jon 25:41
Nice, there you go. So while you're while you're Yeah, right, while you're on this podcast. That's amazing. So are there other questions from from other co hosts or Ben, now you've got you've got a virology expert right here? And anything that you'd like to ask him or is there anything else that I missed that you'd like to say, Grant?
Grant 26:05
No, I've got it all covered.
Ben 26:09
This is Ben speaking. So I, I've been fascinated about this topic because I've wondered, is there a way for us to take like an Elon Musk outsider approach and just come in here with a huge cluster, big compute and just simulate this and optimize it at a whole other scale? Like, what is the self driving discover antibodies for viruses scenario, something that can happen this year? Are we still five or 10 years out from the 24 hour simulation to here's your new antibody? Does that make sense?
Grant 26:37
Yeah, it does. I think that if they were a large enough, a large enough network, and we could bring enough computers together, I think that that is possible. I just think it's very, it's difficult to I guess it's like a really big problem and it's difficult to wrap our heads around it. It's kind of like visualizing four or five dimensional space. It's hard to kind of fathom that that could be possible, but I think that it is possible. I think that we're pretty close to getting to situations like that. One of the key issues right is you can come up with a simulation to model any kind of system you like, really, as so long as you aren't constrained by compute power. So let's assume that we aren't constrained by compute power. You still need to rationally design a good simulation model, you still need to incorporate all of the various variables, you still need to take account for as many of those possibilities as possible. And I think that that's kind of a difficult thing to do still. So yeah, I don't know. It's a tough question to answer, but it's a great question. I think we are definitely getting closer, but we're still a little ways out.
I was reading a little earlier on the topic of prediction about the idea of how to predict how to predict the flu season. As an example, since we're dealing with respiratory viruses, each year, the flu seasons a little different. And researchers try to sort of game the system and get ahead of that curve and try to predict what kind of flu viruses are going to come out each year. And one of the handy things that flu virus prediction has is many, many years of flu virus data. We've got experience on what each of those seasons looks like. With coronavirus, for example, we have nothing. They essentially have to start from from scratch and just work with no information. And so I think that in a lot of these simulation cases as well, if something is sufficiently new, and we don't have any idea of the parameters, the boundaries of that simulation, it becomes increasingly more difficult to, to do that.
Jon 29:00
So that's one thing. So predicting epidemiology like that is interesting, but isn't something something that I've been thinking about with with relation to these viruses, and correct me if I'm wrong here, but so viruses tend to have a pretty small amount of genetic information and, and just and I think like the coronavirus it has, like, in, in one of its forms, it really only has like three functional proteins. And so it seems like have all of the kinds of, you know, if we were talking about bacteria, there's they are vastly more complex than a virus or, you know, you know, other kinds of issues that we might face. Shouldn't it be possible to kind of relatively quickly characterize the function of the relatively small number of proteins, a relatively small amount of genetic material in a virus. I guess it's still just such a challenging problem that you can't, overnight come up with, with a way of attacking one of those.
Grant 29:58
Well, it goes back to to the original problem, which is to say, yeah, viruses do have a very small genome, small number of genes in fact, usually, from a biology perspective, only a couple of those genes are really important to us, because some of them are our structural genes. And, you know, they just create the little like, body that the virus lives in minutes outside of the cell. The ones that we're really interested in are the genes, the proteins that allow a virus to stick onto a human cell and then get inside or the right and maybe, maybe, exactly, and so we can definitely narrow the field down and pick out just a couple that are interesting to us and that we need. But then further to that, in order to really work on those, we need very, very high quality models and how those look. And so now we're back to Folding@home. And we're trying to model these, these proteins and know exactly how they function. Because another thing about viruses is yes, they have a very small genome, but they they evolve really quickly. And so this novel virus is going to have different proteins to the viruses that we've seen before, albeit very similar and they have the same function in generalized structure. They'll be just different enough that you know, a drug that will work on an old coronavirus won't work on this new one or something like that. So, um, it gets really tricky and that's when, you know, these kind of moments happen. So if you go to the Folding@home website, they've got a whole page now devoted to to Covid-19. And they're, they're really sort of moving or I guess giving those coronavirus jobs really high priority on the system. So more of people's computers out in the world are crunching coronavirus related, folding simulations than you know anything else. So, yeah, we're getting there, but it's not perfect.
Jon 31:51
I don't mean to be hogging the questions. I have another question if nobody else has one. If no one else wants to jump in.
