Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Kirill Eremenko: This is episode number 289 with top AI influencer,
Ben Taylor.
Kirill Eremenko: Welcome to the SuperDataScience podcast. My name
is Kirill Eremenko, Data Science Coach and Lifestyle
Entrepreneur. 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.
Kirill Eremenko: This episode is brought to you by DataScienceGO,
which is our very own data science conference.
DataScienceGo is the event where, once a year, we
bring the data science community together, and we
also bring very empowering, impactful speakers.
Check this out, this year, we are bringing you
speakers from IBM, Google, Salesforce, Amazon,
Atlassian, RStudio, Amazon Alexa, Facebook and
more. We actually have 30 plus speakers already
confirmed and coming this year and ranging from all
different roles and backgrounds from analysts to
senior data scientists, from engineers to founders and
directors.
Kirill Eremenko: The beauty of all of this is that you get to interact with
them, you get to see them live, you get to hear them
talk and then come up to them and ask them
questions and connect with them, meet each other.
For example, last year we had people from over 23
countries fly to DataScienceGO, just to give you a bit
of perspective. This year, DataScienceGO is happening
on the weekend of 27th, 28th and 29th of September.
We're expecting 600 to 800 attendees, so there'll be
plenty of networking opportunities. Ticket prices are
going up at the end of Monday, the 26th of August, so
if you haven't secured your ticket yet, head on over to
www.datasciencego.com right now and secure your
ticket ASAP. That's datasciencego.com, and I'll see you
there.
Kirill Eremenko: Welcome back to the SuperDataScience podcast, ladies
and gentlemen. Super, super excited to have you on
the show today because for the third time round, I
have my dear friend, AI influencer and mentor, Ben
Taylor, on the show. It was super cool to catch up with
Ben. In fact, it was funny that we did this episode as a
video and the previous episode we did as a video as
well, which was two years ago.
Kirill Eremenko: We had a look at the previous episode in video, and we
could see how, in two years, we've gotten older. We
had a bit of a laugh about that what AI or what
running a tech startup actually does to you and how it
ages you, and it was quite insane. If you want to find
the video episodes, if you would like to watch that, you
can find it at superdatascience.com/289. That's
www.superdatascience.com/289. You can find it there.
We actually will add a comparison of before and after,
how we looked. In this episode, you will find tons of
value. I'm so excited for you to hear it. Here's a couple
of things that you'll find inside.
Kirill Eremenko: You will find out what Ben has been up to in the past
two years since you've heard from him last unless, of
course, you've seen him in one of his appearances in
international keynotes. Then you will also find some
very cool concepts about artificial intelligence such as
active adverse impact mitigation and what that means
and how that can help train on your dataset without
bias. Then, we talked about AI ethics. We talked a lot
about deepfakes. We talked about Ben's current side
project, passion project. He's building an artificial
intelligence that plays Call of Duty, and he will
actually demonstrate this at DataScienceGO this year
at the end of September.
Kirill Eremenko: In this podcast, he gave us a preview of how he's doing
it. It's such a crazy project that he's working on. Very
excited to hear that. Next, we talked about residual
technology. We also talked about AI startups and how
investors think about them and many, many more
topics. This podcast is jam packed with value. Without
any further ado, I bring to you my dear friend Ben
Taylor.
Kirill Eremenko: Welcome to the SuperDataScience podcast, ladies and
gentlemen, super excited to have you on the show.
Today, we've got a dear friend of mine, a super special
guest, returning for the third time round, Ben Taylor.
Ben, welcome.
Ben Taylor: Hey. Thanks for having me again. We were talking
about that we're both old men now. Looking back
three years ago, we looked like little kids and now
we've got some gray coming in and ...
Kirill Eremenko: Yeah. Yeah, man. I'm glad we're doing this ...
Ben Taylor: There's lines in our faces.
Kirill Eremenko: Yeah. I'm glad we're doing this as a video because, the
last one we did as a video was ... Oh, our previous
podcast was like two years ago, and we just looked at
it. I'll ask our video editors to, right now, [inaudible
00:05:27] put before [crosstalk 00:05:29]-
Ben Taylor: I'll do a ... Look at all this. Every white hair is a
mistake, and I have been able to collect a lot of them
over the years.
Kirill Eremenko: Yeah, that's crazy. It's insane. What have you been up
to that you have so many mistakes in your face?
Ben Taylor: One of the interesting things doing a startup ... My co-
founder, David, says a startup is a series of mistakes,
but hopefully in the right direction. That's very true. I
think sometimes you can beat yourself up about
20/20 hindsight, "We could've done this better, we
could have done this contract this way. We could have
asked for this pricing, we could have not done this or
that." That can be really discouraging.
Ben Taylor: The important thing is you learn from your mistakes
quickly, and you try not to repeat them, and you try to
find themes or patterns or strengths or weaknesses
that give you momentum and help you grow. We've
learned a lot in the last three years, talking to
enterprise for AI, because there's so much hype in AI.
Everyone wants it. It's actually not very hard to get an
executive meeting about AI, but the problem is, it's all
blue sky. It's not very actionable, and they don't really
understand what it is or how it would be useful. You
having an AI background, you feel like you're just
grasping at straws and there's no ... Ideally, you're
looking for pain.
Ben Taylor: We're looking for this specific problem, put a fence
around it and it's worth tens of millions of dollars.
That's what we're looking for. Then everyone's aligned
that, "Okay, solve this problem within this timeframe,
and this is a great relationship for both of us," but if
it's not nailed down to some business objective, then a
lot of times it can be a waste of time. We've learned a
lot of important lessons, talking with a lot of different
companies, a lot of different industries. We've really
focused lately in insurance and in assessments. Those
are our sweet spot verticals right now.
Kirill Eremenko: Insurance and assessments right now?
Ben Taylor: Yeah, video and audio assessments. This would be
something like ... I worked at HireVue, and we did
video assessments for pre hire. This could be
something like assessing English as a second language
or remote proctoring or predicting some type of
behavior or competency. Remote proctoring, there's a
lot of people that need it. This is where you're taking
an exam, and you decide to cheat during the exam.
Ben Taylor: Right now, the only way to catch you is to have
humans watch. They have to watch all the video and
these exams can be very long. With AI, there's an
opportunity to catch these events, which can save
time. Then in insurance, we do loss prediction. We're
predicting loss on a property. Should you insure this
property, yes or no, based on the structured data and
the unstructured data, so images and text. We build
these holistic models to predict that, and that's a fun
problem because then the numbers are big.
Kirill Eremenko: Can imagine.
Ben Taylor: Yeah. You move the needle this much, and it's worth a
really big number. Those are the types of problems you
want to work on.
