0:02
I mean, my first reaction is dabbling ain't going to cut it. Like, we are in
0:06
the middle of this absolutely transformative platform shift. The ones
0:10
who dabbled kind of didn't get that far. Um, and so I think if you're in a
0:14
position to do it, you should bet more aggressively.
0:28
Hey folks, welcome to Agents of Scale. It's a show where we are exploring how
0:33
the world's most forward thinking leaders are operational AI across their
0:37
companies. I'm your host Wade Foster, co-founder, CEO of Zapier. And today I'm
0:41
joined by John. He's a co-founder and head of product at Gamma. If you haven't
0:46
checked out Gamma, Gamma is reinventing the slide deck. and they're doing it
0:52
with AI. Who doesn't want an AI to make all their slide decks? Sounds awesome.
0:56
Um, and uh, they've had some really impressive growth. Scaled to 50 million
1:01
users in just a year. Uh, but there is a big old story uh, behind that. Uh, not
1:07
exactly a overnight success as I think a lot of folks like to portray it. But
1:11
before we talk about Gamma, John, I want you to tell me about Platemate.
1:16
What's the story around Platemate? Uh well, happy to be here. Thanks for
1:21
hosting me. And I love that we're we're going back in my uh history. This is
1:24
great. So uh Platemate is a project that I actually worked on in college way
1:30
before any of the genetive AI would have made it actually good or easy. But the
1:34
basic concept, you could say it was my first entrepreneurial idea, but I just
1:37
did it in in college as a project, was what if you could just take a picture of
1:41
your food, your lunch, your dinner, whatever, and have an app analyze it and
1:46
tell you you had this much protein, this many carbs, this many calories, and add
1:49
it all up instead of people having to like count this up for themselves or
1:52
whatever it might be. Um, great concept, but I would say before its time in terms
1:56
of technology, but it was a neat one to try to build. At the time, AI just could
2:01
not solve this problem. We were so far, this was probably like gosh 15 years ago
2:04
now. AI was just not there to be able to do this. And so we actually used human
2:08
intelligence. We used Amazon Mechanical Turk. Uh if listeners don't know what
2:12
Mechanical Turk is, is this platform where you could basically pay people by
2:16
the minute, maybe even by the second to complete tasks for you online for like
2:20
uh I don't know, probably below minimum wage. Um and so you would just put tasks
2:24
up there and the task would be like label a draw a box around all the foods
2:28
you see on this plate. And another person's task would be identify what's
2:32
in box number three. And they'd be like, uh, I don't know, I think that's mashed
2:34
potatoes or something. And then person number three's job would be uh look at
2:38
the mashed potatoes and say, okay, match this to this USDA food category so that
2:43
we can actually find out how many calories or carbs are in there. And then
2:46
our app would stitch it all together. Um, it was a cool concept. Uh, this was
2:50
when the wisdom of crowds was sort of like this big thing. Um, I had like
2:55
dreams of building a mobile app. Uh, unfortunately, this never got beyond
2:58
research project because it never actually got to be cost feasible at the
3:02
time. I think we came to the point where it would cost maybe, you know, like a
3:06
buck per meal to have someone like label all your food for you. So, you're
3:10
talking three meals a day, 30 days a month, like $100 a month subscription
3:13
just to break even on somebody telling you what's in your food. I think there
3:16
could be some medical applications for it or whatever, but hard to see it
3:20
really taking off as a business. So, unfortunately, had to put it back on the
3:23
shelf. Uh, but I'm psyched to see that now there are companies doing this with
3:26
AI at way lower cost. Yeah. I mean, these these products
3:28
exist. Like if you were to build it today, like how how quick do you think
3:32
you could build a version of this today compared to
3:34
Oh my god, you could b code it so fast. I really think a weekend there's
3:37
actually no magic there now. It's like the vision models just do it. Um, I
3:41
think, you know, the hard part is like tuning the accuracy and making sure it's
3:44
right. Um, and getting people to trust it.
3:46
Um, I also had this whole other idea of like what if it was a social network?
3:50
What if what if we all posted our food pictures, which people have been doing
3:54
since time memorial on Instagram and Twitter or whatever, but then I would I
3:58
would label what's in your lunch and you would label what's in mine like for fun.
4:01
I think I still think there's an idea there. I never figured out the formula
4:04
for it, but yeah, just the version where AI does it. If I code it in an
4:07
afternoon, it'd be fun. Yeah, I've I've heard it's still
4:09
actually pretty tricky because it's hard to see like what are the ingredients in
4:13
this? It's like you need to have like I don't know you have to have like a
4:15
quarter there for like you know to to like scale for scale and then uh yeah
4:21
you just like don't know how much butter did the person actually put in this
4:23
thing or not like it's hard to tell but you know it's it's all about what
4:27
your baseline is. So the thing we did in that paper that I thought was really
4:30
interesting was we actually compared it to professional nutritionists um and
4:32
also to self-reporting of people on their own eating. And it turns out just
4:37
averaging together a bunch of these mechanical Turk people did at least as
4:40
well as the professional nutritionists and often better than people
4:43
self-reporting the nutritionists because it was just one person working off the
4:47
same limited information. Whereas if you averaged five or 10 strangers, you
4:50
actually could get a more calibrated average and better than self-reporting
4:54
because it turns out even when we know how much butter we put in, we lie to
4:56
ourselves about it. And so an impartial stranger without the information often
5:00
does better than you yourself with the information. And so considering that, it
5:04
might still work. Yeah, I've heard that like when people ask like why is
5:08
restaurant like you know food so much better than the homemade one and it's
5:11
like well you're you're paying me the cook
5:15
so you don't have to see how much butter I actually put in this thing.
5:18
I love that. Yeah. Please don't tell me. I couldn't possibly know. Then you make
5:22
cookies at home and you're like, "Dear Lord, that's how you do this?"
5:25
It's like, "Yeah." So that's that's that's why it tastes so good. Um I love
5:30
it. Uh what do you like what do you think you learned about building
5:34
products from platemate? You know since that time you've had a pretty like uh uh
5:39
e epic career across you know Microsoft and optimizely and uh you know now
5:45
Google like have worked on some pretty impressive um products like what what
5:49
groundwork did you sort of uh pick up on working on platemate?
