Skip to content
0011: Evan Prodromou, AI practitioner (part 1) episode artwork

Feb 16, 2016

0011: Evan Prodromou, AI practitioner (part 1)

from Concerning AI | Existential Risk From Artificial Intelligence Hosted by admin

Evan's awesome. Great to talk with a bona fide AI practitioner. Just getting things started. Concerning AI | Existential Risk From Artificial Intelligence is made with Jetpack Podcast. Full show notes and every episode at https://concerning.ai/

Episode details come from the podcast's RSS feed.

Embed this episode

Ready to play

0011: Evan Prodromou, AI practitioner (part 1)

of matches

Transcript Auto-generated, may contain errors

All right, well, this is Brandon. And this is Ted. And we're concerning AI. And today is a special day because.

Very special. We have our friend Evan Padromo. That sounds your name correctly, Evan. That's perfect.

Thank you. We have Evan on the line with us here. And we're so excited to have you here with us, Evan. And why don't you tell us briefly about yourself and why are we talking about AI with you?

Sure. So my name is Evan Padromo. I'm a software developer and entrepreneur. I've been working on developing software and websites for 20-plus years.

I am very interested in ways that software can help people collaborate, help human beings work together, share knowledge. And I'm always looking at ways for us to improve that process through technology. So as a software developer, over the last few years, I've been very interested in artificial intelligence and artificial intelligence techniques for improving the software process. Awesome.

So I do have to say that Evan is a friend of ours that goes back from our early days in the Wiki community. And so Evan has been an inspiration with one of the popular early Wiki sites, WikiTravel, that was one of the first really awesome community collaborative, wide open, create content that's awesome. So Evan has always been democratizing information and leading us in interesting directions for years. Let me tell a quick story about part of Evan's inspiration.

We were in Denmark for the WikiSim conference. And at the time, I can't remember how old your daughter was, Evan. She was just learning to walk. But you were telling us how there was a little tiny threshold between the hotel room and the bathroom where you were staying.

And she was pretty good at walking. But when she got to that tiny little threshold that adults don't even notice, she would fall down every time. And we all drew inspiration from that because that's how we often treat people on the web. Oh, it's just a little annoyance or whatever, with software development or whatever.

And we make things harder than they have to be. But the tiniest little thing could throw people off. Yeah, absolutely. And especially when we're talking about people's free time, when they're working on tasks that they don't necessarily have to do, that they're doing in a volunteer way, that's a little impediment.

It's going to be just enough to knock them off. That's a great anecdote. I forgot not one actually. So I started WikiTravel in 2003.

And then in 2008, I started an open source social networking company called StatusNet. And probably the biggest site using it was Identica. So that's one that many lot of people know. And then a few years ago, let's see, 2013, I was the founding CTO for a company called Breather.

And Breather's where my interest in artificial intelligence techniques really came to a head. Are you guys familiar with Breather or have you heard of Breather? OK, cool. Breather is an on-demand service that rents rooms by the hour.

So it's a network of rooms in cities like New York San Francisco. It's not everywhere in the world, but it's in New York San Francisco, Boston, here in Montreal, where I live, just opening a link. And it lets you rent rooms by the hour from your cell phone. So you can just go on, find a room, book it, and be in the room immediately.

So it's a very short-term room rental system, great for meetings. Or if you need to get some work done, network of private rooms. It's a really exciting project. And I was really happy to be part of it when we were starting the company.

But as a company, as a CTO, I realized that we had a number of problems that were tractable with machine learning techniques or AI technique. Things like identifying where we should put rooms in a city. So depending on what kind of foot traffic is in an area, what the real estate prices are. So you're actually owning these rooms.

It's not that it's not like Airbnb or something like that. Yeah, so it's owning the rooms rather than having a network or a marketplace. So that's exactly it. So we would choose where to put the rooms within a city.

So being able to identify where to put those rooms based on data that we might have about the rooms. Or the rooms are cleaned after every reservation. So we have a team of cleaners that are in each city and scheduling the time for those cleaners so that they're going from room to room optimally, that we're not wasting their time, that we're not rushing them too much. Is that the classic traveling cleaner problem?

It's absolutely the classic category. It's a little bit of a variation because you can have the question is how many salesmen do I need to put on this traveling, on this circuit? It's about the same, right? And then as well as say dynamic pricing.