Ben
I've got a quick one I'll sneak in. So the other the other question I had was, I know that the moonshot is maybe not practical because the moonshot is here's your vaccine 24 hours later computer told us it was. So that's the moonshot. But is there a lower, is there some lower hanging fruit where I've seen press coming out that the malaria some one of the drugs for malaria looks like it's making progress with the coronavirus, but it seems like that should actually just come out of a computer where essentially tests the virus against all known whether you're in the US FDA approved or any known drug or treatment that's been allowed to be injected into humans, it would just immediately do simulated tests against whether it's thousands or hundreds of thousands in the computer should have said you should try the malaria drug. So I'd be really curious kind of what that scenario looks like where it's not the moonshot, it's just something maybe a little bit more practical.
Grant
They definitely are doing that. We did that as part of my, my PhD project, in fact, so it's sort of it's called like “in silico docking” I think was the term if I'm remembering it correctly but essentially what you can do is, is take the molecular structure of a library of known compounds. So typically what people use then is a library of all the FDA approved drugs. So these are all the drugs that the FDA has said, you know, these are safe to give to humans. And also they have this other function that is, you know, that is used medically. There's quite a lot of drugs, I couldn't tell you the number of top my head, but um, they'll they'll run all of those against models of the proteins. And again, you can fish out things that are likely to bind in various ways to all of those proteins and then you can kind of go through those and as you say, you can pick up you say, hey, this, this drug over here that's used to treat this like rare cancer is actually really great against covid 19. That's actually how a lot of the original HIV drugs came out. So AZT was the first one and that was actually a cancer drug prior to it being used for HIV, and it wasn't a very good cancer drug, it just didn't ever really go to market, I don't think (stand to be corrected on that). So it can definitely be done on that respect. Again, back to Folding@home, the better quality you have, in terms of models of the proteins that you're trying to target, the better the system is going to work, the more accurate your predictions are going to be. And unfortunately, that system is a good approach. It's a good kind of first pass because it's easy. But uh, you know, it doesn't actually it's not a very foolproof approach. So what we did sort of subsequently to that was we had this now list I was working myself on on influenza virus drugs, and we then have this list of drugs that are, you know, a smaller list and they might work and subsequently then we'd have to have to actually go put those drugs into cell culture with human cells and viruses on them and see if it affected the way the viruses grow. And so it all ends up coming back down to that with biology. Unfortunately, that's the only really sort of almost foolproof way of doing it. It's, you know, that's a whole other bag of chips that we can get into on another podcast entirely. But, uh, but yeah, it's getting there. And it's a great idea, but it's not perfect. And it's at the moment, it's just used as a first pass to identify potentially interesting compounds.
Jon 35:40
Nice. All right, well, that was a ton about the coronavirus, viruses in general, and how we can use data and machine learning and how you at home, if you have machine learning experience using kaggle, the covid 19 related competition, or if you don't have machine learning experience, at least downloading Folding@home and taking a crack at letting your computer -- it could be your computer -- that you wouldn't ever know it probably, but your computer could be the one that makes a huge amount of headway in resolving the current covid 19 pandemic.
So that was super cool. Thank you so much, Grant. We're so lucky to have you as our resident biologist on A4N, thank you so much.
All right, so that's the end of our first segment. Up next we have another health topic which is completely different and also completely mind blowing. It is prosthetics that you control with your brain. And those are driven by machine learning as well. So that's coming up next.
Segment 2 on Mind Controlled Prosthetics 36:46
Jon
All right, welcome back to A4N! I hope you enjoyed that detailed segments on the coronavirus. Up next we're talking about more health news here. This is Mind Control prosthetics. And so this is interesting because in the US alone, there's nearly 2 million people that have lost a limb and millions more worldwide. Prosthetic limbs that behave in a subtle lifelike way, using mind control would make getting around and going about daily activities much easier for any of those millions of people.
So Cyborg mind controlled limbs. That sounds a lot like science fiction. But Vince Petaccio, our co host of A4N, with both a master's in biomedical engineering, in which he happened to specialize specifically in brain computer interfaces like these, and also his master's in computer science with an AI specialization that provides him with insight into how machine learning algorithms could improve these mind control limbs. We couldn't have a better person on the show to talk about it. So Vince, tell us about this topic.
Vince 38:45
Yeah, so this is actually a really interesting project that comes out of the University of Michigan. Basically, the context for this is, as you said, there are a whole lot of people out there who have limbs that need to be amputated for a wide variety of reasons. And what usually happens when you amputate a limb is you end up with is a piece of the nerve that went all the way from the spinal cord out to the end of the limb, say the fingers and the hand, that just has to be cut at some arbitrary point that the surgeon kind of chooses to do that cut. And what can happen is that over time, that end of the nerve, it's not receiving any inputs and the motor nerves don't have anywhere to go. And so the brain is still has, it still has these representations of a limb that no longer exists. And so you can end up with things called phantom limb syndrome, where you experience pain and sensation in a limb that doesn't actually exist any longer. And over time, the nerves can kind of degrade and just result in kind of generalized pain. And this kind of opens up a lot of opportunity for kind of taking advantage of those existing nerve fibers and using them to control some sort of prosthetic device.