5:53
Gosh that's a good question. I mean the first thing is like work on a universal
5:57
problem that lots of people have. Uh I think all of us struggle with eating and
6:02
want to eat better and want to make it better and I think I was drawn from the
6:04
very beginning to solve a problem that millions and millions of people have um
6:09
uh do something that is like meaningful to their daily life. Um it's something
6:13
that has actually not always been a thread through my career. So at my last
6:17
job uh before gamma was optimizely uh building you know like AB testing first
6:21
for marketers then for developers and that company did well in so many ways
6:25
but one thing I struggled with was TAM the total addressable market. I think uh
6:29
it was actually a victim of its own success. We just got everybody that
6:32
wanted to do marketing AB testing to use us and ran out of people and had to
6:36
think about gosh, what's our next act? What's our next thing? Um uh and so I
6:41
sort of told myself after that, if I start a company, I want to work on
6:43
something with an enormous addressable market where it all just comes down to
6:46
how well can I build the product. But if I build the product really well, it can
6:50
succeed because there's just so many people out there that have this problem.
6:53
So I'm sure we'll talk more about Gamma, but that's what led us to slide decks as
6:56
like another universal. The other thing I learned was I mean this was sort of a
6:59
proto AI product. In this case, it wasn't actually artificial intelligence,
7:02
but it you could say it was imperfect intelligence. It was people who were not
7:05
that motivated with very limited context uh being swapped in and out uh much like
7:10
working with LLMs. And so it was kind of an early preview of some of the
7:13
opportunities and challenges of building products that have intelligence inside
7:16
them. And one thing I learned a lot is that you have to think about how to
7:20
break down a task into its individual pieces and how to measure success of
7:24
each of those tasks. And that is something that's I think served us very
7:26
well at Gamma. Yeah, I mean that that right there like
7:29
we'll we'll probably get into later, but that idea of breaking down task is so
7:33
valuable right now in working with AI. Um but shifting to the start of gamma,
7:39
you pick, you know, slide decks as like, hey, big TAM, right? But gamma is still
7:45
pre this era of generative AI. You're not thinking of gamma with AI as a a
7:51
concept on on the V1. Correct. Correct. Yeah. We started the company in
7:55
2020 when AI was just a twinkle in all of our eye, but it couldn't do much yet
7:58
much like in my blade experience. Got it. And so like what was the
8:04
thinking behind the the sort of first version of of gamma?
8:08
So we started the company in 2020. This is after co had already hit and
8:13
everybody's life is disrupted. We're all like working from home and working from
8:16
like weird configurations of home. like your laptop's on top of the washing
8:20
machine and you're trying to do your your meeting while your spouse is like
8:22
in the other room like with no pants on in on their call. It was just a a weird
8:26
time to even remember of that's that's what life was like. Everything was
8:29
disrupted. But I think the key idea that we had then was because life is
8:33
disrupted this window is open to change how we work. And in particular there's
8:37
this sort of like universal business practice of presentations that uh
8:42
everybody has to do in their jobs and nobody likes doing. Uh if you ask a
8:45
hundred people like do you like PowerPoint? people who've made a a
8:49
presentation, you're not going to get overall very positive responses because,
8:52
you know, they're so sort of performative. They're so frustrating and
8:55
finicky. There's just all these things that are frustrating and yet it's sort
8:58
of a coordination problem. It's hard to move off of PowerPoint when uh that's
9:01
just how your company works and how your business works. And so we thought code
9:05
represented this window of opportunity and in particular like remote and async
9:09
work represent an opportunity to shift the conversation. And so the first
9:13
version of Gamma was actually all about uh PowerPoint for distributed work. So
9:19
uh what if you could make a version of presentations that didn't require
9:23
everybody to actually be in the same room at the same time to consume it. So
9:27
uh the idea we kind of pursued was kind of a hybrid document deck and website.
9:33
Uh so the document aspect is we wanted something that was complete. it would
9:36
actually take all the information, be something you could send around ahead of
9:39
a meeting or instead of a meeting. Uh, but the presentation part is we wanted
9:43
something that was very visually compelling, something that distilled
9:46
ideas into their key points, used a lot of visuals and imagery, and just looked
9:50
professional and polished because it it showed that you'd put a lot of like care
9:54
and thought into whatever you were doing. And then the website aspect was,
9:58
you know, if we're going to reinvent this, what if we brought so many of
10:00
these innovations of web technology the last, I don't know, 10, 20 years. So uh
10:05
interactivity being a basic one like this doesn't have to be a linear story
10:08
you can control the direction of the narrative responsiveness. So like if you
10:13
are taking this uh presentation from a park bench outside your house on your
10:17
phone you shouldn't have to like do this zoom in and enhance thing. You should be
10:20
able to just look at your phone and have it adapt to it the same way websites do.
10:24
So we kind of created this like interesting hybrid format that was a bit
10:27
of all these things. Quite hard to explain to a user to be honest. But I
10:30
think the concept was solid of reinvent the presentation as a medium for this
10:34
new era. So you do this at pretty close to the
10:39
beginning of co you're what I guess uh end of 2020 you you all start to you
10:44
found the company you build the first version of it
10:47
how' the first version work like what what was the success? uh
10:52
middling, I would say. So, what's what's middling like? Like like
10:56
how like Yeah. Like uh yeah. So, you know, I would say the
10:59
the the peak of this first version of the product was probably like maybe two
11:03
years later. We had built up through a couple stages of beta. We launched on
11:07
product hunt. It went really well. I think we won like product of the month
11:09
on Product Hunt. We were patting ourselves in the back. We had maybe uh
11:14
tens of thousands of signups and maybe like a thousand monthly active users or
11:19
something like that. So we had real people using the product um organically
11:24
and they liked it and gave us positive feedback but you know like no more than
11:29
a thousand of them and it wasn't growing exponentially. It was growing linearly
11:34
and it felt like the people who got it really got it and loved it but it was
11:38
very hard to take a new person who didn't get it and bring them through an
11:41
onboarding flow um or actually have them bring themselves through an onboarding
11:45
flow and get to an aha moment on the other side. the ratio of that happening
11:48
was just extremely low because it was like this whole new thing I have to
11:51
learn and like what's the point of it and I can't really tell what it can do
11:55
until I invest an hour into actually making something in the platform which
11:58
most people don't have the patience for. Now I think I heard or read somewhere
12:02
that you were having like a fair amount of like existential angst around this.