So making sure that the prices of the rooms, what's the best price to put the room at for long term benefit of the company and the user, right? And these are all problems that you can point big data techniques at. So some of the machine learning techniques that are very popular right now would work in those situations if you have a lot of data. But for us at the time that I was CTO of breather, we didn't have a lot of data.

We had hundreds or thousands of transactions that we could use for training sets. But that wasn't really enough to really get a good machine learning process going. And so we were a data poor company looking at data rich techniques for deriving information. Can you unpack data rich techniques?

Sure, absolutely. So the term that a lot of us tend to use around this is big data. Companies that have millions or billions of rows or trillions of rows of transaction data, click data from websites, location data, just tons that they want to extract some value from. Google and Facebook come to mind for me.

Are there others that we should be thinking about? Who has actual big data? Anywhere from Amazon is a great example of a company that's used their data from their online site to generate a lot of intelligence and to provide more intelligent behavior in their application. So Amazon is another great one.

Facebook and Google are great ones. Twitter is also one that's doing some really good work in here. I guess Tesla might qualify there. Tesla probably qualifies.

Yahoo has done some great stuff. They did some incredible work on their Flickr corpus. So it's a really interesting work with Flickr. But the thing is, is that you need a lot of data to make most of these techniques work.

And it's the companies that have a lot of the data that are able to hire machine learning experts out of universities and get them to work on these problems. So there's not very many people in that position. Yeah, it's kind of a small click of folks. Now, the techniques aren't, they're not that difficult.

So when we talk about machine learning or deep learning, they're not that hard to do. Smart developers can kind of get into them, but you really need to have a good set of data. And you tend to need to have one or two really smart people to make it work. So some PhDs who worked on machine learning.

Not a lot of companies have those resources. Yeah, Ted and I have talked about three things that we might need for intelligence, one being substrate, like how many cycles you've got, all that sort of stuff, one being training. That's the big data that you're talking about. And then the last one being craft or the techniques themselves.

It sounds like you're saying that the data rich have kind of all three of these because they have a lot of data. And the rest of us don't have that second important thing, which is all those training examples. Right, exactly. OK.

So yeah, and when people kind of in the machine learning world talk about this problem of trying to generate a strong intelligence when you don't have a lot of data, it's called the cold start problem. Because you're trying to get a cold, it's like starting a engine cold, right? It's not easy to get a machine learning process started without a lot of data. So when I was at Breeder and we had these problems, I kind of, I was like, hmm, this is a real opportunity, there are some clearly a lot of people who are interested in getting involved in an AI who don't have the resources to really use these techniques.

So after I left Breeder in 2014 to pursue my own projects, and this was the one that really stuck with me. So my partner and I looked into various techniques. And we landed on one that's called Fuzzy Neural Hybrid or Fuzzy Neural Inference. Basically, the way that you do it is you take, you do, instead of having a lot of data to kickstart your machine learning process, you use rules or heuristics.

So you use some logic. And in our case, it's Fuzzy Logic, which is a kind of vague logic, which means that you take kind of intuitive firm. Oh yeah, that's a good show. Maybe we should call it Fuzzy Shows.

I think that's really good. So anyway, that's the area that I'm working in right now. We have a company called Fuzzy I.O. And we have an API that lets developers get started with machine learning, starting off with these rules sets.

So that's why I'm kind of in this practitioner mode of AI right now. But obviously, thinking a lot about the long-term issues. I think that as you guys have clearly shown, it's an issue that's coming up on society very quickly. And I think that people who are involved in software and involved in the internet see it coming a little bit sooner than the general population.

And we kind of have a responsibility to think about the way that we treat it. And I think that this question about data has and data have not is one of the things that concerns me the most. So I want to make sure that in the same way that there's kind of a democratization of software through the internet that we can have a democratization of AI. And I'm not sure you guys are going to be excited about that.

But it's something that I think is important. It's actually, this is Brandon. It's actually the sort of hopeful place that I've latched on to trying to be hopeful about the future has been definitely involves there being a plurality of AI's instead of one winning winner take all AI that sort of comes to dominate everything. And I just I don't see that happening very easily unless there is more of this open sharing democratization.