Kirill Eremenko: That's crazy. It feels like when you're done, the
completion rates at universities ... Because cheating
will no longer be available ... will like drop by 50%?
Ben Taylor: Hopefully. Yeah, we're just trying to make everyone
more honest, I guess, AI. Some people don't like it.
Especially coming from the HireVue side, using AI to
do pre-hire assessments, there have been some very
negative reactions from that in social media where
people feel like it's Black Mirror.
Kirill Eremenko: Yeah.
Ben Taylor: "It's really happening."
Kirill Eremenko: It is, it is. I have a friend who went through that
recently, and she's like, "I was preparing. I was there. I
was going to start talking. Then I log in and there's
nobody. There's nobody to talk to. Then they give you
these questions, you have to answer them and it's all
recorded and then analyzed by AI," but she didn't
know. She was like, "I didn't know what was
happening. I think they recorded it because maybe
they'd look at it later." I'm like "No, no. I know. Ben
told me all about it. Nobody ... This is going to be an AI
assessing the whole thing, and then this is how it's
going to work." She was like, "Wow," not expecting that
at all.
Ben Taylor: Yeah. I'm biased because I come from that side of
things where I see a lot of the good it can do, where it
can help eliminate bias, because you have racism,
sexism and age bias. You can get ahead of it, so you
can actually protect against it and really try to train
models based on competencies needed for the job.
Kirill Eremenko: But that's only if your data doesn't have bias, right?
Ben Taylor: For most people. What you said for most people is
true. If you take a racist training set, and you train a
resume model, you're going to get a racist model. I
think the point I want to make really clear to your
listeners is you are guaranteed to transfer bias with
traditional machine learning. If you're using bag of
words or some type of fancy neural net to build a
model for video or audio or text-
Kirill Eremenko: Supervised learning, basically.
Ben Taylor: Yeah, supervised learning, you will transfer the bias
right across. There's a special type of supervised
learning where you do active adverse impact
mitigation. What you're doing is you are rewarding
features that predict performance, and you are
poisoning or killing features that predict race or
predict age.
Ben Taylor: It's actually not a complicated topic. The easiest way I
can explain it is imagine a resume. If I just throw
resumes into a machine, and if the last name Garcia is
seen as having any type of lift, it would also have lift
with predicting race because a last name like that can
be very racial. That would automatically be thrown
out. You and I might come up with that idea and say,
"Oh yeah, don't look at name because it can be tied to
race," but AI can automatically figure out that if I'm
trying to predict Black or Hispanic or White or Asian,
this name is interesting. That feature becomes
eliminated automatically.
Kirill Eremenko: Automatically, so you don't have to tell it which
features to ...
Ben Taylor: Yeah.
Kirill Eremenko: Interesting.
Ben Taylor: That's kind of the process. You're not just building one
model. You might building five models simultaneously,
and they're all competing for features. What we find is
if you take that approach, you can actually train a
racist model and ship a model that is within the
guidelines. It still was able to get lift on a performance,
but the bias transfer was greatly reduced. There's
ways to go about doing it.
Kirill Eremenko: Wow. Very interesting. You just need to indicate what
things the model is not allowed to discriminate on like
age, gender, race?
Ben Taylor: Yeah.
Kirill Eremenko: You just need to specify those.
Ben Taylor: Yes. One of the things I tell people is if you can predict
it, you can protect it. If you can't measure it, then how
are you going to protect against it? If there's a genetic
bias or if there's something else going on, how are you
going to protect against it? We've actually found some
really fascinating things that aren't protected right
now. Beauty, there's a really strong beauty bias, so if
you're a woman or a man and if you're more attractive,
you will do better in the hiring process.
Ben Taylor: In HireVue, by the way, they don't use that. That's not
used to give someone an advantage. They've actually
done some internal studies, and they've shown that
there's no correlation, internally, to their AI metrics
around something like attractiveness, but on the
human side, there is. There's some very, very strong
correlation. Some of the strongest correlations we've
seen are tied to attractiveness. It's interesting, but it's
also kind of sad.
Kirill Eremenko: Then how would you measure attractiveness to specify
to the AI that it needs to, as you said, poison those
features that predict it?
Ben Taylor: We already have an attractiveness model. Speaking
about our company, we have one. Attractiveness is a
really fun topic because you hear people say, "Beauty's
in the eye of the beholder." Now, when we're talking
from the AI's perspective, beauty is in the eye of the
training set. If you were trying to put me into a corner
on a hot seat by saying, "Ben, can you tell me what the
AI thinks ... what I would think is attractive,"
technically, the answer is no.
Ben Taylor: I can't tell you that because I would have to build a
training set based on you, but when it comes to LA,
Chicago, East Coast, West Coast, South Korea, Brazil,
yes, AI can tell you because that's trained on lots of
humans that have done ratings in those areas. When
it comes to predicting regional average behavior from
humans, then, yes, AI can predict.
Ben Taylor: But beauty's ... I can't remember if we talked about
this on a previous podcast, but this was an evolution.
When we trained the first beauty model, we found out
that it was rewarding sexualized beauty. It would
actually reward women who were lingerie models or
they were dressed ... It wasn't focusing on the face. It
was a whole-body shot. When we noticed that, we
thought, "Oh, that's not really what we intended."
Then we did face crop.
Ben Taylor: The second time around, the number one ... We have a
million images that we're testing on. These images are
celebrities. They've never been seen before. They're not
part of the training set. This is our sanity check that
we're ranking this dataset to see how well we're doing.
In the second version, the number one pick won Miss
World. It's a million photos, 13,000 unique
individuals, the number one pick won Miss World,
which is like, "Oh, well, that's not random."
Kirill Eremenko: AI picked the lady that won Miss World?
Ben Taylor: Yeah, so like what are the chances that I stick my
hand into a bowl and out of 13,000 people, I pull out a
Miss World contestants? It's not one in 13,000, but it's
probably like five in ... it's pretty good. We noticed in
our top 10, the racial differences were a little ... they
seem to be oversampling on certain minorities. What
we saw when we looked at the racial distributions is
they were very different.
Ben Taylor: You had some races that were skewed high, some
races that were skewed low. The current version that
we have is, we do race norming, where the beauty
score we're predicting doesn't have any racial
differences. Some people disagree with that, but I'm
not going to allow a whole race of people to be scored
low just because someone might argue with me on why
that should be okay. I'm not going to allow that to
happen. It is a controversial model, which was kind of
fun.
Kirill Eremenko: Got you. Yeah, interesting. That's the whole AI and
ethics space, and it's really cool to see that people like
you are really taking that into account and developing
the models that you're creating.