12:06
Now is this related to hey you know sort of this middling success that gamma is
12:09
having at the time or is this more you see chatbt launch and you start to
12:14
realize like ah there's maybe a different way to do this like what
12:16
what's the mindset like for you kind of in and around this moment in time
12:20
the exential angst came before the chat GPT launch um which was a good thing
12:25
because it meant that by the time chatt came out we had already played out our
12:29
exential crisis a little bit um the exential angst actually came from
12:33
the economy it came from the fact that we had started in like the best
12:36
fundraising market ever. And it was super easy to be a startup because
12:40
everybody was like, "Oh, this thing could be the next Zoom. Here's millions
12:43
of dollars, no questions asked, basically." And suddenly we shifted from
12:46
that to high inflation, high interest rates, banks collapsing. And it was very
12:51
clear that uh kind of no matter how good we did, nobody was going to give us a
12:54
series A like the Gulf. So you basically are just like, "Hey,
12:58
this company is not going to work." And so that's where you're like this this
13:02
angst and stress is all sort of just coming from. Hey, you've you've made a
13:07
product that's decent, but it's not it's not a runaway success.
13:11
Exactly. And and I think the biggest feeling was the bar has been raised.
13:14
It's not enough to have traction. You have to be a runaway success. So then
13:17
it's like, well, what can be a runaway success? That's a very hard question to
13:20
answer. Um uh what can really be 10x better than what came before? But at the
13:25
time we're having this exential crisis, AI is just starting to sort of bloom out
13:29
there. Nobody really realizes that Chacht is coming. But uh the first thing
13:34
for us that we see is image models coming out like stable diffusion and
13:37
Dolly and they're making some kind of cool stuff. Still goofy and uh silly,
13:41
but there's a I think what's interesting is that it's goofy in a way that goes
13:44
viral because everyone can see like oh this is like kind of almost good. And
13:49
you know, for us working on this very visual medium of presentations, it's the
13:52
image generators that catch our eye. And we're like, "Oh my gosh, imagine if AI
13:56
could make all the clip art in your presentation. Imagine if you could have
13:59
these bespoke illustrations for every slide." Like that does feel like a
14:03
gamecher. And that like feels maybe viral in a way that what we've built so
14:06
far is not viral. So that's kind of where we first clued into it. And then
14:10
that led us to then looking at LLM progress. Uh checking back in on GPT3,
14:15
which is what it was still called at the time. And I remember logging back into
14:18
my OpenAI account that I had first made like a couple years earlier and just
14:21
trying GPT3 again and noticing like oh this actually got good in the past like
14:26
two years or so. It turns out OpenAI had been making a lot of steady progress
14:30
towards chat GPT before chat GPT launched. Most people just hadn't
14:33
noticed yet and luckily we had enough existential angst that we had and we
14:36
started playing with it and you know we just realized like oh this thing can
14:40
make like a presentation outline and also it's not like making a presentation
14:44
is rocket science. Like it turns out even a pretty dumb AI can like write
14:47
words for a slide. Um do you think that had the economy not
14:53
tanked, had the company been maybe even more moderately moderately successful
14:59
that you all would have been picking up on these trends at that moment in time?
15:04
Obviously on the trends, yes, but betting on them decisively, no. I think
15:08
we would have dabbled in them as a side prototype with you know 5% of our effort
15:14
to like see what could happen. But my guess of what would have happened
15:17
if we hadn't felt the existential angst is we wouldn't have committed to it.
15:20
Somebody else would have and then they would have just won the race that ensued
15:25
after that. Yeah. Do you like curious like what
15:29
advice you have? Like I see a lot of folks even now that have built like
15:34
solid companies um that are more dabbling on the AI side uh of the house
15:40
and that always just like surprises me when I sort of run into that like h how
15:45
would you give guidance to that? Someone who sort of built a company pre you know
15:49
this AI wave and is like trying to figure out like you know should we
15:52
should we not they're kind of sitting on that fence there. Like how do you think
15:55
about that? Gosh, I mean my first reaction is
15:59
dabbling ain't going to cut it. Like we are in the middle of this absolutely
16:02
transformative platform shift. The best analog I can think of is uh when mobile
16:06
took off and there were companies that sort of like dabbled in mobile and there
16:11
were companies that bet on it and the companies that really bet on mobile just
16:15
saw this decisive new market opening for them where they could take over in a new
16:20
place. The ones who dabbled kind of didn't get that far. Um, and so I think
16:24
if you're in a position to do it, you should bet more aggressively. But all of
16:28
that said, I also see companies really forcing AI where it may just not make
16:32
sense for what their product does. Um, and everybody's got like the sparkle
16:36
icon in their product that just opens like a chat GPT wrapper and does
16:39
something kind of unhelpful. So I don't know if betting on AI for the sake of it
16:43
is right. I think the key thing people have to do is think about if this is a
16:47
platform shift, what new market opportunity is open that wasn't possible
16:50
before that we're in a position to win. Um, and hopefully you can find something
16:55
that is adjacent enough and builds enough on what you're doing that you're
16:59
the right one to seize it. If there's not a natural fit, it may not be worth
17:02
trying to force it though. Yeah. So, you all
17:07
you you ship the the sort of AI version of Gamma in what, three months,
17:11
something like that, I think is what I saw. Yeah.
17:14
Um, walk me through what happens next. You do this three month sh uh uh sprint
17:19
and then you launch gamma I don't know 2.0 I don't know what you call it. Uh
17:25
what what happens? Everything kind of goes crazy in a good
17:28
way. So we we we launch the product. We sort of realize it's a bet the company
17:34
moment. We have sort of dwindling runway. Um we don't really have a plan B
17:39
if I'm being honest uh when we launch it. Uh but as we work on it more I think
17:43
we build our own conviction. Uh even before we launch it I remember having
17:48
certain moments just in myself where I'm like oh I think I really believe in this
17:51
thing. And one of the moments I remember is uh we just started uh wanting to play
17:57
with gamma ourselves to make things whenever a question came up. So the
18:01
example I remember was sitting around at lunch with my teammates and we're
18:04
talking about why does the bread from Subway smell so good? Like I feel like
18:08
when you when you go by a subway, they're like piping out this smell and
18:11
it's just got this like very very powerful smell like what are they doing
18:14
there? And that's a question that in the past I might have googled uh to look up.
