Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:20
Back in 1990, Robin Allenson was studying
0:20
AI at university and he fell hard for
0:27
the vast potential of this technology.
0:29
So Robin went on to build several
0:29
AI startups, including similar AI's
0:34
current project, an intelligent SEO
0:34
platform for corporate websites.
0:38
Robin's been in the AI game for over
0:38
two decades and this past year he's
0:43
companies that are
0:43
actively seeking AI solutions rather
0:48
than dismissing them is too risky. There was a joke in AI , in the
0:50
nineties that all of the working
0:53
AI models were called software. because you do some research, you get
0:55
something to work, and then people
0:58
would just assume it was software. Cause software was actually
0:59
the thing that worked but now we have inbound from people
1:02
who are searching for ai for seo, ai
1:07
for internal linking, ai, for new
1:07
page creation and topic research.
1:10
so it feels like the whole
1:10
world has changed around us.
1:15
Join us to get an insider perspective on
1:15
the latest AI developments, and discover
1:20
how Robin found success by defocusing
1:20
on tech and understanding his customer's
1:25
pressing problems and struggling moments.
1:33
Thank you Robin for being here.
1:35
I met Robin years ago when we worked
1:35
together, on an earlier version of his
1:42
startup, which he's gonna catch us up on.
1:44
Robin is one of the smartest
1:44
and most inspiring and,
1:49
forward thinking people I know. He's been working in AI
1:51
for a really long time.
1:54
And I think he's gonna help us
1:54
understand what's going on right now.
1:58
That's very kind of you, Amy, Joe. I don't know, if all of that is entirely
1:59
true, but, I'll try and live up to that.
2:03
, it's like the blind men
2:03
in the elephant, right?
2:05
Like we all can only speak
2:05
to our own, experience.
2:09
So, start by please, Robin,
2:09
introduce yourself and tell us
2:14
about the startup you're working on
2:14
now, who it serves, what it does.
2:19
And then we're gonna drill
2:19
into some of your background.
2:22
Cool. Thanks. so Hello, I'm Robin. so I'm the CEO and the co-founder
2:24
at Similar AI and Similar
2:27
ai, we're automating SEO for
2:27
agencies and for in-house teams.
2:32
So search engine optimization is a very
2:32
manual job typically, and there's a lot of
2:37
folk who are, working perhaps in a large
2:37
enterprise site or perhaps in an agency.
2:41
And, mostly they're taking some of the
2:41
existing SEO tools out there using their
2:46
user interface and spitting out some
2:46
data which they throw into an Excel
2:49
sheet, or a Google sheet, and spend
2:49
a lot of time manually updating that
2:53
and trying to match that up to their
2:53
pages to take some kind of action.
2:56
and what we're enabling is basically
2:56
teams to, to use our no-code, tool, and
3:02
toolbox to automate that, to speed it up.
3:05
So , we are kind of, taking a those manual
3:05
things and sticking all the data together,
3:10
and we call that data ingredients. And on top of that, we allow our, users
3:12
to run, recipes we call them, which are
3:16
no-code programs, which can publish and
3:16
update , their websites to make them more
3:21
user-centric and grow organic revenue. so that's, kind of in a nutshell.
3:25
I could talk for days about it,
3:25
but I'll try and keep it succinct.
3:28
So who's your core target? Who needs this the most?
3:32
so, for the longest time we've
3:32
been, focusing on, in-house SEO.
3:37
And typically there's one SEO working for
3:37
one, search engine optimization, product
3:41
manager working in a big enterprise.
3:44
the site has perhaps hundreds of thousands
3:44
or millions or tens of millions of pages,
3:49
and they, on their own sum have to update
3:49
the site to match it up to how people
3:54
are actually searching and make it more
3:54
intuitive , for search changing users
3:57
or as we call them nowadays, people,
3:57
and try to make the whole site work.
4:02
and so that's just a monumental task. SEO is normally something you
4:04
do by hand, and so it's just
4:06
incredibly hard to do that. Normally the way you solve that is
4:08
maybe you get an agency involved.
4:11
And you get a lot more hands, on deck. or you use some in-house, tools and you're
4:14
using, you're building that yourself.
4:18
But when you use in-house tools,
4:18
mostly you have in-house data.
4:21
And so that becomes a struggle
4:21
to actually match up to how
4:23
users search on the internet. so that, that struggle means that a
4:26
lot of those sites end up with too
4:30
many pages, and those pages aren't
4:30
interesting for search engine users.
4:34
so that's the struggle that we're
4:34
trying to help and we're trying
4:36
to turn those users, effectively.
4:40
So one of our customers, they said
4:40
that we give our, we give our users SEO
4:45
superpowers, which is, I think it's a
4:45
great way of thinking about it, using ai.
4:49
Cool. So we're gonna get to how you came up
4:50
with this, more about how it works.