Ben Taylor: Yeah. I've had some fun conversations with some of
the AI ethics journalists. I see myself as a technophile,
where I love inventing things. Hopefully, I don't invent
anything for bad, and maybe this'll come into the
discussion with deepfakes and some of the stuff we're
going to talk about later in the podcast. There is a
chance that I might make it easier for certain people to
do things just based on having a discussion or
bringing something up or suggesting something that
would make it more difficult to catch deepfakes.
Ben Taylor: Whether or not I create that technology or someone
else does, having the conversation, it's kind of a
moving goalpost because the more you talk about
ways to protect against deepfakes, the better they get.
That's true with any type of AI that you're using to
catch the bad guys. The more you talk about it, and
the more the researchers look into it, the better it gets.
Whether it's-
Kirill Eremenko: Because they all have access to the same data anyway,
the same algorithms.
Ben Taylor: Yeah. The same algorithms. They understand the
incentives. If I'm trying to record your podcasts and
fake your voice for some bank authentication through
your voice, 10 years ago that would have been science
fiction. Today that's becoming easier and easier and
easier to pull something like that off.
Kirill Eremenko: That's crazy. Speaking of deepfakes, tell us a bit about
that. Ultimately, if you think about it, this could be
deepfakes talking to each other. We could not be here
at all.
Ben Taylor: Yeah. I'm actually surfing in Costa Rica and I have
outsourced this podcast to someone in the Philippines
and they're doing a live deepfakes with you right now
in live, so it's very impressive. They're just reading
from a script of ...
Kirill Eremenko: Yeah, imagine the listeners who are just ... they're
listening to the audio and not the video version. It
could be even just two AIs generating natural language
on the fly like having a chat to each other.
Ben Taylor: Yeah. What's the most likely thing that I should
respond to what you just said? Or when does the
laugh track make sense?
Kirill Eremenko: Yeah.
Ben Taylor: I can't remember if I ... There's a lot of buzz right now
around deepfakes where people, they want regulation.
They want us to figure out how to detect them and
stop them from happening. I feel like that's the wrong
conversation to be having. We actually just have to get
to the end of the story, and the end of the story is
there's no way to detect a deepfake.
Ben Taylor: Today, there is. I feel like if there was a very high
profile case, where there were huge consequences for
determining if this deepfake was real or not, there's
some pretty detailed ways that people like us could ...
You or I could figure it out very quickly that this is
fake, but ... I'll bring up some specific things you could
do, but those things are eroding away. Where five
years from now, 10 years from now, I would argue it
would be extremely hard for an AI expert to convince
another AI expert that this is a deepfake or it's not.
Ben Taylor: Today there's an argument to be made, but in the
future there won't be, so let's just finish the story and
figure out what we're going to do when we can't. We
don't know it's real. People talk about going back to
blockchain where you authenticate the source. If I
send a video of you doing something inappropriate, if
you don't trust the source, you don't trust the video,
it's not newsworthy. It should not be shared, but if I'm
a news reporter, you know me, and I'm authenticated
to you. You're able to confirm that, that there wasn't
some type of intercept, that this is me giving you a
video, then that's what we have to go to is source
authentication.
Kirill Eremenko: Interesting, so blockchain could play a big role in that?
Ben Taylor: It could play a big role, and I think it'll change the way
we do media. Right now, for local media, they'll ... I
think I mentioned this too, I personally actually have a
national media mention of me coming from a fake
social profile. It wasn't Ben Taylor.
Kirill Eremenko: Yeah, I saw that, yeah.
Ben Taylor: It was my fake social profile mentioned on USA Today
during the Ashley Madison crap, and they were pissed
when they found out. Right now, there's not a lot of
due diligence on sources because they're just trying to
grab whatever's out there in social media, interesting
videos and different things. That's going to go away
where everyone has to be authenticated, and you have
to know the source, and you have to have
confirmation.
Kirill Eremenko: Wow, important, important. The fastest way I can tell a
deepfake for now ... They are getting better ... is you
look at the teeth. Usually have a third tooth in the
middle that's [crosstalk 00:21:40] or the earrings. Like
if you see a ... A really cool website to test these things
on is thispersondoesnotexist.com. You just refresh it,
and it's a new image every time. Earrings usually,
things that are supposed to be symmetrical sometimes
aren't. Then you can like [crosstalk 00:21:56].
Ben Taylor: We see as the generations of this technology gets
better, those issues start going away. Symmetry and
different things are being improved. I was actually
taking a nap a month or two ago and I woke up from
the nap. When I woke up I thought, "Holy cow, I know
how to catch a deepfake. I know how to catch the
world's best deepfake." When I say the world's best
deepfake, I mean let's say someone in Israel or Russia
or the US spent $10 million to create one deepfake.
You and I are staring at it and we're watching it over
and over and over again and it's high resolution.
Ben Taylor: We're staring at it, and as humans we can't see
anything wrong. You can't see any artifacts, and we
might actually get to a point where we have to give in
and say, "We don't see anything wrong with this, we
think it's real," but the funny thing is mentioning this
... As soon as you mention something, it's no longer a
thing, but I'm fine mentioning it because someone else
will mention it. The fascinating thing with a deepfake
is it doesn't have a pulse. There's no heart rate.
Kirill Eremenko: Oh, okay. Wow.
Ben Taylor: If they're fair skin, you can amplify the heart rates in
the temporal data in the video and you can see their
heart rate in their face, and for a big effort, I would
argue that would have been a detail they may have
missed, where I have a deepfake of you right now, it's
incredible, it looks real. There's no visual artifacts, but
they forgot to give you a heart rate.
Kirill Eremenko: That's crazy. I love that. I was looking at ... There's an
app now that you can point it at a video and it will
emphasize any kind of heartbeat like if it's for baby
monitors. I want to see that the baby's breathing, so it
expands [crosstalk 00:23:51].
Ben Taylor: Exactly, it's that technology. To make the AI
community feel better, I'm pretty slammed right now
just with startup and work, but I love fun marketing
pieces. To make the AI community feel better, I wanted
to show the first deepfake with a heartbeat. I'm too
busy, and I've got other things in the queue, but I
would love to say, "Hey everyone, this is a problem,
and I fixed it for you, and now there's a deepfake with
a heart beat." Then the next step would be you would
actually become very opinionated on this specific
heartbeat signature. Is this Ben Taylor's heartbeat?
Kirill Eremenko: No.
Ben Taylor: Or is this a modified ... You would actually have to go
to that level to ... What you see is the argument starts
disappearing. You and I talk heartbeat, and we say,
"That's a great thing."
Kirill Eremenko: Now, all the dark web is already onto the heartbeat
thing.