18:18
But instead I just typed it into Gamma and I said make me a presentation about
18:22
Subways bread. And I got like a 10 slide presentation with like you know visuals
18:26
and content and everything about the Subway bread smell and I just presented
18:29
on the fly off my phone to my teammates at lunch and it was just fun. It was uh
18:34
we were just enjoying ourselves in a way that I had not before. And I I think
18:38
that that spirit of fun can be such a northstar, especially for a like
18:42
consumer proumer product uh where you're hoping to create some kind of like
18:46
virality and change of behavior. We just had those glimmers. Um they weren't
18:49
solid evidence yet. When we actually launched uh we, you know, we bet
18:54
everything on this big launch. We made a launch video. We we crafted a really
18:58
clickbaity tweet. We did all these things to try to get attention. Um, and
19:02
actually the initial launch day didn't pop the way we hoped for. It was okay.
19:06
Uh, but it wasn't probably as much as we we were looking for. But the moment when
19:10
I started to realize things were going to work is even though our first launch
19:14
day didn't pop, what was different from previous launches we've done was that
19:17
day two was better than day one and then day three was better than day two. And
19:22
we started to see just this steady increase in people coming into our
19:25
product where more and more were signing up even as our launch marketing kind of
19:29
subsided and stopped being the main conversation. There were just more
19:32
people coming in every day. And then in a period of weeks it started to get
19:36
really weird because the number just kept going up and the number started to
19:39
seem kind of ridiculous to me. You know, we we'd been averaging a couple hundred
19:44
signups a day most days before this launch. Suddenly, we were getting like
19:48
5,000 becoming 10,000 becoming 15,000 signups just in a day. And we're like,
19:54
where are all these people coming from? And it turns out they were coming from
19:57
all over the world. I found that out because we had an intercom chatbot on
20:01
our site. And I was the support agent for the intercom chatbot. And it used to
20:05
be something I did in five minutes a day and it became 30 minutes a day and then
20:09
an hour a day of just responding to intercom messages. And suddenly a lot of
20:13
the intercom messages were not in English. They were in like Chinese and
20:17
Portuguese and like Hindi, which I do not speak any of those languages. So I
20:22
was spending a lot of my day literally just copy pasting messages into Google
20:26
Translate, trying to respond to them in Google Translate and pasting them back.
20:30
Um, it turns out they're automations for this, but it took me a while to realize
20:33
that and I wasted a lot of hours doing this. Um, and then as the weeks went by,
20:38
there's just more and more of these. We started to have to have like six people
20:41
in our team in a war room just responding to intercoms. And also
20:45
eventually the tone of the intercom messages changed from like why are you
20:48
missing this one feature? Why can't I export this way to how do I pay you? Uh
20:52
there were all these people that are like my AI credits ran out. How do I
20:56
pay? Why can't I pay? Where's the pay button? And we're like oh my god why
21:00
don't we have a pay button? this is like this is and the reason we didn't have a
21:03
pay button was that it never occurred to us that there was like enough of a
21:08
business here to support monetization. You know, we were just trying to make
21:10
something people wanted. And very quickly, like yes, they wanted it and
21:13
they were willing to pay. And so we had to throw all of our energy into building
21:17
monetization, building a Stripe integration, uh figuring out what our
21:20
pricing should be and getting it all running.
21:22
Yeah, I I I remember those days of uh working, you know, in when we started uh
21:28
it would have been OAR chat. So this that that definitely dates the company.
21:32
Um but uh yeah, I remember you know waking
21:36
up and doing it for like you know 30 minutes or now but then eventually it
21:40
just become like oh my god it's 3 p.m. and we're like still
21:43
we're still just responding to the same things over and over.
21:45
Yeah. And it's such a um like if you've ever sort of like worked on a project
21:49
that doesn't work and gone to one that does work, you you start to realize like
21:53
how obvious product market fit is. Like if you're sort of asking yourself, do
21:57
you have it or do you not? It's like you don't got it. like when you got it, you
22:00
you have all these other problems that you're just like, I don't have any time
22:03
to sort of like reflect on that. I just got to figure out where the payment
22:06
button is, how to translate this conversation. Like all that sort of
22:09
stuff is going on around you. Um
22:12
funny people say that. I've been a product
22:15
manager for my whole career and I always knew people say things like, "Oh, you'll
22:20
know it when you feel it." But still, you know, much early in the journey, I
22:22
think I was kidding myself of like, "Oh, yeah, there's there's product market fit
22:25
here." Like, because I I think I hadn't truly felt it. And it was this moment I
22:30
think I talked about with support where I really felt it. And the best metaphor
22:32
I heard for product market fit is it's the moment you go from pushing the rock
22:36
up the hill to chasing the rock down the hill. And I just felt like I was chasing
22:40
the rock so much those following months. Yeah. So let let's shift the
22:46
conversation a bit. So you you're now not only growing a product, you're also
22:51
having to grow a company. Like at this stage you're pretty clearly a company.
22:55
Um, I'm curious, you have like this distinct advantage of coming of age when
23:03
all these AI tools are are are out and about, which is very different than say
23:07
like how you all were running Optimizely like talk to me about like what are you
23:13
doing differently in terms of how you're scaling the company? how you think
23:18
about, you know, headcount and tools and automation and all these other
23:22
capabilities that maybe, you know, companies who are, you know, sort of pre
23:26
this AI era haven't grown up with it and are having to transform their
23:30
organizations. It's absolutely right. And, you know,
23:34
not only have we grown up in this AI era where you can automate things, but we
23:38
also had a near-death experience as a company of almost running out of money
23:41
from from being probably too big when we should have been leaner. And so I think
23:46
that that got really seared into the DNA of how we operate. And so you know we
23:50
were I think uh 12 people when we hit product market fit. And actually for
23:55
those first six months after we hit it we didn't hire anybody. We just tried to
23:59
see what we could do with the people we had. Looking back I actually think we
24:02
probably should have hired some of those roles sooner. But we were so coming out
24:05
of this paranoid mode that we just couldn't even think that way. And it was
24:08
like what can we do with people we have? And so yeah we we immediately sort of
24:12
like looked for ways to automate things. I think at this time AI coding hadn't
24:17
really taken off yet. We were maybe just using GitHub copilot yet. Uh now we use
24:20
a ton of it, but at the time it wasn't ready yet. Support was one of the first
24:24
areas where we did this where we immediately tried to like automate our
24:27
support. The first one being actually this translation problem. It turns out
24:30
there's just tools that can translate your messages back and forth for you.