So I think I'm a I would say a fan boy of that. I don't know a lot about it. So I don't know. I'm curious.

But the techniques that you're working on now, are they going to lead to artificial general intelligence? I think almost definitely not. All right. Did you get that out of the way right away?

No, that's absolutely the case. So I think that when we talk about AI and AI techniques, probably what AI means to practitioners today are really any techniques that aren't about procedural code. So procedural code being where I say computer, do this, then this, then this, then this, then this. And that's kind of the way that most programming works and has for the last 50 years since programming has started.

AI techniques are ones where we are doing more learning, where we do kind of do what I mean development, rather than do what I say development. So the kind of techniques that are expecting more out of our software than just following directions, one after the other. And so that doesn't necessarily mean that the kind of neural networks or deep learning techniques that we see today are even close to where general artificial intelligence would be. But I think that they're probably going to be something that is part of the, let's say, the reflases of the next AI.

So they'll probably be somewhere in the low level processing. I think they'll be a component, but I don't think they'll be at the same level as a consciousness of those next levels of AI. But at the same time, you have to build your way up. So I think all of us are at the plenary or maybe cockroach stage of intelligence.

But that's a step on the way. And it's an important one. So we call this show concerning AI because we're interested in AI concerns. And you mentioned thinking that we're moving there faster than maybe the ordinary person on the street would think.

What sort of vision do you have? What are you concerned about? What are you excited about? What visions do you have of our future?

Oh, that's a great question. Well, I think that there are a couple of things that I'm concerned about that I'm excited about. I think one of the things that I'm very, that I kind of envision for the future is that when if we have general, generally intelligent software or artificial creatures, however you want to, artificially intelligent creatures, however you want to be able to define them, they'll probably be a lot different than human beings. So they will probably be very significantly different in the ways that they think, in the ways that they are.

They may not have the kind of sense of identity or sense of autonomy that we expect, right? Because we are putting most of our effort right now into using these techniques to solve real world problems. And we're not spending as much time trying to develop my friend the robot, right? So where we're putting our development dollars and our societal effort is into doing things like solving problems for where a cleaner is supposed to be at 225 today or where to put a room or where your Uber is supposed to arrive, right?

So those, the software as it becomes more intelligent is probably going to be acting in a way that serves those purposes. Whereas us human beings, we developed our intelligence for other kinds of problems. How do I eat? How do I find a cave to sleep in?

So that's where our intelligence comes from. How do I interact with other human beings? So it might not look like us. And it might be hard to identify as intelligent, right?

So yeah. And I think that might be also a big issue, right? So it might not be easy for us to look at something like the Facebook feed algorithm or the Google search algorithm and say, okay, this is an intelligent creature, right? And I don't think that we would treat those, that kind of software today as being generally intelligent.

It lacks the autonomy, it lacks the sense of consciousness as you guys have addressed. I think that we will probably cross some of those thresholds as a society without paying attention to it. And I think that that's where some of the dangers happen, right? So.

Well, I think in some ways you're pointing to an idea that we've noticed that every time an algorithm solves a particular problem, we stop thinking about it as intelligent. Right, exactly. So you guys mentioned the Go playing software last week. Yeah, and the check playing software.

Yeah. And I think that the problem with not recognizing things that don't look like us as intelligent is that it can probably let us be kind of inhuman to those things. Once they are achieving some kind of personhood, we may not be willing to acknowledge them as people because they will look so different than we are, right? So I think that's the problem.

You're suggesting maybe we have a human rights is the wrong word, but well, maybe not. I mean, are they human? Are they people? Are they people?

Are they not human? Maybe not human people. Yeah, person for some sort of artificial intelligence entity rather than a human, which I think tends to be more homo sapiens sapiens, right? That tends to be more like R species, right?

But as you guys have also pointed out, there's some crossover, right? I think that will probably have augmentation be a big important part of what, you know, where AI goes over the next 10 to 20, 30, 40 years, right? So. But it's just, yeah, go ahead.

That's interesting. So 10 to 20, 30, 40 years. Like, I think it's fun to have a kind of a real practitioner on the show here and to pick your brains, Evan, for your brain for. So yeah, so here are two questions.

One, do you think that it's close, like artificial general intelligence is close? And two, what does it mean to be close or not be close in terms of time frames? Yeah, so I think close or not close in terms of my lifetime. OK.