Ben Taylor: Yeah, now, you see it's coming.
Kirill Eremenko: [crosstalk 00:24:43]. It's like you said, it's like a race.
Before, they could have used it in a major case where
[crosstalk 00:24:52], but now that it's out there and
you said it on a podcast, now, it's gone. You can't use
that as a [inaudible 00:25:00].
Ben Taylor: I'd love for some people to comment and get mad and
say, "I wish you had just told each other that privately
and not publicly because now someone on the dark
web can have those ambitions," but part of me just
wants to get to the end of the story that you don't trust
anything.
Kirill Eremenko: Yeah.
Ben Taylor: This feels like a waste of time, if you're trying to fight
an intermediate step. Catching deepfakes in 2019,
2020, 2025, it's a lost battle, so why did we spend so
much time when we could have just solved the bigger
problem?
Kirill Eremenko: Yeah. Yeah.
Ben Taylor: Yeah. It's fun stuff.
Kirill Eremenko: Interesting. Speaking of other things that you're busy
on, you're coming to DataScienceGO this year at the
end of September, 27, 28, 29. Your presentation
sounds like a lot of fun. Tell us about that, that
[crosstalk 00:25:57].
Ben Taylor: You have different passion projects, and sometimes
they're spur of the moment where they're literally that
morning or a few weeks you'll think of something. For
an AI company, that can be good because they show
thought leadership and you can kind of drive some
stuff in the AI community, but there's been a very
selfish passion project of mine that I've obsessed about
for years, and I did not think it was possible. That was
playing Call of Duty on the Xbox with AI in a live
environment. I don't mean a modified Xbox or locally. I
just want to be playing full auto live on the web
against people where I have not asked their
permission. [inaudible 00:26:40].
Kirill Eremenko: Rebel. Rebel man.
Ben Taylor: Yeah, rebel man. You know I like to [crosstalk
00:26:46].
Kirill Eremenko: You like to push it. You like to-
Ben Taylor: Yeah, if there's a ripple or a splash, I'd rather go for
the splash because that's more entertaining. Two or
three years ago, I thought, "Man, I really want to see
this happen, but ... This is a very naive number. I'll
just throw this number out. I think Google DeepMind,
they've spent over a million dollars in R&D on ... No,
more than that. They've spent millions of dollars for
specific games. They'll decide, "We're going to go tackle
this game." They'll spend millions of dollars. I was
thinking for this it'd be maybe $5 million in talent and
in time to figure this out because you have the Xbox-
Kirill Eremenko: For quality?
Ben Taylor: Yeah, for quality because you have the Xbox drivers.
It's not meant to play nice to do that. You've got
network protocols through USB that you have to
override and intercept and take control of. It's just a
whole skill set that we don't have. AI researchers don't
have that. For less than $2,000 hardware I was able to
cobble together, I was able to figure out a way.
Kirill Eremenko: Wow.
Ben Taylor: It's a really fun set up. I had to buy a piece of
hardware from France called a GIMX adapter. It does
this man in the middle attack where it tricks the Xbox
into thinking that you are an Xbox controller. It does it
by intercepting a real Xbox controller. I have a real
Xbox control and when I push the button, that goes
into a Linux AI computer and that goes to the Xbox
and it kind of does the handshake. Then once it does
the handshake, then the Linux computer takes over.
Kirill Eremenko: Oh, wow.
Ben Taylor: For the Xbox, it doesn't know. It has no idea. It's just,
"I'm getting these controls from this controller." Then
for the video, we have the video coming out of the
Xbox, and it goes into an HDMI capture card on the AI
computer. The AI computer sees full 1080p. We were
running at 60 frames per second for a while, but it was
a little hard for the capture card. Still 30 frames per
second, that's faster than most humans can react,
especially with the latency. The AI computer sees
everything, and it has access.
Kirill Eremenko: Wow.
Ben Taylor: It's been a really fun project. The thing we started with
is you always want to train from a good baseline
because you want to train. You want to get all this
training data from gameplay so you have stuff to work
with to study. I'm not good enough to be the human,
so I put some social media feelers out for good humans
to come and play on this special modified system. You
would be amazed how many mothers I had on
Facebook-
Kirill Eremenko: No?
Ben Taylor: bragging about their son's kill streaks. I had mothers
saying, "My son has killed 24 people in a row," and
another mother's like, "My sons killed 35 people in a
row." It was so funny because you can imagine being a
young, like 12-year-old kid and your mom's calling
from the other room, "How many people have you
killed in a row on Call of Duty?" You're like, "Oh, she
actually cares about what I do all day." I'm sure that's
the first time that's like, "Oh, this is ...
Ben Taylor: I had a bunch of people coming over to my house and
trying out and I even had some really young kids ... I
think the youngest was 12, which is really funny cause
they're coming over with their really proud parents,
and they're playing on the system. I found this
professional gamer. I think his name is Caden. He was
next level. I'm still shaking my head just watching him
play. He created 3,000 kill events. Kill events, it's not
people killed. It's like fight sequences that were saved
and pushed to the cloud. He's killing people so ...
Ben Taylor: The funny thing is he shows up to the house. We have
this special system set up. He needs to use his
monitor. He's not willing to use our monitor. He has a
special monitor that sits right in front of his face. We
reset up, so it's using his monitor. Then he asks if it's
okay if he uses Gamer Goo, and the answer is yes.
Whatever he asks, the answer is yes.
Kirill Eremenko: What's Gamer Goo?
Ben Taylor: Exactly. I don't know. He pulls out ... It looks like
lotion, and it says Gamer Goo. He squirts it on his
hands, and it makes his hands grippier or something
like that.
Kirill Eremenko: Wow.
Ben Taylor: Yeah. He's in my house playing for four hours, and
he's got all these data scientists and physicists and AI
people behind him commenting. They're not gamers.
They're commenting on like [crosstalk 00:31:25].
Kirill Eremenko: What are you doing Ben? What is your life, man?
Ben Taylor: We recorded some of this. Actually, I should send you
... I'll send you a video element. Maybe you could even,
like [crosstalk 00:31:35].
Kirill Eremenko: Yeah, send it. We'll put it in.
Ben Taylor: It shows him playing with all of this crazy stuff flying
through the computers. He's created an incredible
dataset to study and learn from. The thing you start to
realize pretty quickly is humans, they can't win. They
really can't win because there's ... Maybe a real world
example, let's say there's a gun fight in the future.
You're in a bar, droids come in and they start shooting
up the place with machine guns, and you do too
because it's Terminator days. You've got your machine
gun, everyone's shooting.