24:33
And just by doing that, I mean it sounds silly, but in a previous era of a
24:36
company, and this was true for us at Optimizely, you would have had to hire a
24:40
support person who spoke every single language of your users. And so if you
24:43
wanted to expand to France, you had to hire like a French-speaking support
24:47
person. For us, we just never did that. We have we have an AI translator bot and
24:51
then we had people in the US that spoke English and the bot would just translate
24:55
between them and a hundred different countries and so we just went
24:58
international. U and the same came true when we actually localized our product
25:02
as well. Um we have done this in like a bunch of different places. Uh we do this
25:06
in our marketing, we do this in our well now we do it in our engineering uh a
25:10
huge amount. We do it with our design team. So our design team is constantly
25:13
prototyping things in code rather than just in Figma mockups trying to get as
25:18
far as he can trying to make prototypes they can like show to users and talk to
25:21
people about. Uh I think it's become critical to how we operate.
25:26
What tell me about like are the type of is there a type of like job or person
25:31
that is that is different here? Like when you hire this person is it more of
25:34
a generalist? Do they have like you know more like tech skills? like what what's
25:40
different about like some of these early employees you're bringing on compared to
25:43
just like to your point, oh we need a French-speaking support person?
25:47
Yeah, I would say it's much less specialized and I think we have really
25:50
gravitated to people who are comfortable operating at the margins between
25:54
multiple roles. So uh to go back to our UX design team, I think we are pretty
25:59
non-traditional in the sense that uh all almost all of our UX designers uh write
26:05
code and all of them do their own research. They don't delegate that to
26:09
some other person who like talks to the users for them. Um they're all sort of
26:13
like making prototypes. Um many of them probably could have been engineers if
26:17
they wanted to or still could be, but they actually happen to like the craft
26:20
of like figuring out what users need. Um we we waited a long time to actually
26:24
hire product managers for our company. And now that we have I would say all of
26:28
our product managers are also very multi-talented. Like our our growth
26:31
product manager is also the analyst. He doesn't like delegate to someone else to
26:35
like uh could you write the SQL queries for me? like no no he he has that skill
26:38
and brought it and that was a requirement whoever we brought into that
26:41
role. Um our marketing team is like omnialented and is I think in many cases
26:46
just using AI to multiply their skill sets. For example, when it comes to
26:49
visual design we did a big rebrand a few months ago and there too we could have
26:53
hired a team of you know like uh illustrators and like visual people just
26:58
to make everything. But instead we just used a huge amount of midjourney. We
27:02
really really doubled down on it and we made learning how to use a tool like
27:06
midjourney like a key skill that we developed so that we could scale the
27:11
really creative people that we have to work across multiple media.
27:15
How do you So yeah that that totally makes sense. Now you also said in
27:20
hindsight maybe you should have hired some more folks. Um
27:25
but you you know you have this sort of like baggage from you know uh almost
27:29
dying and like h you know people cost money. I don't want to I don't want to
27:32
run out of money again. Like how do you make that decision of when is the right
27:37
time to go bring a human on versus like let's just use more midjourney.
27:41
You know, we we had this framework that we created for ourselves. I think after
27:45
maybe this near-death experience of hire painfully slow. I don't think we
27:48
invented. I'm sure we stole it from somebody. But it was like a useful
27:50
exercise of like yes, you want to hire, but you should wait till it hurts. Wait
27:54
till you feel the pinch. and only then do we solve a problem through hiring
27:57
that couldn't be solved through, you know, automation or like thinking
28:01
smarter about our jobs. We've actually had to revise that one over time. We
28:04
realized that that actually wasn't the right framework either because I think
28:07
what we realized was that we were optimizing for our own pain and not our
28:11
users pain. So, we were waiting until something was painful for us, but we
28:14
weren't necessarily thinking about what if this is painful for a million people
28:18
using our product. Uh, and this really comes down to things like engineering.
28:20
Like, uh, honestly, it's fine for us to have fewer engineers. It means there's
28:24
fewer meetings, fewer discussions, fewer everything. But it turns out it's not
28:27
fine for our users when our like PowerPoint export is broken for the
28:31
third day in a row that we decided to hire painfully slow on engineering. Like
28:34
they needed us to do that a lot earlier. And so we've had to constantly
28:38
recalibrate this in our head. Um I think for us one lens that we're kind of using
28:43
for this is the metric of AR per employee. um which I think this is maybe
28:49
a number that just illustrates how different the AI company era is from any
28:53
previous era. So when I was at Optimizely I believe this metric for us
28:58
was approximately like $100,000 of ARR per employee and we had this big goal of
29:04
like let's get to $200,000 because that's like top cortile for a SAS
29:08
company. Uh at Gamma we're over a million dollars of AR per employee. Um,
29:13
I think we announced recently that we passed 50 million AR and with fewer than
29:17
40 employees. So that gives you some idea of like what that number looks
29:20
like. And the thing in my mind is like, wow, it almost feels irresponsible to
29:24
let that go above 2 million. Like we should probably be hiring at at least a
29:27
pace that there's someone to support all these things. Like we know there's good
29:30
roles, so maybe we need the kick in the butt to ourselves to hire a little bit
29:33
more. But we're constantly kind of monitoring that one is just one way of
29:36
saying, are we a lean team that's still delivering for our customers?
29:39
Yeah, I I love that. like uh we had a similar mantra uh don't hire till it
29:45
hurts and I like that twist on it which is the like who's hurting like is it you
29:51
or is it your customers um you know I I I still distinctly remember when we
29:56
launched uh workflow multi-steps apps and uh overnight it basically we had
30:02
that same experience where it was like you know uh there's this many thousands
30:06
of people signing up a day and then the next day it was more and the next day it
30:09
was more after that which is like very unusual
30:11
Um, usually it sort of like peaks and then goes back to sort of like a more
30:15
normal plateau. And I remember thinking, oh crap, like our support team is like
30:19
half as big as it needs to be. We got to we got to figure this out yesterday. Um,
30:24
because now uh we were like we're doing double duty. We're doing our day jobs
30:28
and we're doing support all night just to like keep up with all this stuff. Uh,
30:31
and that was kind of our first wakeup call to the like, oh, maybe maybe there
30:36
is a better way to tune this don't tire till it hurts um mantra.