4:55
and before I get into your background to
4:55
figure out, you know, like sort of how you
4:59
got here, I can't resist asking about the
4:59
relationship between SEO and Chat GPT and
5:07
the explosion of chatbots, because a lot
5:07
of people are saying, wow, who needs SEO?
5:13
Websites are dead. Why are websites even gonna matter in a
5:15
few years if we're all using something
5:21
like Chat GPT instead of a search engine?
5:25
I'm sure you've heard
5:25
the same argument, right?
5:27
What's gonna happen to
5:27
websites with all this stuff?
5:30
Is it, you know, what's your perspective on this? You're right smack in the middle of this.
5:35
So I think there's a few things,
5:36
which are kind of, interrelated,
5:39
and where this is impacted. So people have been saying SEO is dead for
5:41
many years, and it, it hasn't died yet.
5:46
and, but there are a few impacts
5:46
where, that the ChatGPT or where large
5:51
language models, could, fundamentally
5:51
change, what the SEO industry is doing.
5:56
one of those is search results. So Google has this Ste so this
5:58
experimental approach to integrating,
6:02
large language models in search. and so, the initial feedback
6:04
on that is it, is that those
6:09
results haven't been superb. large language models have a number
6:10
of, weaknesses, such as their ability
6:15
to, yeah, I call it confabulate.
6:18
so, confabulate is a, a term kind
6:18
of, meaning they bring up false
6:22
memories, is what happens in psychology
6:22
commonly called hallucination.
6:25
So these things have a tendency to lie,
6:25
egregiously, and in search results,
6:30
I think that's a very different
6:30
kind of intent than when you are,
6:33
when you're chatting to ChatGPT. that can be a very bad problem,
6:35
for certain types of results.
6:38
I also think that right now,
6:38
ChatGPT is a toy that's being used
6:42
in a lot of interesting cases. And there's kind of an adage in product
6:44
management is that a lot of the best
6:47
products, start from toys, right? and so, maybe it will grow into
6:48
that if we can solve a lot of
6:51
these, hallucination problems. but then, So that's the first piece is
6:53
it's already a disruption in that we see
6:57
that Google is changing the front page
6:57
of search in order include AI results.
7:02
and actually messing with not
7:02
just seo, but also, the revenue,
7:06
that it gets from an ads, right? So it's willing to sacrifice that, to
7:08
contend with this potential, competitor
7:13
in the form of OpenAI and Bing. That's now, I think there are
7:15
still a lot of unknowns, like
7:19
will that actually work out? our hallucinations a solvable
7:20
problem, for web search?
7:23
Don't know. we don't have a clear answer on that. So the assumption is kind of, yes, it's
7:25
all gonna work out, but it doesn't, there
7:29
are very large categories where they
7:29
just can't use this kind of approach.
7:33
You already see in a lot of, like
7:33
your money or your life categories.
7:36
So those are things where you might
7:36
give medical or financial, information.
7:39
they're not using AI there, cause you
7:39
can imagine that the risks are too great.
7:44
so it could be that actually the
7:44
number of places where you use AI
7:47
shrinks over time, where they're more
7:47
focused on which kind of answers they
7:51
can, reasonably and rationally give. so the second piece of seo, disruption is
7:53
it's a lot easier to, to produce content.
7:59
and so the internet kind of
7:59
disrupted the ability to publish.
8:03
content that's been generated, large
8:03
language models and, AI is now generative.
8:07
AI is now disrupting the ability to come
8:07
up with the content in the first place.
8:12
and, and so that's much in the
8:12
way that the internet democratized
8:16
access to publishing information. so, large language models
8:18
and AI democratizing access.
8:21
So it's no longer just say,
8:21
copywriters who can write copy,
8:25
but everybody can write, copy. and so that, means that there's
8:27
just an awful lot more competition.
8:30
and so that's gonna make
8:30
finding great answers harder.
8:34
So a aka web search. so that's another form of disruption.
8:37
And then the third piece is there
8:37
are a lot of, folks in the SEO
8:41
industry who make most of their
8:41
money from generating content today.
8:44
and so when the price of that effectively
8:44
drops to zero, it's gonna be much
8:48
harder to make money doing that. It's always been hard in a way,
8:50
because, what you were really
8:54
paying for was amazing, content,
8:54
excellent content, fabulous
8:58
content you can get anywhere else. And so I still think that there will
8:59
be folks selling that kind of content,
9:03
and it'll be hard to write that with ai.
9:06
but, yeah, it's got harder to do, because
9:06
there are lots of incredibly cheap, almost
9:11
free alternatives out there using ai. So those are the three
9:14
kind of disruptions I think. Wow, that's really interesting.
9:18
That really helps me
9:18
understand the landscape.
9:21
Thank you. So let's wind it back and can you
9:23
explain to us how you got from
9:29
tinkering around with software as
9:29
a kid to studying and developing AI
9:35
and becoming a serial entrepreneur?