Ben Taylor: If I run in and I yell, "Stop," and everyone stops, and I
ask you to count your bullets, you have no idea how
many bullets you have. You have no idea. You
honestly have no idea. Maybe you think you're almost
out of your clip. You don't know. If I ask AI, AI knows
exactly how many bullets it has, but it also knows how
many bullets you have and how many bullets your
partner has because it's counting. It counts
everything.
Ben Taylor: When it comes to accounting, it's the world's best
accountant in that type of scenario. For a very specific
example, every muzzle flash where the bullet leaves
the gun in the game, the AI is capturing all of that in
real time with perfect accuracy. It's counting bullets
and it's counting its health. It just has a much better
... It has a faster time to react. The amount of data it
can consume is just unbelievable. It creates over a
terabyte of data every hour. You don't think about it
because you're just playing the Xbox.
Kirill Eremenko: Wow, that's insane. What were the results? Did you
manage to train the AI to play like Caden?
Ben Taylor: With these models, they ... It's a huge project because
... I can't remember. Google said they had 18 agents
working together to play their StarCraft, and that's
kind of how you think of it. You don't train one AI. You
train all of these submodels. We have just a bunch of
submodels that have been trained where they hit really
high accuracies, and then they all work together on
one decoder and encoder.
Ben Taylor: We're still working through it. We have a lot of things
that are really exciting where AI is essentially pulling
the trigger. AI wants to shoot you, but there's different
things. There's the gun movement. There's the actual
physical movement. The thing I'm pushing for is by
DataScienceGO, we do have clips of, "Hey look, that
person died, that person died, that person died, and
it's pretty good." You'll see already from these
submodels, their accuracy is unbelievable.
Kirill Eremenko: That's crazy.
Ben Taylor: That's a real passion project of mine. I do like to troll
Microsoft, so I would love to have a Twitch feed with a
life AI bot running and Microsoft just has to watch in
horror. Then you're masking all the gamertags, where
they ... They'll try to blacklist you. If they can see your
gamertag or they know who you are, they're going to
try to kick you off work, but if they can just watch on
the lag, then they're just helpless.
Ben Taylor: Part of me ... No one else is willing to do a shooting
game right now, and part of me actually wants to do it
to raise awareness. I want to just start a discussion
that, "Look, this is actually pretty good. This is maybe
a glimpse into autonomous warfare, and what do you
guys think? What does society think?" If we're having
these discussions 10 years from now, I think it's too
late. Then you're having it based on a real-world
demonstration. Having an unreal world demonstration,
I'd say it's too late.
Kirill Eremenko: Yeah, I agree. I agree. One thing I don't understand in
this scenario is that usually ... For instance, Google,
when they try to model a game or create an AI that
plays the game, they use reinforcement learning. They
don't have of this supervised playing.
Ben Taylor: Oh yeah. Yeah.
Kirill Eremenko: How come you guys needed the supervised data sets?
Ben Taylor: The supervised data sets, it's really, really good to ...
First, you have to study to figure out what the
elements are that you need for gameplay. Me telling
you that muzzle flash and hit indicators were useful, I
wouldn't know that unless we had hours and hours
and hours of footage to review based on gameplay.
Ben Taylor: For the reinforced learning, the thing ... You can
initialize on human gameplay, but the very next thing
you want to do is you want to go to superhuman. You
want to go to some cost metric. The thing that we're
still hammering down is what is that cost? I think the
cost is going to end up being the amount of lives you
take per unit of your health. The good news is these
are very short fight sequences.
Ben Taylor: We're not talking about strategy minutes away. We're
talking about, "In the next five seconds, are you going
to kill someone? If you do, how much health did you
have to give up to kill someone?" That'd be the
reinforced part where then the AI is just rewarded, just
plays, plays, plays. Every single fight sequence is
essentially scored as part of the training set on, "You
fought and you were killed. That's very, very bad. You
should never be killed," or, "You fought and you killed
a few people, but you were hurt very badly doing it."
That just goes back into the training set. Those
sequences and those outcomes become pretty
objective.
Kirill Eremenko: So basically-
Ben Taylor: [crosstalk 00:37:16].
Kirill Eremenko: Yeah.
Ben Taylor: Yeah. The nice thing with something like a first-person
shooter is the objective is even simpler. Something like
StarCraft or these other games are much more
complicated because you have long-term strategies
that are very, very complicated.
Kirill Eremenko: And so many different pathways that can evolve.
Ben Taylor: Yeah, but just AI coming around the corner and
there's three enemies, you have to kill all three, that's
not as complicated.
Kirill Eremenko: Yeah, yeah. We did a simple one for Doom. Remember
that game, Doom?
Ben Taylor: Yeah, yeah. Doom would train on random
initialization. It just starts, you have a navigation, and
initially it shooting at the sky, shooting [crosstalk
00:37:57]. Then eventually it kills the monster. The
problem with live gameplay on the Xbox as if you try to
start with random initialization, the gameplay's too
complicated. You're not going to kill someone, but if
you start with initialization trained on a professional
player, the likelihood of you shooting someone and
killing them is high.
Kirill Eremenko: Okay.
Ben Taylor: If someone walks in front of your gun, you're going to
shoot him. Guaranteed, you're going to shoot him.
That's come from the human gameplay.
Kirill Eremenko: Got you, so this is like a combination of supervised at
the start to get you up to speed and then
reinforcement learning to take you superhuman?
Ben Taylor: Yeah. A lot people don't know this about me, but I
really geek out about high-performance computing.
The thing I'm the most excited about this is just the
high-performance computing element. The number of
models that have to run at 30 frames per second and
keep up is very impressive. That's something that I'm
excited to show off at DataScienceGO.
Ben Taylor: In my career, my favorite thing to do when it comes to
AI research is I love to show someone something that
is so unbelievable, it's not believable, especially like
where I'm accused of lying. If it's, "Hey, these are my
benchmarks. This is how many models I'm running at
30 frames per second on this hardware," I love to have
numbers where there's someone in the audience that
says, "No, I don't believe that." For me, that's kind of
icing on the cake because then I can meet that person,
say, "No." I don't have to meet him. I just like that. I
like when people don't believe me.
Kirill Eremenko: Yeah. Wow, that's really cool. It's interesting that you
set yourself that challenge, that DataScienceGO, have
some of these things to show. That's really gonna push
you to get there.
Ben Taylor: I've been showing this stuff off for a while. I presented
this at Amazon's Palo Alto office, and I presented it in
Minneapolis. We've been working on this for a while.
We made really good progress, but we're actually
getting to the live gameplay elements that get really
exciting because there's actually some historical ...
People might not think they're that historical. I think
they're pretty historical. The other wonderful thing
about this is this is all recorded in full 1080p.