30:42
Um you mentioned that your designers write code. I I want to
30:48
come back to that. Um when you say your designers write code, are they is this
30:53
like vibe code or are they coding? Uh you know it's a mix of both but often
30:59
real coding. I think this started with our head of design, Zach, who has just a
31:02
really cool background. He started his career uh in the Marines doing like
31:06
avionics and then he ended up as like a Apple store guy and then like a
31:12
WordPress developer and then a UX designer and so he's just someone who I
31:16
think has always just done it all and like done what's needed and that became
31:19
a big part of the culture of what design was at Gamma. We just didn't know any
31:23
other way. Like he he's a Swiss Army knife and so
31:25
if for a given idea the right way to build it was to do a Figma mockup that's
31:29
what he would do. But if it was like an LLM thing, it was just so clear that you
31:32
had to actually prototype it. You had to actually like let the user type in
31:36
whatever they wanted, have the LM respond, and have real data inside this
31:40
UI so you could actually know what would work realistically. Uh, and so I think
31:44
that set the tone. We didn't make it a requirement of design hiring that
31:47
everybody must be like a former engineer. That would have made it
31:50
really, really hard to find designers. Um, but we did want everyone to have the
31:54
curiosity. And it turns out there's a lot of people out there who another of
31:58
our designers, Nick, uh, you know, he just like makes mobile apps on the side.
32:01
He has like a camping trip packing app that he always thought was cool because
32:04
he was curious and he wanted to learn Swift and he came to Gamma and you know,
32:08
we set him up with like code sandbox and he just started coding more and more of
32:11
his prototypes and now I think most of his prototypes that he makes are coded.
32:15
Um, for other folks they have no engineering background and so they are
32:18
just vibe coding. they're just like taking their ideas and putting them
32:21
into, you know, lovable or bolt or something and they're just finding that
32:25
as another tool in their toolbox to express interactive ideas. Um, we even
32:29
have a design engineer on our team, uh, and her role is to basically help
32:33
designers bring their ideas to life. So, if they're not as strong with developing
32:37
or prototyping or some of the extra visual polish, she can actually help
32:40
them take those to the next level. Yeah. So, you're hiring these like
32:45
multi-disciplinary folks that are, you know, fairly fluent with AI. what what
32:49
lessons learned have you found uh or had trying to find these folks? Um you know
32:55
is there different approaches you take in the hiring pro process compared to
32:58
you know maybe in the past where you're looking for you know specialists in
33:02
particular areas. Uh you know passion is actually a hard
33:06
thing to fake. That's something that I've learned over time is that when you
33:09
just talk to people and say what are some things you're trying to learn? How
33:13
are you pushing yourself in your career? Uh why were you interested in applying
33:16
to us? It's just amazing the kinds of things people tell you and uh you can
33:20
tell when someone just has this sort of like overflowing passion and care for
33:24
their craft. Another thing that we do in our interview process just because of
33:27
what our product is is uh for most roles we make everyone make a presentation. So
33:32
they will actually make a presentation in gamma for like 30 minutes about
33:36
usually it's like an intro to themselves and it's also something about like a
33:40
project they've worked on that's relevant to us or a specific homework
33:42
assignment. And you can just tell when someone uses Gamma, which is itself an
33:46
AI product, to make something, uh, how much extra little craft they put in, how
33:52
much they discover all the weird little features of our product that are
33:55
sometimes buried or not even very good, but do they push them to their limits?
33:59
And the act of just like making that presentation really lets people show us
34:03
who they are. And I've been amazed by the correlation of the people that go
34:07
all out in their presentations and put way too much work into it who then show
34:10
up at work and go all out in their work and put way too much work into it. And
34:14
that is just such an amazing sign of passion and craft and hard work that
34:18
matters in the job. H how so I co-sign like that passion that
34:24
hard work etc. Um, how do you balance that with like, you know, experience,
34:30
practical skills? You know, I think of like the, you know, the intern who like,
34:35
you know, gives it everything they've got,
34:38
but everything they've got might not be that much.
34:42
We've skewed pretty senior in our hiring, I would say. um you know like
34:46
especially uh most of our early team like everybody on our team of the first
34:50
maybe like I don't know six or so people we hired were like at least maybe 15
34:57
years into their career or more so like quite experienced quite solid and I
35:01
think we're lucky that that helped us create a very solid foundation in terms
35:04
of like everything from like architecture to design systems to
35:08
cultural practices um and because we've now built that sturdy foundation we've
35:13
been able to start to hire are more like younger people earlier in their career.
35:17
We still are mostly not hiring people directly out of college. Uh
35:19
unfortunately, we're more going after folks that have three or four years of
35:23
experience, usually at a bigger company, who are now ready for that uh new
35:26
challenge. But I think when those people come in, they're now stepping into a
35:30
solid system. And they bring this tremendous energy and like new ideas of
35:34
like, wow, what if we choose Claude for everything? I think if you've been in
35:36
your career for 15 years, it's still hard to even internalize that idea. Um
35:42
uh and now increasingly we're also just bringing in AIs to do things which also
35:46
I think like AI coding tools and AI tools work so much better when there's a
35:50
solid system in place built by experienced people for them to start
35:53
from. What have you learned from one of those
35:55
folks that come in and are just like ah let's just use cloud for anything like
35:58
I'm curious how that's like changed your own point of view and your own way of
36:02
working. I've realized that I am a dinosaur. It's
36:06
like the feeling I have watching people use these things is the same feeling I
36:11
had maybe a couple years ago when Snapchat was like first big and I was
36:14
like I think Snapchat was the first moment I had where I'm like oh I'm maybe
36:18
too old to like understand this technology like I don't get it and and
36:21
that it turns out is just is just life as you get older. This just happens with
36:25
increasing pace with so many new products. And so the biggest thing I
36:29
realized was holding me back was I'm not even asking the right questions of AI.
36:32
It's not like I'm a bad prompter or I don't know how to integrate the tools.