9:37
yeah, when I was, I wanna say
9:37
seven or eight, but, I'm probably
9:41
getting the timeliness all wrong. But my dad came home, one on Christmas
9:42
time with a computer called a ZX81.
9:47
So one of the earliest computers. And, I started, as a kid
9:49
to learn, to program.
9:52
so I'm thinking ZX81, probably
9:52
in 81, so I was probably 10.
9:56
but then, I started to learn to
9:56
have a program in basic, and then
9:59
in Assembler, when I was about
9:59
11, I read this book called Gödel,
10:02
Escher, Bach by Douglas Hofstadter. and I was smitten.
10:06
I read that a couple of times. I just completely fell in love with it.
10:09
and I started writing programs as a
10:09
teenager to do things like, translate.
10:14
Between two languages. So I'd learned Esperanto as a kid,
10:15
and so I tried to build a machine
10:19
translation program failed miserably, but
10:19
it was a very interesting, experience.
10:23
and I also built a program as
10:23
a kid to, predict which
10:26
horse was gonna win the race. So you could send away in those days
10:27
for books, in the post I sent away
10:31
for a book, explaining a horse racing
10:31
system where you gave different points
10:34
depending on how the horses are traced. but then I took that system and I
10:36
turned it into a computer program
10:40
and I would enter in the, the
10:40
horses that won the races, and which
10:43
horses were coming up each week. so you could just, find
10:45
that in the newspaper. And I tagged that all in.
10:48
And, so, that program worked
10:48
and it was profitable.
10:52
there were two downsides at the time
10:52
was that, we made about, we would've
10:55
made about a 15% year betting tax
10:55
in the UK at the time was 12%.
10:58
So that wasn't, amazing. And then, yeah, I was a minor and,
11:00
my parents didn't really want to go
11:03
to the, to the bookies every day. so that project kind of failed, but
11:05
I took those, ideas and, aspirations.
11:10
And then I studied, ai, artificial
11:10
intelligence and computer science
11:13
and, uh, joint honors at University of
11:13
Edinburgh, which at the time was the only
11:16
place in the UK where we could study ai. now, AI is, everywhere.
11:20
and then, at, yeah, at the time my dad
11:20
was a research physicist, and he kind of
11:24
said, I want to do AI and linguistics,
11:24
and he said, don't do AI and linguistics.
11:28
You're never gonna find a job. there's no real future in that.
11:30
So I took his advice and did a
11:30
joint honors with computer science.
11:33
but like a couple of years after that,
11:33
Google was hoovering up every AI and
11:37
linguistics PhD , in the country. but then I didn't, study a PhD.
11:42
I went straight into, programming's
11:42
job and gradually did some different
11:45
jobs in, programming, kind of
11:45
technical architecture, and then,
11:49
kind of running, larger commercial
11:49
teams, and more and more on the kind
11:53
of, online business side of things. And then fast forward to
11:54
2009, did my first startup.
11:58
and then, we worked on lots of really
11:58
interesting machine learning problems.
12:02
that didn't really pay the rent. and then we found a really boring
12:04
software problem, that everybody wanted,
12:07
and we started to scale very quickly. We got bought out by a competitor of ours.
12:11
so I took some of that, money and then
12:11
in, 2016, a long, long, long time ago,
12:16
I sometimes say when at university I
12:16
studied AI when it was mostly philosophy.
12:21
it's funny to think that a lot of
12:21
the essentials of what we're doing
12:23
now with large language models,
12:23
the way that, AI works way, deep
12:28
learning works was still what we
12:28
studied at university in the 1990s.
12:32
So back propagation and neural networks,
12:32
still the principles are the same,
12:36
but we've had a few, fundamental
12:36
restorations in how the algorithms work.
12:40
Like reinforcement learning as an example. but also we've got, I don't know, a
12:42
trillion times more compute and data.
12:46
and so that's mostly what's actually,
12:46
I think given the power into,
12:49
into current life language models. anyway, that, that was kind of a
12:51
whirlwind back and forth and kind
12:55
of all kind of other things that, I dunno if I actually answered. It's fun though.
12:57
It's interesting that there's
12:57
a whole thread of probability
13:02
running through your story. Yeah. Messaging probability.
13:05
there's some inevitability to, to
13:05
ending up doing what I'm doing.
13:09
And, interestingly, my, 16 year old
13:09
is about to study statistics and
13:14
she's like, what's it all about? And I'm like, well, probabilities.
13:18
Yeah. It's, that's kind of it so much
13:22
probabilities. Right. And you know, when you talk
13:24
about reinforcement learning,
13:27
it's, you know, sophisticated
13:27
messing around with probabilities
13:30
if you learn to think in, in kind
13:30
of in more dimensions, a lot of this
13:34
is kind of, if you can think about
13:34
weight spaces, A lot of the stuff we
13:38
were thinking about then was how you
13:38
would think about an AI system now.
13:42
so I was gonna say, so my son is
13:42
studying AI at university, today.