Now, this particular study, there are kind of two interview innovations, innervations, a lot of innervations, two main innovations in this case, the real big one here is actually more on the, the interface that they made on a hardware level that connects the the nerve to the technology. So just for a little bit of background, some of the existing systems, what they do is they might take the nerve that gets cut and kind of reroute it to a muscle somewhere else in the body and use the nerve to kind of make some muscle say in the chest twitch, and then record that muscle twitching and use it to control some prosthetic. Another solution is to stick like a cuff electrode around that nerve, or to put a needle inside the nerve and and all of those techniques they work at first, but over time, they tend to degrade. You have a really high and low signal to noise ratio.
Jon
Yeah, so can we get better somehow? I don't know. Like just that. I don't know by constantly using it. It would cause you know neuroplasticity, somehow to engage more, but I guess because it's outside of, you know, it's kind of in the peripheral nervous system I guess there's not really neuroplasticity like there is
Vince
Yeah, exactly. And that's the major challenges. It's not an issue of the central nervous system. It's the peripheral nerve that you're that has just been severed, and it doesn't have its normal terminal terminal points. And you just are trying to interface with it in some way that's not really consistent with how it grew. And you know, just how it evolved in general. So this what they use in this study is something called the RPNI, the regenerative peripheral nerve interface. And the way this one works is they basically take the nerve that was cut, and they can do this years later after there's already been some healing involved, which is pretty interesting to me. And they'll take a piece of muscle graft. In this study, they use the vastus lateralis, which is a big long muscle in the thigh. And they basically just wrap that muscle graft like a lollipop or a Q tip around the end of that nerve and that allows the nerve to eventually innervate that little ball of muscle. And that doesn't really do much on its own except occasionally twitch this little ball of muscle, but they've implanted electrodes in the muscle. And so what happens when you when you move your muscle is there's a nerve impulse that goes through the nerve and transfers to the muscle and then the muscle fibers activate. And so they're basically using these small balls of muscle as amplifiers to amplify these nerve impulses to the point where they can reliably and consistently record those responses with a computer.
So what they can actually do as well as they they do, they cut the nerve up into individual bundles, and kind of splay them out. And then we'll send each one of those bundles, it's called a fascicle to a separate little muscle graph. So now they have a kind of multi channel a set of these balls of muscle graphs that they can use to control a multi channel system. And so and so this gives them a lot of opportunity to, to record raw signals in ways that can be used to control some prosthetic device.
And so this is where the kind of second innovation ties in here. And this innovation is the use of some machine learning algorithms that allow them to, to really get some nuanced control signals out of these, these muscle grafts. So they have two approaches that they mentioned in this study. The first is a naive Bayes classifier. And so in this situation, they're basically looking for, from what I can tell from the from the article about this it's basically a binary classifier. And what they do is they record 50 millisecond windows of activity from those little muscle graphs. And they take the mean absolute value of the EMG or muscle, electrical muscle activity over that 50 millisecond time window. And that that actual, that value, which is just a float value from that 50 millisecond window becomes a feature to a naive Bayes classifier.
And so what they do is at training time they have the subject or the patient, just imagine making a fist or try to make a fist. And so you end up with these nerve impulses coming down the nerve into those little muscle grafts. And they record those 50 millisecond time windows over and over and over in that time that the the person is imagining making a fist. And so you can do this for a few minutes. And you'll have many, many, many thousands of training examples. And then you can just in a couple clicks in the same day in the same session, use that to quickly train a naive Bayes classifier which are generally extremely quick to train and and then immediately start using the prosthetic right there and then it's pretty cool.
Jon 44:02
And so you would train like one action at a time like you would train them? Like, hey, you know, for the next five minutes, we're just going to do making a fist and then 4 to 5 minutes after that we'll do like pointing your index finger or something like that.
Vince
Yeah, exactly. There's a whole range of activities that they gave them, I use the example of making a fist, and you give another one of pointing, but they also do things like, like touching your forefinger to your thumb and your middle finger to your thumb, giving a thumbs up, all kinds of things like that. So they can get different different features or different training labels, basically, and then train that classifier to identify what pose the the user is trying to make.
