SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks...

23
SDS PODCAST EPISODE 67 WITH RISTO MIIKKULAINEN

Transcript of SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks...

Page 1: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

SDS PODCAST

EPISODE 67

WITH

RISTO MIIKKULAINEN

Page 2: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

Kirill: This is episode number 67 with Vice President of Research

at Sentient AI, Risto Miikkulainen.

(background music plays)

Welcome to the SuperDataScience podcast. My name is Kirill

Eremenko, data science coach and lifestyle entrepreneur.

And each week we bring you inspiring people and ideas to

help you build your successful career in data science.

Thanks for being here today and now let’s make the complex

simple.

(background music plays)

Hello and welcome to the SuperDataScience podcast. Today

we’ve got a very interesting and exciting guest, Risto

Miikkulainen. So there's a lot of things you need to know

about Risto and we probably won't have enough time to go

through them in this quick intro, but here are a few. So

Risto is a professor at the University of Texas in Austin. He's

been working in artificial intelligence for the past 40 years.

He is also a published author and he's published book titles

such as Computational Maps in the Visual Cortex, Sub-

Symbolic Natural Language Processing, and others. And also

Risto is the Vice President of Research at Sentient

Technologies, an Artificial Intelligence startup in the Bay

Area.

So as you can see, our guest today is an expert in the space

of AI, and that's exactly what we're going to be talking about.

And what was interesting today was that we just met with

Risto today online and we were going to record a podcast

some time later on, but we started getting into such

interesting topics already, so we decided to record a podcast

Page 3: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

already. So this is going to be a quick one because it was

quite short notice, Risto didn't have that much time, but

hopefully we'll be able to get Risto on a future episode some

time down the track.

And another thing that I wanted to point out is Sentient AI.

So Risto is working for Sentient AI as their Vice President of

Research. This is a great startup in the space of artificial

intelligence, it's been around for 10 years, and the whole

purpose of this podcast was to help Sentient get their word

out there and possibly influence people to get more into AI.

It's very cool how passionate they are about artificial

intelligence, and we're very grateful that they've "lent", so to

speak, one of their top talented people to come on the

podcast to share some of the top and most cutting edge

things that they're doing in the space of artificial

intelligence. So you'll get quite a few cool applications of AI

from here.

And also from this podcast, you will understand, hopefully,

or see for yourself, why it's so important to slowly start, if

not getting straight into the space of AI, but just keeping it

on your radar and considering what possibilities and

opportunities might exist for you in the future in the space

of AI and how quickly it is coming ahead. Because

ultimately, this is indeed something to look out for if you're

building a career in this space, the space of data, and the

space of data science, and in the space of technology.

So there we go, we've got quite an exciting chat ahead. And

without further ado, I bring to you Risto Miikkulainen, the

Vice President of Research at Sentient Technologies.

Page 4: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

(background music plays)

Welcome ladies and gentlemen to the SuperDataScience

podcast. Today I've got a special episode with Risto

Miikkulainen from Sentient AI. Welcome Risto, and thank

you so much for joining us on such short notice.

Risto: Thank you. Happy to be here.

Kirill: So tell us a bit, where are you calling from? It was such a

surprise, I was expecting you in San Francisco, and you're

completely on the other side of the world!

Risto: Yeah, today I'm in Helsinki, Finland. The weather is just nice

here, the sun doesn’t go down so you get a lot of work done.

Kirill: Yeah, that’s fantastic. I heard at this time of the year it’s still

sunny at midnight.

Risto: Yeah, it is kind of a little dusk, but it comes up right at 2:00

A.M. again. It’s hard to get any sleep.

Kirill: That’s so cool. It’s interesting as well, at the start we had a

quick chat about your background. You come from, or you

were a professor at the University of Texas. Is that right?

Risto: Yes I am. Actually, I’m on leave, have been on leave for a

while. I’m at UT Austin in the computer science department

there.

Kirill: Okay, cool. And give us a bit of a rundown, if you don’t

mind, about the work that you do at Sentient and what is

Sentient all about.