13:46
So my son is, 20. and I was kind of surprised that a lot
13:47
of what he was studying, I recognized
13:51
what was going on in the textbooks. and it wasn't something fundamentally
13:53
different than the AI that I was studying.
13:58
what is it like, 20,
13:58
30 something years ago?
14:00
and so the fundamentals have
14:00
really stayed very similar.
14:03
It's just do an awful lot more with
14:03
the compute and the data we have now.
14:07
and there are some, yeah, there have
14:07
been some big leaps forward, but
14:10
those big leaps forward are still,
14:10
still based on, , on the same ideas.
14:14
It's pretty fascinating. So we'll see where this
14:16
whole craze goes, right?
14:19
Like in a year what's
14:19
gonna be happening with
14:21
. Ai. You nailed it. There's so many unknowns,
14:25
I mean, like you also said it,
14:25
there's a continuity here as well.
14:28
It's not like large language models
14:28
have popped out of nowhere and
14:30
a suddenly this amazing thing. We've had, oh no, we've had
14:32
kind of big data being a thing.
14:35
We've had AI being a thing. We've had deep learning being a thing.
14:38
We've had reinforcement learning being a thing. We've had gans, all kinds of other
14:40
kind of things where people were
14:43
going, wow, this is the moment.
14:45
Now we have large language models. I don't know if, I don't think the
14:47
large language models are the moment.
14:51
But it is pretty clear that there
14:51
is a, there's a line going through
14:54
all of that, where things are
14:54
getting more and more capable.
14:58
and the downsides to, to these kind of
14:58
connectionist systems, still remain,
15:03
but many of them are getting improved. yeah, so I dunno if this is the
15:05
moment when AI kind of, let's say
15:09
takes off, but then, it does appear
15:09
to be taking off in a way that it
15:12
hasn't in the last, six months. but I expect there'll be another
15:14
evolution of AI in, in a few years.
15:18
and it will just keep on going. So
15:20
as someone who's been inside of this
15:20
and wrestled with the difference
15:26
between the promise and the reality
15:26
of ai, what is your take on.
15:34
The doomerism and the, oh my God,
15:34
I'm gonna quit my job and go around
15:40
the country telling everybody
15:40
how dangerous this moment is.
15:44
there's obvious things
15:44
around, misinformation.
15:47
Right. what is your take?
15:50
How do you feel both as a technologist
15:50
and just kind of as a human and a parent?
15:55
So, so I think, so there are a
15:55
few different, parts of my answer.
15:59
right. So one of, there are folks in the
15:59
industry who've been studying the
16:03
risks of AI and how to make AI
16:03
fairer, and work better for, everyone.
16:09
and mostly what I see, they've been
16:09
kind of, I dunno if it's literally
16:14
fired or just ousted from, the big tech
16:14
companies that we're employing them.
16:18
not a great sign really. and then, but, and they are
16:20
really looking at say, how.
16:23
generative AI is using deep fakes
16:23
for misinformation, uh, campaigns
16:27
in actual stuff that's going on
16:27
now that can incite violence.
16:30
That's scary worrying, stuff that
16:30
I think we could be, actively
16:35
doing something about now. But I see people not doing stuff about
16:36
that and talking about how, AI is
16:41
gonna bring about an extinction event. So I think that seems a bit strange
16:43
to me Also, , I would prefer to
16:46
focus on the present day dangers. and how, LMS can make yeah, for
16:49
instance, misinformation, but also bias.
16:54
So these models concentrate the bias
16:54
that's naturally occurring in data, which
16:59
is data that's trolled from the internet. and, that concentration, I
17:00
think, can be very risky.
17:03
and we see models hallucinating in ways
17:03
that could be, could be bad for people.
17:08
that doesn't mean that, that we're gonna
17:08
have an AI that's gonna be Skynet and
17:12
take over the planet and kill everybody. it's just more like, this is a very
17:13
powerful tool, that can be misused
17:17
and it is currently being misused. so I, I think those things are risky.
17:20
A lot of the guardrails that have
17:20
been placed onis, and large language
17:25
models, yeah, like there are pretty
17:25
easy ways of getting around them.
17:28
and, it's not clear how that,
17:28
could work or should work.
17:32
I think that, I think that, say OpenAI
17:32
calling for some kind of centralized
17:37
agency, to look at AI risks, is
17:37
also a little bit self-serving,
17:41
cause I think that's gonna push
17:41
towards, more centralization.
17:45
only the big tech companies will
17:45
actually be able to afford to, conform,
17:49
to whatever regulation is put in place.
17:51
It's incredibly early in large
17:51
language models to be thinking
17:54
about that kind of regulation. and I think there's a burgeoning open,
17:55
software scene for large language
18:00
models, which will be effectively
18:00
killed off by that kind of regulation,
18:04
which seems pretty handy for big tech. Right.
18:07
who are mostly interested
18:07
in selling those, models.