Another algorithm they've tried and they describe in this paper is a Kalman filter. A Kalman filter is an optimal estimation algorithm. And basically, it's a way to in a control system to try to predict some parameter in a noisy environment using indirect measurements. And kind of one of the canonical examples is, you know, you're trying to figure out when to turn a hydrogen powered rocket engine on and off, because if you leave it on too long, you can melt the engine components. But if you don't keep it on, the rocket can stop moving or fall out of the sky. And you can't measure the temperature from inside the rocket engine, you can only measure it from outside of the engine. So this is another example here where it's a very noisy environment. And you're using very indirect measurements. Here they're measuring the muscle activity from these different balls of muscle, and they're trying to actually predict the correct pose of a prosthetic limb. And so using this, this Kalman filter, they're actually able to produce a much wider variety of more nuanced and complex movements than the naive Bayes classifier.
Jon 45:53
Yeah, I understand. When I used to trade at a hedge fund, we used to use Kalman filters to try to predict like, whether we were in like a bear market or a bull market with like a particular asset or you know, something like that, you know, just kind of regime changes. And so I can see how that would be like You're predicting when you flipped from wanting to have a particular hand gesture to another one. That's cool. Sorry, I spoke over you.
Vince 46:15
Yeah. Oh, no, not at all. I was just going to say that what I think would be really interesting to explore in these types of problems is, you know, my background in brain computer interfaces was not so much controlling prosthetic limbs, but allowing users who were locked in, we worked with ALS patients, primarily to control a computer in a way that allowed them to communicate. So we would just be recording neural signals while they looked at a computer monitor. And our goal was to try to figure out, you know, what letter are they trying to type out on a computer. And what we found is that over time, you have this kind of multi agent game model that happens where the the system our brain computer interface system is trying to learn how to interpret the user's neural signals. But at the same time, the user's brain is incorporating this new device as though it's a new limb. And so you have two systems simultaneously learning in a cooperative way at the same time, but for each one of them, it doesn't necessarily feel cooperative because the other system is changing while it's trying to learn it. So I would be really interested in the future to see some reinforcement learning algorithms that that kind of are designed for this sort of multi agent game type of model applied to these sorts of problems. That's super cool. And so Ben, we were talking earlier today, you mentioned that you recently met someone who was using devices like these. Do you want to tell us about him?
Ben 47:40
Yeah, so this guy, his name is Gabe Adams. He's an inspirational speaker. He actually has no limbs, so he's not currently using these devices. You can look out to if listeners are interested, you can find him on Twitter (@no_limbs)
Jon
...and we’ll provide the YouTube video that you mentioned as well in the show notes.
Ben
Yeah. So he potentially someone like that in the future could have access to something like this where he has no limbs and he could, you know, get mobility. He's, he's actually an inspirational speaker because he's surprisingly mobile today. So he's able to do a lot of things that should be impossible. He can dress himself, feed himself, do all these things, the thing that we were talking about was, how can AI, if it's kind of, you know, what do you want? How could he make his life better? It involves a lot more beyond just the limb replacement to a lot of its voice activation. You know, can AI assist with opening doors, can they can he program with his voice, stuff like that? And so that it was more blue sky conversations, but I'm, I'm hoping in the next year or two, we find an opportunity to work together and do something that's exciting and inspiring. But yeah, very interested in the stuff that Vince is talking about.
Jon 48:59
Wow. Yeah, that's cool. So and if there aren't any other questions I have, I have just one more. So there's, I've read a lot, Vince in the last couple years about Elon Musk (he's come up a couple times in this podcast) -- about Elon Musk's Neuralink project, which is a project designed to, to... it’s brain computer interface project where, you know, he thinks and I think it's probably an inevitability some way or another where we're connecting directly between our brains and the Internet, and we're kind of connected in that way. And so how does this kind of stuff this mind control prosthesis relate to that work?
Vince 49:43
Yeah, so whenever the subject of brain computer interfaces come up, it's, it's inevitable that someone will will kind of reference something like the matrix where, you know, an operator can pop a floppy disk. For some reason they were still using floppy disks into a drive and upload instructions on how to fly a helicopter into somebody's brain or teach them kung fu by just uploading data directly into their brain, and I think that's that kind of level of depth of interconnection between computing systems and biological brains is kind of the inspiration inspiration behind the Neuralink project. And I, I'm skeptical. I will never say that anything is impossible because I don't know what the future is. But it seems unlikely to me. And here's why: I think that we have over the last few decades actually pretty well optimized our interfaces with computers. We evolved in a in a regime where we had to learn to communicate as efficiently as possible with other humans, because that benefited our evolutionary ancestors. We also evolved the nervous systems that are really good at telling were telling where our limbs are in space and being able to identify what things feel like and interact with the physical world.