It's probably, you know, constrained to about 25 feet would be my guess, you know, so like first half of the century. And I think that it's likely that we will have that kind of, we will have something that you and I would acknowledge as a person by 2050. I think that's a very likely situation. We're in a very fast, we continue to be going through very fast technology cycles.

It doesn't seem to be slowing down. This is an important place for us to be going. We are pushing the limits of what, of how human beings can deal with this level of technological change. And we are leaning more and more on software agents to do that or artificially intelligent agents to do that.

And we're going to go through a few iterations of that, probably five, 10-year cycles of that. And it's likely that we're going to hit that personhood level sometime before 20, 20. And that's just, that's just for me, from a practitioner point of view. That's mind blowing.

I think that is mind blowing. To hear you say that, just so baldly out, just boom, laying it out there. That is mind blowing. Yeah, and I think that a lot of people who are in the practitioner community, they tend to be very focused on what's happening today on how things are going today.

And I think that a practitioner today would say something like this software that just identified features, billions of features in Flickr images. And was able to say, this is a beach ball, and this is a spaniel, and this one is an ash tree, or an oak. That's right. That's right.

And there's a difference between an ash tree and an ash tree. But even though it's able to do that, because it's not Robbie the robot, it doesn't quite feel like it's general artificial intelligence. And I think that we would all agree on that. But it's happening so quickly.

There's a lot of competition happening right now. It's clearly something that has business value. So it's clearly something that has value for these larger governments. I think the big thing for me is making sure that there's a lot more participation across the board.

So from the same kind of people who are able to make websites or to develop mobile apps for phones, are able to be playing in the same space. And that's not where we are right now today. That's very smart. Tell us.

You mentioned that earlier the data has, and the data have nots, and fuzzy IO, and trying to democratize access to artificial intelligence, and learning techniques, and stuff like that. Layout a roadmap. Inspire us, give us a vision for how that could happen, what the path might look like. So one thing is, I think that your typical web developer today is very tied up with procedural code.

That's what they've been working on or a mobile developer. So they work in programming languages like Objective-C, or Python, that are very much about do this, then do this, then do this, then do that. And I think that probably the next big step is to be introducing non-procedural techniques into the normal tool set of that kind of developer. And they need to see some value from it.

They need to say, I can do tasks with these kind of techniques that I couldn't do before. And I can look good doing it. It's good for me. It's good for my customers.

It's good for my boss likes that I do it, right? So that kind of thing. What's missing in that progression? Just what are the ingredients that are missing today?

Yeah, you tend to, like I said, just a slight shift in the way that the techniques are, right? So being able to do a more of a mix of logic and learning. So we're in a very learning mode right now with AI. It tends to go up and down.

And learning is definitely where things are at right now. I think logic is an important part of being able to democratize that process and mix logic and learning. So that's an important thing. But I think the more people that we have involved across the spectrum of small organizations, nonprofit organizations, individuals, groups, big company, small companies, that means that we have a real diverse set of software developers who are playing with these kind of technologies.

And I think that's really the first step. And I think that all of us, our experience with the internet, our experience with collaborative systems, tends to, well, it led me to think that open systems with a lot of participants tend to be more generative, to tend to create more value, and create more interesting results than ones that are just within smaller ecosystems. So I think that's probably the biggest thing right now. Sorry, I got a little.

Would that, you know, the open systems? I'll admit, this podcast has made me want to start dabbling in AI again after 10 years of not being involved in it all. And one of the first things I wanted to do was start working on some speech recognition software. And the piece that was missing is the data, more than anything, the training data.

But that said, there are all of these public domain books and actually audio book versions of them out there that I think could be turned into something that would be useful to people. Otherwise, you have to pay for these data because they aren't even very good. Or else you're a big company that that's part of your secret sauce that you don't want to license. Oh, you lost it.

For the conclusion of this conversation, please download the next episode.

Frequently asked questions

When was this Concerning AI | Existential Risk From Artificial Intelligence episode published?

This episode was published on February 16, 2016.

Is there a transcript available for this episode?

Yes, a full transcript is available for this episode. You can read the complete transcript on the episode page.

Can I download this Concerning AI | Existential Risk From Artificial Intelligence episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.