36:36
It's that there are things I just assume in my job are not automatable and I'm
36:41
like totally wrong. Like every time that I'm writing like a SQL query, I'm
36:46
realizing now like, oh, I actually should have let AI do that. Like there's
36:49
just so much finicky stuff that I'm doing that I should have let someone
36:52
else do. Um uh so many things that I write I'm like, oh man, maybe this could
36:57
have been an AI. And so I think you have to work around people who just reach for
37:01
those things because it's their instinct and very quickly you realize they run
37:05
circles around you. Say say more about that. So you've got
37:09
like the the example of the SQL query. What like what are the other areas where
37:12
you're like, "Oh, I'm not reflexively reaching for this thing and I should
37:15
be." Well, let's I'm just going to dig into
37:17
the SQL query one because it's such a good example. I think intellectually I
37:20
know that I can go into chat GPT and I can say write me a SQL query that does
37:24
XYZ. But as soon as I say that out loud, I just have all these ideas of what's
37:27
going to go wrong. Like, uh, I'm probably not going to bother with that
37:30
because it doesn't really know our database schema. It doesn't really know
37:33
like where the bodies are buried in the data. I should really just do it myself.
37:37
Uh, but, you know, then we hired uh, this new guy who's like really into
37:40
using AI with SQL and he just plugged uh, Snowflake directly into cloud code
37:45
and he just has it run these things in a loop. He says, you know, like find the
37:49
promo code that's causing problems uh in like Indonesia. And it turns out Claude
37:54
can just run itself in a loop, run the command over and over again, fix its
37:57
query, get there, and for like 28 cents and in about like 2 minutes, it has not
38:02
only like solved the entire problem, but created a reusable framework we can use
38:06
to solve this problem again in the future.
38:08
And so like seeing that, I'm like, wow, I I need to totally step outside my
38:12
understanding of how to get work done. Actually,
38:14
I I mean, I feel the same way. I was talking to somebody yesterday who on the
38:18
same like SQL discussion, they had taken their entire schema and they'd added it
38:24
to a cloud project and they were like, "Hey, we're going to outline it all in a
38:27
very specific way." And now they just have this project they talk to to write
38:29
all this stuff. And it's just it it's one of those things that once you hear
38:32
it, it's so obvious. You're like, "Oh, of of course that's how this should be
38:35
be done." But to your point, like if you've got a decade plus of working one
38:39
way, it's you don't you don't instinctually go there. Yeah.
38:44
And to go back to your earlier question, uh, constantly had this question of
38:48
like, should we hire a data analyst for the team? Like, it sure seems like it'd
38:51
be useful if anytime anybody has a question, there's someone who can answer
38:55
it. But I've also always had doubts about that. I think there's this weird
38:58
disintermediation that happens when the person asking the question doesn't
39:01
actually know how the query is written and and the person who's writing the
39:04
query doesn't actually have the business context. Like things can get lost along
39:07
the way. And so, I've kind of like resisted hiring it. And who knows, we
39:11
may still hire it. But now I'm like, maybe we don't need a data analyst. We
39:14
just need someone to maintain the cloud project that is like the mega data
39:18
analyst and has the schema in it. And we just need to train every person on the
39:22
team how to use cloud code to do these things. We might get way better results.
39:27
Yeah. And I mean, you still need like good like statistical foundation. You
39:30
don't want people making like errors like that. So that's where it does get
39:33
like kind of kind of fuzzy where it's like you still benefit a lot from
39:36
understanding these concepts. Uh and being versed in that
39:42
but the like act of actually you know artisal sequel is like not exactly a
39:49
thing anymore. God I love that term artisal. Yeah. So
39:52
it's so true. Um so I want to shift back um to the
39:59
product gamma. One of the things that you're doing there it that you know I
40:03
think we try and do a lot at Zapier as well is how do you make this
40:09
thing that everybody like theoretically should do like just that much easier,
40:15
that much simpler, you know? I I think to like all the folks building like
40:18
PowerPoints back in the day that would like whisbang PowerPoints and then I'd
40:22
show up and try and do it. I'm like I just suck at this. Like I just don't I
40:25
have no enjoyment. I get no enjoyment out of it and I'm bad. And now I'm sure
40:29
if I practiced and tried and put effort like sure I could get there but just
40:33
never felt worth it. Um and I think in a lot of ways what you've done at Gamma is
40:38
you've brought those tools closer to a person like me where I'm like oh maybe
40:43
I'm not as bad as I'm thought. I still may not be as good as you know uh Jane
40:48
or Jill over there but I'm like passible uh in a way. And so like what are the
40:53
what are you learning about how you build tools and products that sort of
40:56
close that uh skill gap? You know, one funny thing it makes me
41:01
think of is you do have to think about your your target customer and for us our
41:04
target customer was never a specific like job function or industry. It was
41:09
always more of a like uh skill level and it was someone who sounded like you who
41:13
said you know I'm not a designer but I have good ideas. uh I I know I have
41:18
something important to say, but I get lost in all the formatting. And the more
41:21
you can actually like zero in on a description of that person's pain points
41:24
and really align on that across the company and make sure everyone knows
41:28
that's who you're building for, it's actually immensely clarifying and it
41:31
lets you really build a lot for it. And so for us, we've always thought about
41:34
like, well, okay, what are the tedious parts of this that everyone hates? We
41:37
kind of had this north star at Gamma of this quote we've heard over and over
41:41
from people where people say, "I'm making this presentation and I realize
41:44
that I spent 90% of the time on the formatting and 10% of the time on the
41:48
content." What would it feel like to invert that? To actually spend 90% of
41:52
your time on the content and still 10% on making it look nice and making some
41:55
choices. But if you really hold that up as a product principle, it's like
41:58
immensely clarifying. And you know for us one of the things that we've held
42:02
true in gamma actually from the very start even preai is this idea that like
42:06
constraints are really helpful. And so unlike a lot of you know traditional
42:10
tools like PowerPoint we are not a full drag and drop design tool. And that's
42:14
actually not what we aspire to be. We we kind of have this framework that making
42:18
a presentation should feel more like writing where you're just typing stuff
42:22
into boxes and making a couple choices along the way but you're not actually
42:26
rearranging and realigning things. And I think that creates a really nice
42:29
separation of ownership where the user can focus on what they want to say and
42:33
we slash the AI can focus on making it look nice. Um, and it also means that
42:38
the AI can really let it rip. We can really give the AI a lot of space to
42:42
grow. I think another sort of like mental model we've had to constantly
42:45
readapt to is how do we plan for a world where AI gets better and better. And
42:49
there's all these categories of things that AI can't do just yet. Uh, for
42:53
example, AI can't make a diagram to save its life. If you ask AI to draw you a
42:57
picture of a ven diagram, it's actually shockingly bad at it. Even though it can
43:01
do so much more, you know, it can solve like international math olympiad like
43:04
physics problems, but it can't draw in a diagram. But it'll get there. And so we
43:08
have to plan for a world where it'll get there. And we have to think about what's
43:10
the user interface where if we just play forward a world where AI is pretty good
43:15
at almost everything, what does the user actually want to own? What do they want
43:19
to be giving input on? Um, another really interesting design principle um
43:24
that we've come up with recently is this idea of abundance. Uh, so the thing that
43:28
AI unlocks is abundance, especially visual abundance. So when you're
43:33
designing a slide, it's not like you should have one choice of design. You
43:36
actually could have 100 choices of design and it's almost free to give you
43:39
a 100 choices. And so a lot of the design challenges are how do we plan for
43:43
a world where it's almost effortless to give you a hundred choices but still
43:47
make it easy for you to choose between 100 things without being paralyzed by
43:50
all that abundance. Yeah, that's fascinating. So what say
43:55
more about this? Like I I you know I think we think about this the same way.