So things like mice for computers and keyboards, and, to an extent voice control interfaces, these I think, are the optimal channels for getting information into and out of the brain. To suggest that a brain computer interface could be, you know, an electrode that's like a grain of rice that you kind of like stick under your skin behind your ear, that could somehow have a more efficient means of access to information in the brain suggests that there's a lot of low hanging fruit in terms of how information gets in and out of the brain. And I don't necessarily know that's the case. And just on a technical level, something like a brain computer interface, I think, viewing that as something that could be small and just implanted in one spot, I thinks kind of misunderstands the way that the brain works. We're still have a lot to learn about the brain, but so much of what the brain does is distributed across the entire organ. You know, just having one memory can involve everything from motor systems, to sensory systems to vision to all the different parts of your brain. And if you really want to have that sort of deep, interactive connection with neural systems, then you would have to be directly tapping into neural signals throughout the entire brain. And I used to actually work I worked about seven years in brain and spine surgery, and I have never met a spine surgeon or brain surgeon who would ever crack your skull open to implant a big electrode net over your over your entire brain surface for this type of procedure, nor have I ever met a patient who would want it, so I just don't know that it's likely in the foreseeable future.
Jon
One thing that I find really interesting about this idea is that let's say that in 20 years, this Neuralink project is super successful. They come up with a way of somehow non invasively having a mesh go like through your like around a hole around your eye or something. And it like gets implanted in your brain and it covers your whole brain and it's really effective at accessing the internet, like somehow we managed to get this to work. And then two years later, they're like, oh, man, we just came up with like a way better way of doing it. And we're gonna need to like, but you can't get like the old way out because it's like this like mesh on your brain. So they had been like, he certainly tried to, like, insert another one. And this is like, every couple of years. It's like the like, Apple iPhone program, like I'm on the program where I get a new iPhone every year. So I'm gonna have to, like every year get a new like mesh implanted on my eye or in my nose. And it kind of like there's just all these meshes getting stuck in there.
Vince 53:41
Yeah, I mean, I definitely have enough cobwebs in my brain as it is. But yeah, I can definitely see that being an issue. You know, I think a big part of this as well, like, as you said, I can imagine, let's say we get around the technical challenge of getting something into the skull. Well, that's still only going to get us that outer surface of the brain and that part, which we kind of called the neocortex because of its evolutionary and the newest part of my brain is really directly tied to our consciousness in our cognition, but it doesn't tie into things like emotion, which is a big role in how our brain works. And one final word on this that makes me even more skeptical is that the brain works by changing itself. It has to change itself. There's a fantastic book on this by Dr. Norman Doidge, called The Brain That Changes Itself. And it totally changed the way I thought about brains and they're, they're plastic, they change over time, you cannot learn without your brain physically changing. And in order to learn, you have to forget some things and you have to restructure your brain. So any system that relies on physical interactivity with direct neurons will immediately become deprecated. The moment you learn something or forget something, which happens many times every single day. So I think there there are major challenges to this maybe insurmountable fundamental challenges.
Jon
Vince, I stopped learning years ago, but whatever. All right, so we've gone over on this segment. I loved I love talking about this. So it went on longer than than I intended, but I just couldn't resist asking more questions. This is such a fascinating topic. All right. So we've got, we went on. We talked about the coronavirus for so long. We only have one more topic left to cover. And it's really exciting. It is not in healthcare, but it is about how AI based startups today are so vastly different from the other kinds of startups that VCs were investing in historically. And we've got more on that coming right up. Oh, man, that was great, Vince.
Vince 55:44
Thanks. Sorry, I went a little over time.
Jon
It's okay, I let you!
Ben
It also tied back to the beginning when we were talking about simulations because first you have chemical simulation, protein simulation, cell simulation, organism simulation, where somewhere in the future, maybe we're not alive, we're doing organism simulation and you actually simulating, you know, what is the brain IQ based on these, you know, theis gene sequence? And I know that's like total moonshot, we'll probably never live to see that. But theoretically...maybe that’s possible. And we're definitely gonna see human engineering, like we're already seeing it with the Chinese use case.
Vince
We'd also have to solve the nature versus nurture problem. How much of a brain is, is just the the sum of the experiences that brain has had in its lifetime? You know, post postnatal, yeah.
Ben
And you you hit the nail on the head, because it all goes back to: You have to simulate their environment, and now we're all in a simulation.
Vince 56:48
That proves it.
Ben
Yeah, that proves it fact. This was discussed on the podcast 1000 years ago. Fact.