Risto: Yeah, sure. Sentient is one of the best funded start-ups

currently in artificial intelligence. We started with stock

Page 5: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

trading and have added intelligent commerce and digital

media to that. We build products that are based on cutting

edge artificial intelligence, in particular deep learning and

evolutionary computation, so that is machine learning that

not only predicts but also optimizes what you’re doing on the

Internet or in the stock market.

Kirill: Okay. And I had a quick chat with Babak who is the CEO,

as I understand, of Sentient AI, and he gave me a quick

overview of the three products that you have. Could you talk

a bit more about them, those current three products that

you offer?

Risto: Yeah, sure. So, the training product is what you might

expect. It’s stock trading. We evolve, using generic

algorithms, stock traders that are completely autonomous. I

mean, it takes a long time to evolve them. We’ve actually

built a whole grid computer-like system where we evolve

these traders for several years. 40 trillion of them were

tested in this process over one year, and some good ones

were found. And these traders are sets of rules that tell the

system how to trade stocks. They observe the time series

and then they make trading decisions on their own. In this

way, they are completely autonomous. We of course monitor

them and make sure they are not getting out of their league,

but that very rarely happens. They actually have been

trained and tested so well that when we actually deploy

them, they do quite well. So that’s stock trading.

The second product is intelligent commerce. The idea there

is to build an interface to retail, for instance, some kind of e-

commerce site where there’s lots of items in the catalogue

Page 6: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

and the users, the customers, would like to see them. They

would like to be able to navigate that space of all possible

products conveniently. Not by using keywords, having to

guess what keywords are relevant, but by looking at the

product and saying, “Oh, I like something like this,” “I like

something like that,” and AI will figure out what it is that

you want.

For instance, if you are buying sunglasses or shoes or other

apparel, you get a first window that shows you what kind of

shoes there might be and then you can click on those that

are more to your liking and you get another sampling. And

this way, after a couple of clicks you usually find what you

are looking for. The beauty of that is that in a normal

keyword kind of search, the customers might see 15% of the

catalogue, typically in a week 15% of the catalogue is shown

to the users. Using this kind of search that’s based on users’

input and feedback, 70% of the catalogue is actually shown,

which is much better for the vendor, it’s better for the user,

for the manufacturer of shoes, they get more visibility. So

this is an intelligent interface. It is based on user’s feedback.

In a loop, AI figures out what it is that you’re looking for.

That’s the second product, that’s Sentient Aware.

And the third one is Sentient Ascend, which is in the digital

media space. It means that, for instance, we are optimizing

web interfaces. So, again, in an e-commerce site, it’s a very

important concept for the landing page, how good a landing

page – is it going to convert the user to a paying customer or

somebody who signs up for the membership or something

like that? There might be a goal for this website to convert a

user into something. And it turns out it’s actually quite hard

Page 7: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

and quite difficult and it’s a whole art of it and so on, how to

design a webpage so that this will happen.

And there are general principles of design, of course, like

you have colours that are consistent, you should be clear

where the action is. But also things that are harder to

understand, like how do we relate to the user, what kind of

images to show, what is a good call to action and so on. And

there’s a whole area, discipline, called conversion science

that has emerged that tries to undercover these principles.

And it turns out humans can do a pretty good job, but if we

let AI do the design, we can do quite a bit more. Human

design – you can only think of a couple of designs and test

them and evaluate them. But if you let AI do the designing,

particularly evolutionary algorithms, you can test a lot more.

And you can test combinations and you can make decisions

quicker and find pages that convert that humans would not

come up with very easily. For instance, one of the sites that

we’ve built, we only optimized the widget, which was the call

to action and a button, and they called it “the ugly widget”

because the colours were very flashy, but it turned out that

was what the users wanted. They wanted to see where the

conversion was and they clicked 43% more often using that

widget than the human-designed widget. And this can be

extended to many other things, but that’s a great example of

Sentient Ascend, this kind of digital media optimization

using evolutionary algorithms. Evolution is doing the design,

not humans. So those are the current three products. Yes.

Kirill: Very interesting. It reminds me of that situation they had

with—remember AlphaGo beat the world champion of Go

last year in South Korea? And it had a few moves where it

Page 8: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

was putting the pieces not on the first, second or third line,

but on the fourth line, and nobody has ever done that in the

past 3,000 years of the history of Go and now this has

revolutionized the whole—everybody is now thinking and

analysing that because it led the computer to win. So

everybody is now “Whoa, maybe we should do that as

humans.” That’s really very different.