43:58
I'm curious if you've had like any heruristic or tips like everyone is
44:02
trying to figure out like predicting what the next model can do, predicting
44:06
what those next capabilities are possible. So they're sort of designing
44:08
ahead of that where they're like ah you know when this next new model lands this
44:13
thing that we can barely do now actually just becomes amazing and but also at the
44:18
same time like not actually building a bunch of features that the next model is
44:21
just going to make totally obsolete like so how are you thinking about like do
44:25
you have any horistics for like assessing when and when you're in which
44:29
camp? You know, one kind of magic wand thought
44:32
exercise we do is if money was no object and you can outsource all this job to
44:36
another human, which parts would you outsource? Which parts would you keep
44:38
for yourself? Um, and you know, for presentations, there is precedent for
44:42
this. Like I don't have a professional presentation designer. Maybe you do. I
44:46
know like Steve Jobs did. So there's like there's these jobs that are like,
44:51
okay, really make this premium. And so a northstar we use a lot is,
44:55
you know, what would a world look like where everybody has a Steve Jobs level
44:58
presentation designer? available kind of at their beck and call and and we ask
45:02
ourselves well what do like the McKenzis and the Steve Jobs have in their
45:06
presentations that people don't have now and that's led to all these interesting
45:08
ideas like it turns out really premium presentations have a lot of like video
45:12
and animation in them there's like subtle motion throughout um they have a
45:16
lot of visual consistency where there's like a visual metaphor that's used
45:19
throughout or very like custom styling and so that's actually driving a lot of
45:23
our road map it's saying like well let's just assume that AI will be able to do
45:26
all these things um and like video is a good example video is not cost effective
45:30
right now. It's actually like high wateringly expensive to offer AI video
45:34
in a product like ours. You know, it's like for example, like I think a lot of
45:37
these models are like $5 for a 5-second video. Good luck monetizing that right
45:42
now. It's like my plate problem all over again. Uh but I think we're playing for
45:46
a world where the cost of that will just fall and fall over the period of say 2
45:49
or 3 years. And so let's just line ourselves up and build all the places in
45:53
our UI where generative video will just fit when it becomes good. and let's try
45:57
to do the the barest inkling of it in different places, which in our case is
46:01
we have an animate feature where you can take any image and animate it like kind
46:04
of an animated GIF. And that's just been like the first place where we're trying
46:07
that interaction out and seeing what it can do. Yeah, I I really like that
46:12
framing of think about what you know like the the people with infinite
46:17
resources have the Steve Jobses of the world and what role do they actually
46:23
want in making a presentation or in do in anything like they're going to
46:26
outsource big chunks of it but clearly there's parts of it that they aren't and
46:32
it's like okay that's the surface area to go build for and then AI we're going
46:36
to have them take care of this ice like this iceberg underneath of all the tasks
46:40
that the human doesn't. It's a that's a very like human way to design a product.
46:44
Another version of it that I find interesting is to ask the question, what
46:48
would you do if you had infinite patience? Uh I'm realizing in all of my
46:52
LLM usage that that is like the single biggest edge that the models have over
46:56
humans is their patience. It's also their undoing sometimes. I think
46:59
sometimes they get stuck in loops because they're too patient. But when I
47:03
think about presentations, like what would I do if I had infinite patience?
47:06
Man, I would really spend the time to make sure all the boxes were perfectly
47:09
aligned. And if I was perfectly patient, I think I would like I would find the
47:14
perfect image for every slide. And I would really really put the effort into
47:17
making sure that they were all kind of colorcoordinated. But if I used like
47:20
pinks and blues, pink and blue just appeared in every image. And if I had
47:24
infinite patience, I would go pixel by pixel and just make sure that it all
47:27
felt right. And if I was infinitely patient, I might even make two or three
47:31
versions of the same presentation with different versions of the story and like
47:34
see what I would find. I'm not though. There's no way I would actually do that.
47:38
But but I can delegate my infinitely patient agent to go do those things for
47:41
me and serve up just some morsels of ideas so that I with like a breakfast
47:45
burrito in my mouth on the go on my phone can say like, "Okay, yeah, option
47:48
two. Let's do it." I love it, John. Well, that sounds like
47:52
a perfect place to end, which is uh with an AI that's perfectly patient, making
47:56
every last little headache in our lives. Just like just that much nicer. Um,
48:00
thanks for coming on. Uh, I appreciate you sharing the behind the scenes on the
48:04
Gamma's journey. Uh, I think there's going to be a lot for folks uh to to
48:07
learn from this one. Uh, if you're listening and want more stories, uh,
48:11
make some recommendations. Uh, like, subscribe, do all that good stuff and,
48:15
uh, follow Agents of Scale wherever you get your podcast. I'm Wade and I got to
48:19
see you all next time.