Segment 3 on AI Startups 56:58
Our final topic for today is brought to us by Andrew, who has been a software developer and data scientist at data focused businesses, mostly in the financial industry for over 20 years is on the differences between the AI startups often invested in by venture capitalists today, relative to these Software as a Service tech startups that were the most popular investment category until recently. So we're talking about this because the prominent venture capital firm, Andreessen Horowitz, recently published a blog post on this where they suggested that there are three main issues facing AI companies today. Andrew, could you tell us about that?
Andrew
Sure, Jon. It's actually a great blog post, I am not gonna be able to do it justice. But it was quite interesting to hear their take on on how they value startups in general and then AI startup specifically. The idea of a software driven like a software as a service startup gets a lot of valuation given the fact that the costs are usually pretty low and they have a large gross margin. Especially now with the cloud, you can spin up an idea for a software as a service in an hours given how many tools and stuff that are out there, and then sell it to as many people as you can, so you get a high valuation from that perspective. And in this blog post, they sort of compare that with AI businesses that are a little bit more obviously data intensive and compute intensive. So that margin decreases. And and you really have to take care of them or value them a little bit differently because of that.
Jon 58:42
Right, right, right, right. So we have lower lower margins due to heavy cloud infrastructure usage. And then there's other issues as well, right with, you know, scaling challenges you have, you know, you build your machine learning algorithm and it works in a bunch of cases, you know, it works in 90% of cases. Well, guess what, that's a huge number of cases that are that are challenging and a require gonna require typically expensive data scientists to tackle these things.
Andrew
Right. So your margin is is is degraded by needing humans to constantly look at the model to try to capture what the edge cases are, once you get the model into production. I mean, it's very different from knowing that what you want your software to do and writing the 10 heuristic rules and then once you code them, you're pretty confident that you've covered you know, the the edge cases with it, you can kind of put that live and and until you change something that most the time that functionality is going to continue. With machine learning models, obviously a lot of things can change, and you have to keep an eye on it.
And then the third thing they mentioned, is sort of this, this concept of a weaker, defensive moat. So a moat is this financial term that's come up I don't know in the last five years maybe about how well protected is your revenue stream from other competitors in the business. So the wider the moat, think of it as a castle, the more protection I have the right to some new companies not going to come and take away all my revenues. And the idea is that AI actually has weaker defensive moats maybe because there's so much commoditization of AI at the moment. So where it used to be you needed, somebody that knew the deep, the deep weeds of the mathematical model to build your neural net, and deploy it in some performance c++ version. Now you can kind of just throw up your features into a Google platform and and let it sort of iterate over the different hyperparameters and you have a model and that works somewhat decently.
So there's some of that and then there's some concepts around data network effects, where just because you have your data is driving your model, that data can become stale. It can become cost prohibitive to acquire new data. That's not really adding value to your to your model. So those are the three sort of things that they've outlined as differences in AI software startups, so lower gross margins, scaling challenges, and then weaker defensive moves.
Jon 1:01:26
Right, well, so Ben, how, you know, is there anything else you think they're missing? Or how do we get around these issues? How does someone like yourself, build a successful AI business?
Ben
Yeah, I love this topic. And I think like you said, before we started, we could talk about this for hours and hours and hours. The funny thing, so I've been three years into the startup. The funny thing about doing a startup and actually getting pounds of scar tissue all over your face and on your shoulders, is if you could go back in time and talk to yourself three years ago pitching VCs, you'd want to punch yourself in the neck. You don't know. You don't know anything. And, and all these things resonate. And one of the things I was gonna throw on there was with SAAS, you get this and it’s not just SAAS, it’s just software, you get the multiplier effect, where you have like the React Twitter clone. And so if I told you 15 years ago, I was going to build this Twitter ecosystem, that'd be very daunting and intimidating. But today with serverless, Docker, Kubernetes, these these wonderful languages, you can do stuff like, you know, go back in time, a decade, and you would blow the socks off people what you can do in 24 hours.
But speaking about AI, one of the things we talk about is your model. Building a model, it's not, it's not your IP. It's not, it's not a competitive advantage. Your data is your IP. So if you can get access to some unique data sets that can become a significant moat, where it's not a script kiddie away. What I mean by that it's not like a TensorFlow blog post away from imitating your business. And we talk a lot about that. I'll pause there.
Jon
Yeah, that's a really good point. I mean, there's a lot of you know, even a lot of the biggest companies will open source their models and even their tools for building AI systems but the most successful AI companies out there they do not let you get access to their data, that's for sure.