18:09
That's also maybe the reason why the
18:09
folks who are researching this kind
18:14
of, the potential misuses and what
18:14
could go wrong, when they actually
18:17
wanted to publish the results of those. Yeah.
18:19
Big, big tech broadly. So I think in that case, Google was
18:21
not very interested in them doing that.
18:24
and so that was the conflict. so I think it just feels very, it
18:26
feels a little bit one-sided right now.
18:30
and so I'd prefer to listen to the people
18:30
who are actually, have been studying
18:33
this for many years rather than, so there
18:33
are a lot of kind of, one of my first,
18:38
my thesis at university was looking at
18:38
some of the work that Jeff Hinton does.
18:42
he's an amazing researcher. I don't know, just because he's an amazing
18:44
researcher in AI means we should listen
18:48
to, his ideas, on like what AI will bring.
18:52
I don't know if being a great
18:52
science researcher means that you
18:56
are good at thinking about the
18:56
conflict, with the science and AI
18:59
and technology and, and society. there are other people who have
19:01
been studying those things. I think we be giving more,
19:02
credence to what they say.
19:06
yeah. So again, long rambling answer.
19:09
I dunno if that helps you. Well, I like it because it helps me think about
19:10
the space, which is what we're all
19:16
trying to do is have a position, right?
19:18
Because so much info coming
19:18
at you all day, every day.
19:23
It's overwhelming. So many certainties, right?
19:26
I found it all. Like, I don't know.
19:29
And so I think, from what I read,
19:29
there was a questionnaire going around
19:33
asking Ai lumin and, researchers.
19:36
The probability of some
19:36
kind of doom scenario.
19:39
And the vast majority, I don't
19:39
remember the figure, 90 something
19:42
percent just abstain from answering
19:42
that cause it seems unanswerable.
19:46
and the, you know, 1, 2, 3 something
19:46
percent that did answer it, then they
19:50
summarize those answers and then suddenly,
19:50
that's what people are talking about.
19:54
I think it'll be more useful to talk
19:54
about like, that these are unknowns,
19:59
rather than these are quantifiable
19:59
in some way cause they're not.
20:02
and I think , forcing people to
20:02
quantify it seems, seems strange cause
20:06
then obviously you take some tiny
20:06
percentage and multiply it by infinity.
20:10
and I think there was some
20:10
kind of, pascals paradox,
20:13
something like that, right? so you end up with, oh, we have to
20:14
dedicate a lot of resources to that.
20:17
I think. We should be dedicating resources and
20:18
most of the people who are, campaigning
20:22
for that are the folks who are actually
20:22
founding these, these large language
20:25
model, companies in the first base. they're also people who are using all
20:27
of the crawl data without thinking
20:30
about, say, copyright issues, or
20:30
thinking about, you know, how they're,
20:35
displacing some of the people, who
20:35
first came up with the content.
20:38
and just publishing that as quickly
20:38
as possible, publishing those
20:41
models as quickly as possible. and then saying, Hey, we should
20:42
have some kind of regulation. it, yeah, it feels very one-sided to
20:45
me, like, as if we already know what
20:50
the problem is and how to solve it. I think we should, look at the
20:51
people who are, like you say,
20:54
talking about the space, and talking
20:54
about the space of, of questions.
20:59
yeah, I think their questions are really interesting. I don't purport to have any of the
21:01
answers, but I think the questions
21:03
are a lot more interesting. Yeah, I think it's something
21:05
we're all trying to understand.
21:07
We had, Douglas Hofstadter, the
21:07
author of Gödel, Escher, Bach, which
21:11
you mentioned earlier, as a guest
21:11
a few weeks back, and he's pretty
21:16
upset and depressed about it all.
21:19
Yeah. I couldn't get a positive angle
21:19
out of him even though I tried.
21:23
But you know, everyone's
21:23
got their own point of view.
21:26
And again, just cause you wrote
21:26
a bestselling book and worked as
21:29
a researcher and as a theorist,
21:29
doesn't necessarily mean that
21:33
you have insight into everything. But I like that you're in
21:36
there working positive.
21:39
The reality is the book's not written.
21:42
The final book on AI is not written. We're writing it.
21:44
We're all writing it. Indeed. So, I'm glad to be able to
21:46
hear your point of view and,
21:51
you know, give it some oxygen. So
21:53
I mean, take it with a
21:53
pinch of salt, everything.
21:55
Like my curse as a founder is
21:55
to be perpetually optimistic.
21:59
and so I always think the world
21:59
is slanted towards, good things
22:02
and a happy, hopeful future. so in my head there's always those
22:04
kind of good things going on.
22:08
So I tend not to focus on
22:08
the, cataclysmic, potentials.
22:11
I also just think we have plenty of those. Like, I heard a quote, uh, this,
22:13
um, Zi Maza wrote a bit about this,
22:17
and, he quoted in his newsletter,
22:17
somebody else, saying, climate change
22:21
should have, um, AI's, PR team.