Risto: Yeah, that’s exactly right. The computer, the AI, does not

care. It does not have the same prejudices as humans do. It

plays very cool Go and it makes very cool optimization. I’ve

got to tell you another story. In addition to those three

products, we of course have various proof of concept projects

going on. And some of them are not even immediately

products, they are something that’s just cool to do. So one of

them is cyber-agriculture, which we are doing with MIT

Media Lab, Caleb Harper’s lab and Phenom, their for profit

arm.

This is a very interesting project. The idea is that they can

build containers where you can grow plants – all kinds of

plants, food, maybe fibres, maybe medicines, and you

completely control what’s in them: temperature, water, light,

nutrients, everything is controlled. Now, vertical farming

existed for a while, but what if you can completely control

everything? It turns out nobody knows what to do. There are

some rules of thumb, you know, you should have sufficient

water, for instance, sufficient light, but how do you really

optimize the growth? So we let the same kind of evolutionary

algorithm figure this out and we ran a loop so they planted

what our algorithm suggested and three weeks later we get

basil that’s growing according to that recipe. And it turned

Page 9: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

out that we discovered some things that we already know, AI

discovered it.

For instance, if you try to optimize the size of the basil plant,

you kind of lose on flavour. But if you try to optimize the

flavour, you get smaller plants. That’s kind of what you

expect – strawberries, tomatoes are like that – but the more

interesting thing was that we were optimizing the light, the

kinds of light, how much light, how long the light period was

during the day. And exactly like you suggested, nobody

thought the plants would thrive if there was full daylight all

the time. Everybody expected that 8 hours of night is

necessary, turn the light off for 8 hours. Well, the algorithm

very quickly discovered, “No, you turn on the light and it’s

on 24 hours and the basil grows much better.”

So this is something the people did not anticipate, but the AI

discovered it again very quickly. So, yes, these kinds of

opportunities exist now, especially for the algorithms that

discover solutions, don’t just imitate and mimic and predict

data as it exists, but they have an opportunity to be in an

exploratory mode. They try out things and then we can use

predictor to see what actually might come up and they can

optimize and discover things that we don’t already know.

Kirill: That’s very cool. I just actually yesterday read a quote by

Mark Cuban. You know, Mark Cuban is on “Shark Tank” in

the U.S., and he said the world’s first trillionaires are going

to be AI entrepreneurs and AI start-ups. It sounds like you

guys are right on track for that, right?

Risto: Well, that’s the plan, definitely. AI is everywhere, and it will

be everywhere, so currently a lot of the excitement is really

Page 10: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

in the algorithms and methods, but it’s time now to build

real products, things that people use. The techniques are

powerful enough. For instance, this website optimization

that I mentioned, we have now consistently applied it to

many, many customer sites and consistently they beat

human design, routinely. And this is now in a real world

application, where there are real users who are not AI

experts, but are actual users who want to buy something or

want to sign up for something. This is what’s really exciting

to me.

They’ve had in the evolutionary algorithm conference,

GECCO—for a long time there’s a competition called “human

competitive results” competition. And for many, many years

people have shown that evolutionary algorithms can discover

these things that we don’t know how to do as well. But this

is the first time to my knowledge that it’s actually out there

beating human design over and over again every time it’s

applied. That’s really exciting.

And yes, the sky is the limit. That’s kind of cool. And it still

requires human ingenuity in identifying the problems and

giving evolution enough space to work with, to explore. So

we have to think about that, but that's the kind of jobs that

will be created, is to work with the AI systems that create

new things. You give them the problems, give them the

parameters and give them tools and the AI will then

discover, given that space, something really good.

Kirill: Okay. That’s fantastic. And you mentioned evolutionary

algorithms a couple of times, whereas something that we

have already talked about in the courses that I teach with

Page 11: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

Hadelin de Ponteves is reinforcement learning algorithms

which are also AI. As I understand, those are two different

types of artificial intelligence. Could you comment on that

for our listeners to just give it a quick overview of what’s the

difference between the two?