Everything is recorded all the time.
Ben Taylor: The very first autonomous kill on an Xbox against
someone online without their permission will be
recorded. I will know their gamertag. The world may
not know their gamertag. I will know their gamertag.
That video can go on YouTube and just be shared to
the world that, "Hey, this person was the very first
person killed online with autonomous AI, and they had
no idea. They're just coming around the corner and AI
saw them and activated and shot them.
Kirill Eremenko: Yeah. Well, if anybody watching this sees Ben Taylor
in their game [inaudible 00:41:01].
Ben Taylor: Yeah. I would mention my gamertag, but that's
[inaudible 00:41:09].
Kirill Eremenko: No, then Microsoft will take it down. Don't mention it.
Don't ruin the exercise. Wow, that's very cool. Very
cool. Ben, our listeners might be getting a bit of a false
perception of you that you are just like this gamer,
crazy gamer who creates AI to dominate the online
world of shooters. Your company, Zeff or Zeff, right?
Ben Taylor: Mm-hmm (affirmative).
Kirill Eremenko: You are consulting, is that correct? I think that we
mostly-
Ben Taylor: We have a platform. We have an AutoML platform. We
specialize in image, audio, video and text models, but
very specific kinds. The types of models we build that
we do well are called holistic models, where it's
structured data interacting with not just one image
type, but multiple image types. Imagine predicting
loss, and I've got images of roof, dwelling, satellite,
Google Street view, structure, maybe text descriptions.
Those types of models, the industry is starting to catch
up, where they're starting to think that way. A couple
of years ago, we felt like we were the only ones
thinking about that way.
Ben Taylor: We're seeing some open-source projects like Ludwig
from Uber, where they are starting to think about
encoders and decoders. "How do I take a hybrid or
mixed dataset and build these types of models?" That's
our specialty. We allow engineers to build those
models. The Xbox model is actually showcasing how
complicated some of these AMLs can get in real life.
This particular model is going to have north of 10
submodels or encoders working together to drive a
final outcome, and we see that in industry too.
Whether it's insurance, house-price assessment. It
does showcase some of our capabilities.
Ben Taylor: The other thing I want to throw out there ... Maybe
just go with me for a second on a scenario. Let's say
you're an investor or you're a VC. I'm going to pitch
you right now. Got a good startup idea, and I say, "I
need $25 million, and I'm going to go hire a team of
PhD physicists, data scientists. We're going to work for
the next two years doing Xbox gaming with a AI, and
we're never gonna make any revenue. We'll never make
a dollar of revenue, but in the next two years, we will
sell for 50 to a $100 million." That scenario to
someone who when they see that for the first time,
that sounds ridiculous.
Ben Taylor: It sounds like, "Why would that ever work? Why would
that ever produce any value that'd be worth buying?"
The thing we're noticing when you tackle a project like
the Xbox, you actually get residual tech. You just get
tech that comes along for the ride. You just get things
that are invented. You didn't know you needed them
and suddenly you have them. Some of the models in
the throughput we have for these higher-resolution
video feeds are kind of groundbreaking, but if we
didn't have the passion piece, we wouldn't have
discovered them. It's kind of crack cocaine for nerds.
Ben Taylor: If you tell them you're going to do autonomous war on
Call of Duty, how many nerds can you get rushing to
your side if you can pay them market comp to work on
that problem? In the end, you have to pay the bills.
You have to make money. You can't just ... I don't have
the swagger in the VC community to swing that stick
yet in my career, where I could say, "Hey, I need $20
million to goof off for three years."
Kirill Eremenko: Yeah, yeah. Got you. This AutoML thing is a way for
you to supplement your research?
Ben Taylor: Yeah, yeah.
Kirill Eremenko: Passion projects.
Ben Taylor: Yep.
Kirill Eremenko: What kind of AutoML do you offer? Any company can
come sign up and start using the platform?
Ben Taylor: We are specializing in insurance. If there are insurance
companies that want to predict loss on a property ...
Like a residential home, they're trying to decide,
"Should I insure your home right now?" They have a
human underwriters that will go through that process.
With our capabilities, we allow them to unlock the
potential of the unstructured data because it's very
awkward and clumsy for these companies to try to do
that internally. They really struggle with it.
Ben Taylor: We make that very easy. Their engineering team can
build their own models on our platform. They don't
need to know data science or AI or neural networks.
We actually have an adjacent schema where they can
submit property records through our system, and then
we take care of the model building and the automation
for them. Those are the types of customers that we
would be going after, and the good thing with
insurance is there's a lot of them. There's a lot of
insurance companies that care very much about loss
prediction or price prediction.
Kirill Eremenko: Awesome. Makes Sense. What's your exit strategy for
this business?
Ben Taylor: We've had acquisition options in the past, so there's
always that scenario I guess, assuming that the
market keeps up that appetite, but with some of these
insurance contracts, there's also an opportunity to
just grow the business and become cashflow positive
and self sustaining too. We are not VC backed right
now, so we don't require an exit strategy today, but
you never know what's going to happen in the next 12
months or six months.
Kirill Eremenko: Or Microsoft might come along and buy you so you
stop destroying their Call of Duty product.
Ben Taylor: Yeah, yeah. It could be like a ransom. If you don't buy
us for $15 million, we will play for another 24 hours
and kill a thousand people online.
Kirill Eremenko: Yeah. Got you, got you. Ben, I wanted to ask you
another thing. One of the best places ... You present at
many different conferences and people can meet you in
many events, but one event is DataScienceGO. For
those who are listening to this and are still on the
fence about coming to DataScienceGO this September,
what would you say to them? You've been to two
DataScienceGOs now. I love you for this. You're such a
great supporter. You always come and do amazing
speech, everybody loves you. What has your
experience been so far from 2017 to 2018, and what
are you looking forward to in 2019?
Ben Taylor: I speak at a lot of conferences all over. I spoke at
Dublin Tech Summit. I'm actually speaking in Madrid
right before DataScienceGO, and we had to figure out
the flight pattern where it works. It's going to work
out. One of the things I really like about
DataScienceGO is it's a really tight-knit group of AI,
data science professionals and people trying to break
into that space. Out of all the conferences I've
presented to, I've never presented at a conference that
has the energy and the excitement and the nurturing
that comes with the attendees that I see at this
conference.
Ben Taylor: Because a lot of other conferences, the audience is not
that engaged, honestly. If you had to like measure
excitement from the crowd, there really isn't any there.
They're just kind of there, and DataScienceGO is
completely different. Last year, there's people cheering.
You guys do a great job with the DJs and stuff, but
people are cheering. Actually, I think if you listen to
my talk, there's people whooping and cheering during
the talk.