Ben
Yeah, and in there's a general con... it's really surprising to me how bad AI companies are failing and this hopefully this doesn't come across as arrogant because I just admitted I was an idiot, like a major idiot years ago, like a lot of this has been learned.
Jon
But not anymore. Three, three years from now you'll look back and say, Yeah, I had it all out then.
Ben
All my good ideas are stolen. So all my good ideas. I take it for customers not paying us or like whatever. Yeah, there's a lot of confusion around: How do you build a successful AI company? And a lot of people, they chase AI is the shiny tool. So imagine me going into a prospect. Hey, do you guys want to go into business? Yes, we do. What do you want to do? Let's do an attribute model. What are we gonna do? We're gonna look at an image and say, socks sweater. And I think for a new founder, they're gonna run down that vein. But for a seasoned founder with scar tissue, they're going to hold up the stop sign immediately. And they're gonna say, we're not working on that problem. But the underlying truth is, what is it worth predicting a swimming pool from space for an insurance company? What is that worth? The joke is it's worth more than zero. But if you get into these other problems, you could find AI problems that are worth amazing amounts of money. 10s of millions of dollars, I would argue, in petroleum and oil and gas discovery, you could be looking at a $40 million SAAS contract if you can find the right problem. And so I'm not saying I found those problems, but it changes your perspective. Hunt for value. AI is a shiny toy. Everyone wants it. And that's a very dangerous thing.
Jon 1:05:02
Yeah, that is really good. Hey, Vince, you you really want to be exploiting petroleum, right?
Vince 1:05:10
That's not funny. Not funny at all. At what cost? Forty million dollars is it's not worth it when you're destroying the planet.
Jon 1:05:24
You're always saying coal is a future.
Ben 1:05:31
Don’t you want to run your droids on coal?
Jon 1:05:35
The dirtier they get, the better they'll run.
Jon 1:05:41
Yeah, no, I, you know, that's a very I mean, so actually, that's a really good you know, another way of kind of spinning that is there's a lot of value in a lot of energy sources. That's actually something I hadn't thought about: Even in solar or wind or other renewable sources, in ways that may be actually tying back to our discussion of viruses in the first segment of this program, you could imagine using machine learning on on particular proprietary data sets for, you know, a solar or wind manufacturer to devise ways that the, that the, the panels or that the, you know, the turbines can be designed in a way that is that much more efficient and you can extract, you know, you can get that kind of incremental value out of the system. There's a lot yeah, and it's a really good way of approaching problems. And so I guess that's exactly so that's what you'll be thinking with your next company, I guess is you know, Where are the values? Where are the million dollar contracts?
Ben
Yeah, well, and it's a little different than that, because for my next company, which won't be for a while, like four or five years, or or maybe never, but I think I'll get the itch again to go do it again. But for the next company, I love the idea of developing a product backwards because I don't really value opinions, including my own. And this even includes customers. So if I'm talking to a customer, hey, what do you want? What do you want me to do? I don't really value their opinion, I value their behavior. I want to measure their behavior.
Jon 1:07:14
What do you mean?
Ben 1:07:18
Sorry, I got a kid that's knocking. I'll try to get the point across, then you guys can chew on it. So the idea is, if I'm building AI, for engineers, and engineers say this is great. We really like this. These API's makes sense. I would rather measure the engineer’s behavior. So you and I would decide ahead of time, if this is really amazing. What should the engineers be able to do? There's, there's a very scary thing that happens in software, and that's called the recipe playbook, or the template repeat. So if I'm doing cloud, if I if you're cloud certified for Amazon or cloud certified for Google, I actually have no confidence that you can solve my problem. Because my problem’s new. You repeat, repeating a recipe or repeating a template is not a sign that you've learned something. And so when it comes to building a product or building an API, you really want something with grassroots, grassroots adoption, where you see intrinsic motivation for them to show behavior, share the product. Does that make sense? You're really looking at it. And the funny thing with that is you don't even have to build a product. You can build it backwards. You can do mocks, you can come with a PowerPoint, you can do a play by play scenario, and try to show the behavior. So yeah, I'll stop there.
Vince
Yeah, so really, really actions speak louder than their words. If they're not using it and buying it doesn't matter how much they love it.
Jon
Yeah, I think that's right. I think it sounded like Ben was dealing with kids showing up in his recording studio. So the beauties of recording remotely.
Andrew, do you want to run us through quickly the kind of the six solutions that Andreessen Horowitz recommended, and we can use that as kind of our endpoint for this program?
Andrew
Absolutely. So the six things they went through real quick: Eliminate model complexity. So that, you know, you need to build this trade off between whether you're doing custom model builds for all your clients versus you having one complex model that you expect everybody to use.