Risto: Yes, absolutely. AI is a very broad area, there are many

methods. And even machine learning has many methods in

it. Most of the machine learning that you see currently is

based on supervised learning, which means that you have a

dataset, “This was the situation and this is what happened.”

It allows you to predict, given the ground truth. Like meta-

prediction, for instance. You know what happened the next

day and you can learn to predict that.

Now, reinforcement learning and evolutionary algorithms are

a little different in that you don’t know what the right action

is. You have to explore and find that action yourself, or

really, the algorithm has to do that. So in that sense,

reinforcement learning algorithms are in the same category.

Not in the supervised learning, but in the reinforcement

learning discovery exploration category. So what you require

is some kind of a system where the algorithm can exploit,

can try out things and see how well they turn out. You don’t

know what they should do, what move they should play in

Go, or what trades to make in stock trading, but you can tell

after a while how well they are doing, how often you win in

Go, how much money you make in stock market. That’s the

feedback. It’s not the optimal action, it just tells you how

good your action is where you may try it. And then you

explore many different actions.

Page 12: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

Now, in reinforcement learning, the approach is to evaluate

values for the actions in different states and then try to

interpolate between those actions that you see. In

evolutionary algorithms, you are evolving the entire system

at once. You are testing it at once. You are not assigning

values to individual actions. And in that sense, it’s still

different. And what happens then is that reinforcement

learning is perhaps best suited for learning over the lifetime

of the agent, when you’re learning while you’re performing

and it counts, what you’re doing.

In evolutionary algorithms, they are better in discovering

good final designs when you can explore very broadly and

spend a lot of time and effort into exploring avenues that

might be more wild, perhaps, and there’s not big cost for it.

So, for instance, if you have a robot that has to learn in real

life on your physical system, you might want to use

reinforcement learning. If you have a simulator for that

robot, you can learn in simulation and you can try things

that you would not try in a real robot. And in the end,

evolution can then discover in the simulation something that

can be done on the robot and works really well.

So they answer a little bit different questions, but they are

both examples of this kind of learning, where exploration is

the key.

Kirill: Okay, gotcha. Very interesting. And if somebody was new to

AI and wanted to get started into the space and completely

new to AI, even new to machine learning, where would you

recommend for that person to get started?

Page 13: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

Risto: Okay, so this is very interesting. It used to be that you could

pick up a textbook that was written a couple of years ago

and be pretty much up-to-date, maybe look at a few journal

papers. And that has changed. Now the field moves very,

very fast. Three months is a long time now in AI. And it’s a

bit of a problem, it’s very hard to keep up with that, but

there’s no question you have to know the fundamentals first

and there are certainly textbooks and books that do that.

Right now, perhaps the most productive, hottest area is

machine learning.

So you might want to pick up a book that’s about machine

learning as opposed to more broadly in AI. But there’s much

that the rest of AI has also to offer, including various

reasoning systems and logic representation. And I believe

strongly that this machine learning that has been very

powerful in the last couple of years is going to stay powerful.

But eventually we need to connect what is being learned

with these reasoning systems. So that will also be part of it.

And if you want to know where the future is going, yeah,

learn about machine learning, but also learn more broadly

in AI. There are excellent textbooks in both those areas.

Now, if you really want to do the latest, then you have to go

to the web. You have to look at archive, you have to look at

blog posts. And there are some really good ones. It turns out

that sometimes the latest papers are very technical, kind of

hard to follow, hard to see the picture, but it turns out

there’s also a whole other community of people who write

these blogs that describe what these inventions in these

papers actually are about and put in perspective.

Page 14: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

Sometimes you can find those, they are mostly tutorials,

perhaps. A tutorial on generative adversarial networks, for

instance; a tutorial on probabilistic soft logic, various

aspects; tutorial on evolutionary computation or

neuroevolution. After reading the textbooks, pick up some of

those tutorials and you can pretty much catch up relatively

easy and new. But the field is very broad. So don’t assume

that you can catch up with everything immediately. That

takes years. But certainly catch up so that you can read the

literature and understand what the excitement is about.