22:24
I think right now being so sure
22:24
that AI is gonna bring this terrible
22:28
future when we have science talking
22:28
about climate change, like with this.
22:34
Enormous impact. I feel like there should be some
22:35
enormous evidence, that, that this
22:39
is really coming for climate change.
22:41
I think we have a lot of that
22:41
enormous evidence and we should
22:44
maybe be focusing on that rather
22:44
than this very speculative, stuff.
22:48
yeah, I would prefer to focus on the stuff
22:48
we, we really know about and we think we
22:52
can actually change, rather than this,
22:52
it feels way too theoretical for me.
22:57
Spoken like a true entrepreneur
22:57
who likes to get his hands dirty.
23:02
So, I wanna follow up on
23:02
your entrepreneurial mindset
23:05
that we're talking about. You're a serial entrepreneur and
23:07
having now worked with a number of
23:11
serial entrepreneurs, there's something
23:11
qualitatively different doing it
23:14
multiple times versus the first time. So from your own perspective, What
23:17
are some mistakes maybe that you
23:23
made that you, perhaps you see other
23:23
entrepreneurs making common mistakes,
23:29
especially for first time entrepreneurs,
23:29
for folks that are interested in maybe
23:33
having an awareness of those, what are
23:33
some of the things to watch out for?
23:37
Both for AI entrepreneurs, but
23:37
also entrepreneurs in general?
23:41
I was gonna say, daughters
23:41
entering living rooms when you
23:44
are, trying to, have a conversation
23:44
is one that keeps on coming up.
23:47
But, um, so, I think one of the things
23:47
that looks like, initially it's gonna
23:51
be unique to AI entrepreneurs, but is
23:51
actually, very common across everybody.
23:55
And it's also for me is, falling
23:55
in love with your secret source.
23:58
so that's kind of focusing on the solution
23:58
space rather than the, problem space.
24:03
And I think, there've been a lot of,
24:03
recent AI companies which have exploded.
24:08
Which are something like,
24:08
a thin layer on top of t.
24:12
And, kind of the answer has
24:12
been AI is gonna solve that.
24:16
I remember when I first started,
24:16
similar ai, we were starting to use,
24:20
deep learning and we had angle on how
24:20
to create a lot of training data, and
24:24
come up with, more training data means
24:24
basically better models at scale.
24:28
and so we had an angle on that, but
24:28
we had a couple of, advisors to the
24:31
company and they said, but Robin,
24:31
deep learning, you don't need training
24:34
data cause it's deep learning. and so they said it in, in a way that
24:37
made me think that deep learning was
24:39
actually in italics when they said it. so I hear something very similar about
24:41
large language models now where people
24:44
are like, what does your company do? We do large language models.
24:47
yeah, but what, I mean,
24:47
what exactly do you do?
24:49
And so I think some of those
24:49
companies are being disrupted
24:52
effectively by ChatGPT now. and also.
24:56
a lot of folks are finding it very
24:56
easy to integrate, large language
24:59
models into the existing incumbents.
25:02
and that's actually, it's kind of harder
25:02
to just do a, I dunno, an AI startup.
25:07
so we fell into that. Initially, we were building
25:08
a lot of trading data.
25:11
We'll come out with a whole bunch
25:11
of, ideas about how to, come up with
25:14
large, well, basically multimodal,
25:14
models is what we're building.
25:17
So combining, images and texts to
25:17
understand product pages and turn
25:21
them into the language that people would use to search for them. and so that was the core thing
25:23
we were doing, but we actually
25:25
found it really hard to sell that,
25:25
because we found that a lot of other
25:28
people were claiming they used ai. and so we over time just
25:29
stopped talking about ai.
25:34
We didn't bring up in the conversation, it was in our names. It was a bit, kind of hard to
25:36
avoid, but we just didn't bring
25:39
it up in the conversation. And we started really focusing on
25:40
the problems that our customers were
25:43
struggling with, but our users were
25:43
struggling with and kind of explained
25:46
how it was different, and how we, help
25:46
them make progress, in, in that struggle.
25:51
And that was a night and day shift. and so what's been, miraculous to me
25:53
in the last six months or the last
25:56
nine months is actually, there has
25:56
been a shift , in how people are
26:00
approaching us, because now, like.
26:03
I know three years ago, five years ago,
26:03
if we actually, and certainly one of
26:06
the first a studying ai, people would
26:06
say things like, why do you do that?
26:09
Because it doesn't work. And so there was a joke in AI in
26:11
the nineties that all of the working
26:14
AI models were called software. because you do some research, you get
26:16
something to work, and then people
26:19
would just assume it was software. Cause software was actually the
26:20
thing that worked a few years ago.
26:23
If we talked about using ai, which
26:23
was in the solution space, then
26:27
folks would just be, yeah, we're
26:27
not looking for anything innovative.
26:30
We're just trying to solve this problem. Right? And so for them, AI was
26:32
something that the, no, the
26:35
business innovation manager did. and not something that
26:37
the e-commerce team, used.