And the second one was, choose your domains very carefully, kind of what Ben was talking about. It was really, actually they were they were stressing like solve the narrow problem. Focus on high scale and low complexity problems. So you don't run into the education issues.
The third one was plan for high variable costs. Treat maintenance of the model as sort of your first order problem. You must account for that in your in your budget.
Embrace services. So, you know, you guys know that we've always try to back custom service work versus product work. They said embrace it to learn your customers, but pursue one strategy and very committed way.
Plan for change and tech stack. To me that just sounds like an obvious thing to do.
Vince 1:10:20
I don't know. I mean, we use COBOL.
Andrew
In your in your gas and oil well, right? No but plan for changing tech stack, especially around AI because it's definitely still in infancy. There's so much out there, but you have to remember that some of this stuff is 5, 10 years old at the most. And then they talk about building defensibility the old fashioned way: good products, get proprietary data that only you have. That's gonna that's going to be the selling point on your product.
Jon
That's right. Exactly as Ben was saying. All right, sounds good.
Ben 1:10:58
I was gonna add real quick, I loved what you said, Andrew. New founders, they get very ambitious that we're going to build a platform. And they they go after multiple industries. And the VCs are always saying focus on one thing, focus on one vertical. And, and even if you think you have a platform, or even if you think you have a wide product, if you have to come up with a different sales deck for like every sales meeting, that's, that's a waste of time. But you also don't understand your prospect and you don't understand the problem. So something I talk a lot about is find me a problem with a customer behind a customer behind a customer behind a customer where there's actually this sense of momentum. And so I think in the future, five years from now, 10 years from now, 20 years from now, if I ever do another startup, I'll actually just grab five customers immediately, and they don't even have to pay me. I don't even care if they pay me for a year. I will be hip to hip with them, take them out to drinks, take them out to dinner. I will just fix their problems for a year but there'll be big companies like Oracle, Amazon, whoever they are. And I'll just do the you know that five customers is huge. But I think young founders don't, they don't think about that because they want to take over the world. They've got a billion dollar startup, everyone has those ambitions. So I love what you said, Andrew.
Jon 1:12:11
That's great advice. Thank you very much. So it's been wonderful having you on the show, Ben, you had a lot to add. And it's too bad that we couldn't get you in person last week. And you know, that, you know, this, this Corona, social distancing situation. Not ideal, but we made the most of it here and we're lucky to be able to, to have you call in then from Utah and actually because of the way that we're currently under lockdown even in New York, everyone on the program today, all the untapt co hosts we're all having to call in. We do intend on having this podcast continued to be in the future. As soon as we can be together physically again, we will go back to recording live and having that nice video feed for all of you to watch if you're interested in that.
And yeah, so thank you very much for joining today's program. We talked about how data and machine learning are leading the fight against Covid-19 so that we can get back to recording, live broadcasts, and how you can help whether you write software or not. If there's one takeaway from today's podcast, it should be that you should download the Folding@home software so that your computer can make strides against the coronavirus while you sleep. We also talked about mind control prosthetics powered by machine learning, and how AI businesses differ from the tech startups that predominated yesterday. So please do like, subscribe and follow our podcast. We're across Apple podcast, Spotify, Google podcasts, and we also will we provide video content on YouTube although I guess for this episode, it's going to be a relatively static YouTube experience. And then we also a few of us have handles for reaching out to us. For me for Twitter, I'm @JonKrohnLearns. Grant, do you want to give us your Twitter handle for in case people have ML applied to virology questions?
Grant
Sure. It's @grantbey.
Jon 1:14:20
Great. And I suspect Ben, you must have a Twitter handle.
Ben
Yeah, it's @bentaylordata. That's also my LinkedIn handle. If you send me a message on Twitter, I might respond a month later. LinkedIn, I'm pretty active. Yeah.
Jon
Yeah, it's interesting. Um, we were talking about that earlier on the program that really for what we do in software development and data science, it's interesting how rich the LinkedIn experience can be. And that statement is a testament to it. Yeah, and feel free to sign up. I have an email newsletter at my website, jonkrohn.com. There's an email newsletter where I post every podcast as well as other video tutorials and such that I publish, which are all free. And do feel free to add any of us on LinkedIn. We'd love to talk to you there. Many thanks to untapt for bringing us together as a company, to Sangbin and Maria who produce the show, and of course, to our guest, Ben Taylor, thank you very much for being on the show. It's really been wonderful.
Vince
Thanks Ben
Andrew
Thanks Ben
Grant
Thanks, Ben. See you soon.
Ben
Yeah, it's been really fun.