Kirill: Okay, gotcha. I totally understand. And I actually agree with

that. It’s incredible what we’ve seen. Like, when we’re doing

the AI course which we’re creating now, there’s been so

many changes even while we’re creating the videos. You just

have to keep up with things.

I recently heard an opinion that AI is not only going to get

more sophisticated and more powerful in the challenges that

it can solve, but it will also get more accessible to people. It

will become more kind of drag-and-drop, there will be more

tools for people to use AI. Not the most cutting edge AI, but

over time things will become more accessible to everybody,

even the people who don’t want to pursue the technical side

of things. With that in mind, people who are philosophy

majors, who have the capacity for critical thinking, who have

creativity, we’re going to see an uptake in AI among those

people. Would you agree with that, or do you have a different

opinion?

Risto: I absolutely agree with that. Tools are very important and

this is a big part of why AI is now moving at the speed that

Page 15: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

it’s moving and so many people are getting into it. Because

there are tools that people share, including software

packages like TensorFlow, for instance. It’s open source so

people can also contribute to that, and people have. So in

that sense, you can take TensorFlow and download a model,

run it, and see what it’s doing within a couple of hours. And

then you can maybe build your own dataset and apply the

same model to your dataset and it won’t take that much

work to do that and you might already be doing quite well.

If you’re a retailer, if you have some medical data, whatever

it is, you can build a model very quickly. You still have to

understand what its limitations are. I mean, it’s too easy to

attribute a lot of intelligence into what’s happening, and you

have to understand what it can and cannot do, and therefore

you should read those textbooks and understand it. But

then actually running an experiment has become much,

much easier.

It used to be that my PhD students would have to take

about a year to build the software before they could run a

really good experiment, and this is not true at all. Same is

true of robotics. Now there are all kinds of robotics kits and

hardware that’s available and you can get started and that’s

how robotics is also accelerating rapidly.

So yes, I totally agree this is the best time and we need those

philosophers, we need the psychologists, and we need the

medical people. That’s how we push AI forward. The

applications are always exciting and more people get into the

field. And when new applications come out, there’s always

challenges and then us who develop the algorithms will have

Page 16: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

to look at it and see how can we improve the methods and

actually make this work. So I think it’s exciting.

I’ve never in my career, 30 years I’ve been working in AI, it’s

never been like this, that we are so close to application. We

are working with people who have the original data, they see

the patients, or they collect the statistical data from the

customers, and we can talk to them and we are immediately

in the thick of things. And I think this is the most exciting

thing about this cycle of AI hype, that it feels different

because it involves a much broader base of people, both

academics, people who develop algorithms, and those who

apply them, and those who use them and benefit from them.

So, I think if this is going to stay a little longer, we

absolutely need the people from outside, who are not AI

experts, to use these tools and do what they can.

Kirill: Okay, fantastic. That’s very interesting. And what are your

thoughts on the other side of the spectrum? So lots more

people are going to have access to AI, and we want more

people to use AI, but what about the people whose jobs are

threatened by AI, like chartered public accountants seem to

be getting slowly edged out by AI, even truck drivers and

slowly car drivers and so on? So there’s going to be lots of

change in the world. I know this is a very philosophical

question and there’s no one correct answer to it, but I just

wanted to get your opinion on it, being a person who is

driving this change in the world. What are your thoughts on

that?

Risto: Yes, that’s a very important question. And there’s no

question that it will require change in society and in jobs.

Page 17: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

This is a discussion that has to go on alongside with these

innovations and we cannot really have an immediate answer

to any of these issues. Now, one important realization is that

we’re not going to have AI replace people completely in all

areas. For a long time they’re going to be assistants. They’re

going to be making the job of the humans easier and more

productive for quite a while. That’s what we develop AI to do.

And even if you take, for instance, self-driving cars, they’re

going to have to work with people for quite a while and on

people’s terms. Same thing in all kinds of medical decision-

making. Stock trading – we have autonomous stock traders,

but we’re always watching over their shoulder and we decide

where to deploy them, what markets and so on.