26:40
But now we have inbound from people
26:40
who are searching for ai, for seo,
26:45
ai, for internal linking AI for,
26:45
de-duplication, ai, for for new
26:49
page creation and topic research. Those are the kind of things we do.
26:52
We also do AI for content. but you know, who doesn't?
26:55
so, it feels like the whole
26:55
world has changed around us.
27:00
so now people are coming out to us
27:00
and saying, Hey, we found out that
27:03
actually, you guys do AI for seo.
27:05
Well, you guys do AI for internal linking.
27:08
We're like, yeah, we do. That's amazing. Can you explain it to us?
27:11
Whilst nine months ago, if we
27:11
had said that they would've ba
27:13
basically gone, why do you wanna
27:13
use AI what's the point of that?
27:16
Now, their expectations about what
27:16
AI does have shifted, and with those
27:21
expectations, yeah, it suddenly means
27:21
we have a new and much larger market.
27:25
Again, I feel I answered
27:25
two of your questions
27:27
in that is, I wanna reflect back a little
27:27
bit what you said and really highlight it.
27:33
Cause it's such good advice. One, don't fall in love
27:35
with your secret sauce.
27:38
That's so hard. That's really hard.
27:42
But what great advice,
27:42
because customers don't care.
27:46
They don't care. They just want you to solve their problem.
27:50
Exactly. Yeah. So in my first startup, I had
27:52
a startup before this one.
27:55
We worked on lots of interesting
27:55
machine learning problems.
27:57
We were so interested in, in,
27:57
those machine learning problems.
28:00
They weren't customer problems. there were machine learning
28:02
problems that we could apply and we
28:04
thought there were great solutions. And then, when we found, and so, those
28:06
are building kind of sales reporting.
28:10
and then, one of the reports talked
28:10
about a problem we found we could build
28:14
some software to solve that problem. And it, it wasn't easy, but it was
28:15
pretty easy to get off the ground.
28:19
Once we got that going, suddenly
28:19
we found, a lot of those customers
28:22
were super interested in that
28:22
problem, in solving it more simply.
28:25
we were off to the moon, right? in the second startup.
28:28
So in similar ai, I basically
28:28
fell into the same trap despite
28:32
advising other startups. So despite saying you really should be
28:33
asking, you know, what problems your
28:37
customers are struggling to solve, and
28:37
have you thought about these kind of
28:40
questions, and, kind of, teaching them
28:40
about how to think about that and then
28:44
in our own business, I've found it
28:44
very hard to take that advice myself,
28:48
basically cause I don't think this
28:48
is something you can simply learn.
28:51
I think this is something that is
28:51
often, you can think of it more
28:55
like a cognitive bias as a founder.
28:58
Even if you know intellectually
28:58
what you're supposed to do when
29:00
you get up in the morning, you're
29:00
in love with the solution again.
29:03
and so you need to actually correct
29:03
course correct each time, to go back.
29:08
but when you do, it's enormously valuable.
29:11
So, I've got one more question actually, I've
29:12
got a bunch more, but I'm gonna only ask
29:16
one more So, um, you're talking about
29:16
understanding what problems your customers
29:22
have and that you struggle with that as
29:22
a founder, even though intellectually
29:27
you know how important it is, by the way. Me too.
29:30
I struggle with it too. So, what are some of your go-to methods?
29:35
Robin, you mentioned that you
29:35
like the magic wand question from
29:39
the game thinking methodology. Can you just share a little bit about
29:41
how you've used elements of game
29:45
thinking or maybe that particular one
29:45
to unlock customer value for yourself?
29:50
Yeah, sure. so, so we actually, we did a big piece,
29:51
but it was a few years ago before we're
29:56
actually working on, what we're doing now,
29:56
to talk to fashion designers at the time.
30:01
and we were planning on building a
30:01
model using gans, using generative
30:05
adversarial neural networks. But they turned out to be
30:07
kind of too bleeding edge. but what was amazing to me at the
30:09
time was, I was able to reach out
30:14
to, to, some fashion designers and
30:14
we got, a lot of responses, to an
30:19
initial survey within a day or two. and well, initially we didn't, and
30:21
then we tried again a couple of days
30:25
later and we got a lot of responses
30:25
with some screener questions.
30:28
and actually, some very quick five
30:28
minute, kind of 10 minute interviews.
30:32
We started like hitting gold, very
30:32
quickly when we started asking, you know,
30:37
what are the things you struggle with? and they started telling us, and from
30:38
10 conversations, I don't know, eight
30:42
of the things they were struggling
30:42
with were the exact same things.
30:45
And we were like, what on
30:45
earth is going on here?
30:48
And so we kind of fed in
30:48
love with those problems.
30:51
and we got deeper and deeper into the
30:51
conversations and we started talking more.