Kirill Eremenko: Yeah. When you took that Selfie, everybody was like,
"Yeah, yeah, that's great."
Ben Taylor: Yeah, yeah, people were doing that, but even little like
whoops and stuff in the background. Just you say a
statement that people see as truth or they agree with
it, and your confirmations are whoops from the
audience. I've never been to an AI conference that does
that. It's a lot of fun for me.
Kirill Eremenko: That's very cool. Thank you for the words. Did you
meet any interesting people last time?
Ben Taylor: Yeah. I always have fun interacting with a lot of the
other speakers like that. That's fun for me. I met some
people there from Red Bull and SpaceX I was able to
follow up with and go onsite to their locations.
Kirill Eremenko: Wow.
Ben Taylor: I've kept in touch with a lot of people that aren't local,
even international folks. I've really enjoyed staying in
touch with them. I've always enjoyed the contacts that
I see there and meet there.
Kirill Eremenko: That's good, Ben.
Ben Taylor: I've had people already message me that are attendees
that are coming back, and they're excited to reconnect
and say hi. That's fun. It starts to feel more like a high
school reunion.
Kirill Eremenko: That's really cool. We do have some people coming
back for the third year on. It's really exciting to have
returning guests to the [crosstalk 00:50:36].
Ben Taylor: Out of the other conferences, I'll put in a few weeks of
thought before the talk, but DataScienceGO, for the
second or third year in a row, I definitely am thinking
like six months before the talk like, "What do I want it
to be? What's the wow factor? What's the messaging?
What's the takeaway?" I get excited about that. I think
it's important to have the one talk that you get realy
jazzed about and maybe you get over your skis a little
bit. You set a goal or you set some expectation and
you've got most of the year to kind of stew on it and
hopefully motivate and deliver where-
Kirill Eremenko: That's really cool. That's very cool. A lot of speakers
just reuse the talk in many different conferences.
Ben Taylor: Yeah. I don't do that at DataScienceGO. One of my
commitments to the people listening that will go to
DataScienceGO this year, there will be things in my
talk that I will be showing that no one has ever seen
before. Like ever. They won't just be, "Look at this AI
application." They will be benchmarks and numbers.
This is the reaction I want. They're just like, "We see
those numbers and we see what's done. How?"
[inaudible 00:52:08]. That's the icing.
Ben Taylor: I don't want to be like a mystique or magic or I'm
withholding. It's a lot of work. It's really, really hard,
and you have to do a lot of stuff to kind of plow
through these milestones. I'll talk a lot about it in the
talk, but some of the things could end up being trade
secret and stuff where I can't roll back the full kimono
and say like, "This is why we're going 30 frames per a
second on a single CPU thread," or stuff like that.
Kirill Eremenko: Got you. What I like about your talks is that they're
different every time. This time, Call of Duty. Last time,
you were talking about passion and obsession, and,
was that “There's a transition on who you want to hire
and how to get hired.” Really cool. [inaudible
00:52:51].
Ben Taylor: I feel like I get bored easily. If I had to give the same
talk again, it would be really boring for me.
Kirill Eremenko: Yeah, I can imagine.
Ben Taylor: It'd be boring for the audience too if they're coming
back. They want to see something new, something
inspiring, something different. The talk this year is
really focusing on the models that industry needs.
They're so much more intimidating than what I
thought industry needed as a data scientist. I'll be
going through some of these models and thankfully
they're becoming easier to build. You don't need our
company to build these models, but they're becoming
these very complicated, mixed dataset models where
YouTube advertising or cheating detection or ...
There's just a lot of different data elements floating
around.
Ben Taylor: The idea of you building an image classifier that is
game changing for a companies is kind of laughable
today because it's hard for me to think of an
application where that would be that important. If I
can predict do you have a swimming pool from space
for insurance, we could build a deepnet that looks at
an image of your house and it predicts swimming pool,
no swimming pool. AI can do that. Deep learning can
do that, but the problem is what is that worth?
Literally, what is that worth? It's only one thing. What
is it worth? It's worth more than zero, but it's not
worth $10 million for that. That model's not worth that
much. When you get into these mixed models, the
numbers get really big because they typically have a
big impact on the business.
Kirill Eremenko: Yeah and then have a compounding effect as well on
each other.
Ben Taylor: Exactly. When you start combining these different
datasets, the amount of lift ... We do those
benchmarks internally. We're benchmark structured
only to these models and we'll see some significant lift
differences. That's actually what we get paid on. We
get paid on the delta.
Kirill Eremenko: Yeah. Makes sense. That's the best way to do it, right?
Ben Taylor: Yeah.
Kirill Eremenko: Add value, you get paid on value. Ben, we're slowly
approaching the end of this amazing third session that
we're having now. What is your one piece of advice
that you can give to our listeners who want to do
things that you do, who want to get into AI, do passion
projects, create cool stuff, maybe start a company, do
amazing things? Then next time, when you're come
here for the fourth one, there'll be a new piece of
advice, but until they hear from you next or until they
see you at DataScienceGO, what's your one best piece
of advice for them to succeed in their undertakings?
Ben Taylor: I think the best piece of advice I have for them is to
take some risks. Don't work at the same company for
very long. I'm sorry for their employer, but try to work
somewhere for a few years and go to a different place
to challenge yourself. It's really important for you to
figure out what your strengths and weaknesses are.
There some things you're really good at and there's
other things you're not in. The sooner you can figure
that out, the better because maybe you can find a co-
founder to compliment your weaknesses or you can try
to protect yourself from them. It's really important that
you know what your weaknesses are. If you know
where your weaknesses are, then you can protect
yourself from some of these pitfalls.
Ben Taylor: Maybe this will sound cheesy. You only get one career,
so knowing you only get one career, why do you want
to go work for a company for 20 years doing something
that wasn't ... it didn't impact the industry. It wasn't
something you can look back on. The other thing I
want to throw out there is a lot of times we think
about our resume, but there is a startup resume. As
you go and raise capital, sell a company, raise capital,
sell a company, there's some life-changing
opportunities that can come from that. Not just the
money, but your momentum and your ability to tackle
a new idea.
Ben Taylor: Here's an idea that no one's tackling. They have
autonomous mowers that you can buy today for
$3,000. It's like a Roomba. If you have a backyard that
has an electric fence, this mower will come out and it'll
just go around the yard. It's very quiet, mows every
single night. It's really dumb, but it works, and people
pay for it. I think there should be a company today on
the market that has an AI system on top of that mower
that is killing weeds with lasers at night with AI, and
the technology elements are very doable. It's not
science fiction. It's like, "Hey, I'm going to give you $1
million. You go do it. You give me $1 million, I'll go do
it," but we're busy, so we're not doing that thing.