So, at least for the foreseeable future, they are just simply

better tools for people who are working in those areas. But

in a longer period, yeah, there will be a change that we don’t

need humans to make those low level decisions, the actual

decisions anymore, or driving necessarily the actual cars, we

need people to plan where the car should go and like I said,

what markets to trade. So there will be a little bit of a

change. Those jobs that are doing the nitty-gritty actions will

be raised a little bit higher so that they can utilize the AI to

be more productive.

And there will be entirely new areas of jobs as well. For

instance, cyber-agriculture. That means that we will grow

plants anywhere. They can be grown underground, on the

roof, in the subway. And there will be jobs and markets and

maintaining in this kind of [indecipherable] agriculture

system. That’s what makes it interesting. It’s kind of hard to

imagine what the jobs will be like and it’s very difficult to

Page 18: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

imagine whether it will happen so quickly that there will be a

displacement of people who used to do the old job and it’s

very difficult to learn new jobs, or if it will be slow enough

that they would gradually be phased out and new people

trained for new jobs.

This is a discussion that’s on-going and we continuously

have panels and discussions, the government leading some

of those, there’s publicity events on this. And I think it’s an

important discussion. But my perspective, as in really my

personal perspective, is that I believe that there will be a

slow change where people are always working with the AI

and gradually learning to become more effective because

they use AI as tools and in the end we will be better off.

Kirill: Okay, gotcha. I understand. So the future doesn’t look as

bad as a lot of people are portraying it. There’s hope,

basically, for the human kind.

Risto: Yes, there’s definitely hope. Humans are very resourceful

and will figure out how to take advantage of this new tool

just like we’ve taken advantage of all the other tools:

electricity, steam engines, all those things. They all caused

transitional periods that were sometimes more difficult,

sometimes easier, but we emerged better off in all those

cases.

Kirill: Gotcha. Any thoughts on the Universal Basic Income, the

UBI, that’s being pilot-tested?

Risto: That’s getting a little bit outside my expertise. It would be

interesting if people could be motivated to do something

good with their time which they should be planning out

rather than just watch TV – be more creative or travel. I

Page 19: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

think the Universal Basic Income only works if there’s

something to keep people busy and engaged and interested

and motivated. And that piece is kind of interesting, how

that might happen. It might require cultural change, but it’s

certainly something we should research more.

Kirill: Okay, gotcha. Well, thank you very much for the overview. I

know we went a bit to the side. I’d like to pull back into

Sentient AI. Could you give us a quick overview of the team?

How big is the team? I remember Babak told me that you

guys have been around for a long time, like 10 years or so,

you’re coming up to your 10th anniversary.

Risto: Yes, that’s right, this year. The beginning was the trading

system and this is a great first application because it

requires nothing. You just have to have money and

algorithm and you can do it. You don’t need to have support

people, you don’t need to build anything. It’s a very good way

to get started. Totally the opposite end is this e-commerce.

You have to have customers, users, you have to have web

optimization, databases, cloud, everything. So it really takes

a village. It’s a huge team and not all of those are AI

researchers obviously.

Big computation [indecipherable] so we had to build a grid

computer system to just run these algorithms, 40 trillion

traders. And now we are running these deep learning

networks that we are evolving. They run on 5,000 GPUs

simulated in the grid, which makes that computer I think

the sixth most powerful computer in the world because

those GPUs have so much power when you put them

together. So, there’s a dozen people working on the compute

Page 20: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

and then there’s maybe another dozen doing the core

research, underlying algorithms, and then maybe double

that, 25 or so who are doing the development based on those

algorithms, how you build applications based on that, new

features in the products, and then our product development

team is maybe about the same size. And on top of that you

need to have customer support. You need to have

salespeople. That’s something that we are kind of different

from many AI companies in that we have real products and

we have real salespeople instead of just generating the

technology, cool demos. There are actual products and

therefore you have to have a very large sales force as well,

and marketing and PR and so on.

This is how maybe trillion dollar companies come about. You

have to have a well-rounded team. But interestingly enough,

the algorithm development is a relatively small team, like I

said, 12 people. The algorithms are powerful and you have to

have power algorithms, but then to build a product is ten

times harder. And this is what I’ve learned. Coming from

academia, having great ideas, having published papers,

algorithms, demonstrating they work, that’s 1/10 of the job.

Then that’s when the real work begins and I’m really

thankful to be able to be in a company that has those other

facets also covered. And a team that works together to

achieve those goals.