30:54
That was an amazing experience,
30:54
mostly because the value we got
30:57
to, in the time we spent on it. And so, it was just a such a short
30:59
time compared to building a product,
31:03
which we'd already done and spent a
31:03
huge amount of money on and wasted.
31:06
compared to building a product,
31:06
this was remarkably easy, right?
31:10
And then later in the journey down,
31:10
now, into, when we're doing what
31:14
we're doing now, we had really
31:14
focused on these big enterprise teams.
31:18
So science that might have, I
31:18
don't know, 150 million pages say,
31:22
or hundreds of millions of pages. And, we thought, wow, this is a big
31:23
market enterprise seo, talking to
31:27
heads of seo, in these big companies. and we started doing some market
31:29
research where we tested smaller sites.
31:34
that only had I, no, a few hundred
31:34
thousand pages or only had 50,000 pages.
31:38
And we kind of asked 'em about the
31:38
problems, like at least the way that we
31:41
marketed the problems that we solved. And they were effectively blank stares.
31:46
And they didn't really get what we were
31:46
talking about, but we continued the
31:49
conversation and we started asking about
31:49
some of the problems that they struggled
31:53
with, and it all opened up again.
31:55
Right? And so when we actually really focused
31:55
more about their, about their experience,
32:00
what we found was that everybody
32:00
struggles with, with automating seo.
32:05
they don't always think about it in those terms. and so the way we were talking about
32:06
it, they just didn't get that at all.
32:10
but they would say something like,
32:10
wow, internal linking is like, we
32:14
dedicate time to that every month,
32:14
and it's painful and it's tedious.
32:18
Is time consuming and
32:18
it's incredibly valuable.
32:20
isn't there a way that you
32:20
could like automate that and
32:23
we can still get the value? and so we could talk
32:25
about what the value was. They were doing it like every month
32:27
they had people working on it.
32:29
They, like, they knew exactly
32:29
why they were doing that.
32:32
and what that was true for a site
32:32
that had a few hundred pages in the
32:35
same way that it was true for the fact
32:35
they had a few hundred million pages.
32:38
and that was kind of a light bulb moment
32:38
for us when we were like, oh wait,
32:42
these guys have the exact same problem.
32:44
They just don't, express it in quite
32:44
the same way, but there's a way in
32:48
which we can, kind of reframe that
32:48
job, where suddenly it applies to
32:51
this whole enormous market basically
32:51
to everybody trying to, everybody
32:55
trying to do, customer acquisition. and so, yeah, so that was a light
32:57
bulb moment for us and that really
32:59
just came across, from having a few
32:59
of these, these customer interviews.
33:04
Before sales. Like not actually reaching out and
33:05
saying, Hey, we're gonna do sales.
33:07
Reaching out and saying, we just
33:07
wanna learn from your experience.
33:10
Can we have some simple conversations? and people just opened up and and
33:12
told us, not what we wanted to hear,
33:17
cause we really wanted to hear. Yeah. What you're doing now is great for us.
33:19
We'd love to buy that. Please. but what we needed to hear,
33:20
and it, yeah, it transformed
33:24
the trajectory of the company. I love that.
33:26
And just to frame and put it on a wall,
33:26
this piece of advice, if you can get
33:32
the right customers in a room and not
33:32
talk about the solution, but understand
33:36
their problem, you will hear the language
33:36
that they use and it might connect, it
33:42
might not be the same language you're
33:42
using, but understanding the language
33:45
your customer uses to talk about their
33:45
problem is the gift that keeps on giving.
33:50
That's it. It's what? So Amy Jo, what you taught me is sometimes
33:52
it's better to have a, your scientist
33:56
hat on, than your salesperson hat on.
33:59
So sometimes founders find it very hard to
33:59
talk about their company without pitching.
34:04
Right. And so, like the antidote to that is,
34:04
stopping talking about your company,
34:09
asking about your customer, asking
34:09
about their problems day to day,
34:13
and then, shutting up and listening. Right.
34:16
And listening is just so, so powerful.
34:18
cause at some point, even though most
34:18
people can't get a word in edgewise
34:22
when they're talking to me at some
34:22
point they'll, you know, they'll
34:24
actually explain what they're struggling
34:24
with and it will make so much sense.
34:28
Awesome. Thank you so much, Robin, for
34:30
leveling up our thinking and giving
34:33
us some really , exciting and
34:33
challenging visions to wrap our
34:37
minds around. And fun.
34:39
I think Amy, Joe, like there is a
34:39
lot of fun, uh, ahead of us as well.
34:43
Right. So it's super cool to have the power of
34:43
computation in the hands of everybody.
34:48
Right. I think it's, um, it's a it
34:48
could be a very fun future.
34:51
It is. And I'm working with more and more AI
34:51
startups and I'll be sharing that for
34:55
all of you over the next few months, so thank you so much for joining us.
35:00
Yeah, thanks, uh, for inviting me and having me here. it was just amazing.
35:03
Thanks everyone. Let's get smarter together.
35:06
There's nothing better. Bye.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More