Ben Taylor: I would love for your listeners to eventually get to a
point in their career where if that sounds exciting, they
will go do that thing. They'll just go do it. You have to
take risks. You have to tool up, challenge yourself, get
to where you can do that. I feel like that project I just
suggested, if that became your passion project, you
have the resources to figure it out, you'll figure it out.
Kirill Eremenko: Yeah, just resources, being resourceful, exactly. I feel
excited about that. As you said, I'm busy, but if I have
the time, no problem. Give me 1 million bucks, give me
a year, it'll be done.
Ben Taylor: Yeah, but if I told you that project 10 years ago, you'd
be like, "I don't even know where to start. I don't even
know how to tackle that project." Today with a little bit
... Not a little bit ... with a lot of experience, a lot of
mistakes, a lot of things deployed, value added all over
the map, maturity and then some reputation attached
to that, you could pull that off. How fun would that be
if that was your passion project? If you just went head
down for the next two years and you changed the
world, where there's never another blade of crabgrass
... Some people might think that sounds kind of
stupid, but for me, I think that's amazing. If you did
that, that is amazing, and I will buy that lawnmower
from you for the $10,000 or whatever it is.
Kirill Eremenko: Yeah.
Ben Taylor: Because that's amazing.
Kirill Eremenko: That is true.
Ben Taylor: And I don't want ... I think it's important for people to
plan their five and 10-year goals, and there's no
reason why someone can't have that in their sights or
something similar.
Kirill Eremenko: Yeah, and it doesn't have to turn into a company,
right? You can do it as a passion project and through
the recognition you get, Microsoft will come and give
you an offer or I don't know, Google will want to take
you and your team on board even before you
incorporate. They'll just see the potential, the value
that you're bringing through passion, through what
you're working on, and that's it. You have a new job all
of a sudden.
Ben Taylor: Yeah. Another point I want to shoot out there is I think
sometimes people think that this is the time for AI
startups, this is the time to do AI, but you and I will
live the rest of our lives with more AI opportunities
than we can handle or think of. There are tens of
thousands of startups from now until we die that are
related to AI, that are niche applications like the
mower example or something else. Huge impact. Plenty
of opportunity for your listeners, and I would definitely
recommend going that route. It's not an easy route.
You could probably tell from the look on our faces. It's
not ... It's not [crosstalk 01:00:15].
Kirill Eremenko: The beards. [inaudible 01:00:17].
Ben Taylor: Yeah, it's not an easy out, but maybe the hope is a
year from now, our paths cross in Costa Rica and
we're surfing and throwing mangoes at coconuts for at
least a week before we go do the next thing.
Kirill Eremenko: Yeah.
Ben Taylor: It's still worth it, but yeah, anyway ... Always happy to
answer questions and I know you are as well based on
your availability to people as they have these ideas or
questions.
Kirill Eremenko: Yeah.
Ben Taylor: Perfect.
Kirill Eremenko: Yeah. When are you going to write a book? We're
waiting for a book from Ben Taylor.
Ben Taylor: When are you going to write a book?
Kirill Eremenko: I wrote a book last year. [inaudible 01:00:55].
Ben Taylor: Oh you did?
Kirill Eremenko: Yeah. Your turn.
Ben Taylor: I need to read that book. Man.
Kirill Eremenko: I will give you one at DataScienceGO. It's the purple,
Confident Data Skills, but when's yours coming out?
Ben Taylor: Six months ago, I would have said never, but then I
ran into someone at Dublin Tech Summit. He was a
New York Times best seller, and he was telling me
about the process. I thought, "Man, that sounds
terrible."
Kirill Eremenko: It is.
Ben Taylor: That sounds terrible to go through the work to write
that book. He said, "No, it was actually really easy." I
said, "What do you mean it was [inaudible 01:01:32]?"
Yeah, you reacted too, like, "What do you mean? How
does that sound easy?" He said, "No. Every one
morning, I woke up, I turned on my recorder and I
spoke in the car. I'd literally say, "Chapter one," and
I'd just ramble. Then the next day, chapter one,
chapter two, chapter three. He just had hours and
hours and hours of just whatever was in his head. He
gave it to a writer, and he paid them a lot of money. He
paid them $70,000 or whatever. It sounds like a lot of
money. Maybe that's not ... I guess you could pay
someone hundreds of thousand dollar. They wrote him
a New York Times bestseller and put his face on the
book.
Kirill Eremenko: Fantastic. Actually, it's the same way that I did it.
Ben Taylor: [crosstalk 01:02:09].
Kirill Eremenko: I recorded everything with audio. I'm [crosstalk
01:02:11]. I'm no good at writing. I just recorded the
sentences in the audio, what I wanted to convey. Then
you find a writing partner who helps you put it into
text and you review it.
Ben Taylor: I didn't know you could do that, and that sounds
[crosstalk 01:02:26]. Yeah, so you're a step ahead.
Kirill Eremenko: The point was there to get your thoughts out there, to
get a medium for people to read it. It's just one of the
ways to do it. You should. You totally should.
Ben Taylor: Maybe if I control Microsoft really, really bad with the
live Twitch feed, then I'll write a book about that,
about how angry they were, but anyway ... If we do get
a liquidation event, then I'll tell you all about it.
Kirill Eremenko: All right. Okay. Ben, it was a pleasure having you on
the show for the third time around. Thank you so
much for everything you shared, and I look forward to
seeing you at DataScienceGO.
Ben Taylor: Yeah, you as well. Excited to go there. Month and a
half away.
Kirill Eremenko: Fantastic. All right. See you.
Ben Taylor: Okay. See you.
Kirill Eremenko: There you have it, ladies and gentlemen. That was Ben
Taylor. I hope you enjoyed our conversation as much
as I did. If you'd like to hear more from Ben, if you'd
like to see how that AI playing, Call of Duty project
turns out, then come on over to DataScienceGO this
year. It's happening on the weekend of the 27, 28,
29th of September. You can still get your tickets at a
discounted price. The prices are going up on the 26th
of August. If you jump on www.datasciencego.com ...
That's datasciencego.com ... you can secure your seat
for the conference there and meet Ben along with other
speakers from companies such as IBM, Google,
Salesforce and many, many more.
Kirill Eremenko: As always, you can find the show notes for this
episode, including the video version of this episode at
www.superdatascience.com/289. That's
www.superdatascience.com/289. We'll include all the
materials mentioned in this episode over there. Once
again, thank you so much for being here today. Don't
forget to grab your DataScienceGO ticket before the
prices go up on the 26th of August. I look forward to