Kirill: Very, very cool. And you guys seem to develop products

mostly for B2B and then these businesses then apply them

to make the world better and maybe serve customers. Have

you ever thought of getting into the B2C space and creating

products directly for the end consumer?

Page 21: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

Risto: Yeah, that’s a good question. Maybe sometime in the future

if we find an application to those. Right now I think that’s a

little bit different space. What might be an application of

that? For instance, you could do optimization for

individuals, for their calendar maybe, for their navigation.

There are perhaps some such tasks that these techniques

could be applied to. But it is easy. Like I said,

[indecipherable 33:54] stock trading because you need

customers. And here we have B2B, which a major

investment can be done that way, and gradually build up

and achieve more of those kinds of knowledge and knowhow.

But eventually, I think when people get more comfortable

with having AI doing discovery for them, I think those

applications should also come up. That’s still a couple of

years out.

Kirill: Okay, gotcha. Well, I’m a bit cautious of your time. I know

you’re a busy man, you have to run somewhere just very

soon. I just wanted to ask you quickly, if somebody wants to

know more about Sentient AI and get to learn about the

company, are you guys open to having those conversations?

Are you possibly hiring at the moment? Or what are the best

steps that somebody who is interested in AI could take to get

more involved with Sentient?

Risto: We have a website like everybody and there’s a blog that

describes various things that are going on, especially some

of these algorithms and methods. And yeah, we’re always

hiring. Everybody always is. But there’s also

sentient.ai/careers, it lists the currently opened positions.

We are growing and we will grow in the future. So I would

suggest first take a look what we have there. We also have

Page 22: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

papers, they’re really technical. We have technical papers

published in GECCO and other conferences and those are

the real core algorithms. But also the blog posts give you a

general idea what’s going, for instance with the CyberAg.

Publication is still a cycle. They always take a long time to

come out. The blog posts are a little faster and a lot of times

those job announcements are tied to what’s happening right

now, so the blog posts might be a better idea. And if there’s

no job today there that looks a good match, come back

tomorrow. There might be.

Kirill: Yeah, I totally agree. Well, thank you so much for coming on

the show. I know this was a quick episode, but it was short

notice and I really appreciate you sharing the insights.

Hopefully Sentient will cover off some really cool applications

in the future and we’ll hear about the first trillion dollar

company very soon.

Risto: We’re definitely working on it. Thanks a lot for having me.

Kirill: Okay. Have a good day. Bye.

Risto: You too. Bye.

Kirill: So there you have it. That was Risto Mikkulainen, Vice

President of Research at Sentient Technologies. My personal

favourite part of this podcast was when Risto talked about

the cyber-agriculture. I think that was a really cool

application and I’d love to get more information or insights

into what comes out of it and we’ll possibly see it in the

news some time. This looks like a really cool idea that’s

going to reshape how we do agriculture and hopefully

produce more food for the world and solve one of the world’s

Page 23: SDS PODCAST EPISODE 67 WITH€¦ · help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple. (background music plays)

biggest problems of not enough food or not enough food in

the right places at the right time.

And apart from that, I hope you enjoyed how Risto spoke

about AI and what insights he gave into the different types of

AI and how quickly AI is advancing. Hopefully that gave you

some ideas to consider for your career for the future. Once

again, I’m very grateful to Sentient AI for encouraging or

spreading the word about artificial intelligence into the

world. So if you’re interested in learning more about this

company and their mission and what they’re up to, definitely

check out their website. It’s www.sentient.ai. There you can

find out more about what they’re doing, also you can find

their blog, there’s a link at the top and there you can read

some of their latest research and papers, as Risto

mentioned, maybe even some research that hasn’t yet been

published in papers so be on the forefront of artificial

intelligence technology.

And hopefully we’ll hear more from experts from Sentient AI

in the future on the podcast. And in the meantime, don’t

forget to connect with Risto on LinkedIn and follow his

career. You will find the link at the show notes on

www.superdatascience.com/67. And on that note, we’re

going to wrap up today’s podcast. I really appreciate you

taking the time today and I look forward to seeing you next

time. Until then, happy analyzing.