YT YouTube2Text

YouTube2Text → Transcripts → The Untold Story of Higgsfield | Burning $4M a Month on AI Models | CEO, Alex Mashrabov

The Untold Story of Higgsfield | Burning $4M a Month on AI Models | CEO, Alex Mashrabov — Transcript

by 20VC with Harry Stebbings · 12,584 words · 1,855 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00On average at Hicksfield, person on the
0:01team spends over $10,000 a month on
0:04various models. So internal usage of
0:07models a month is over 4 million.
0:10Hicksfield. This is the story that no
0:12one has told in startups yet. The
0:14company has just hit a billion in
0:16revenue. It is the fastest growing
0:18company in consumer land to hit this
0:20milestone. It even surpassed Cursor.
0:22Alex, the founder, is an incredible
0:24genius. This is the story that you don't
0:27know that you need to know. My parents
0:29told me that I must get to the United
0:31States cuz this is the place where
0:33technology matters. By the age of 19, I
0:36was able to get to top three in the
0:37world in competitive programming. I just
0:39caught a guy who spent over 30k in a
0:42week on Astra model. Many people spend
0:44over 10,000 in a week. Ready to go.
0:57Alex, I am so excited for this dude. We
1:00were talking downstairs and I said, I
1:02don't think the Higsfield journey has
1:04been told before and it's it's an
1:06amazing journey. So, thank you so much
1:08for joining me today.
1:09>> Uh that's very special opportunity for
1:11us. Thank you for having me. Obviously,
1:14your story is inspiring as well, like
1:16how social media has become like an
1:18elevator for you, opportunity to create
1:20fun and so on. Dude, it's very kind of
1:22you to say. I do just want to go back
1:24though because you're not the Stamford,
1:27Silicon Valley, born and bred engineer.
1:31You were a competitive programmer in
1:33Kazakhstan. Can you just take me back?
1:36How did you first find and fall in love
1:37with computers and become a programmer
1:39so early?
1:40>> So, first you need to understand where I
1:42come from. So my father is from
1:44Usbekiststan. Usbakistan is a country in
1:46central Asia where like if a family of
1:50five people makes $1,000 a month, it's
1:53considered to be wealthy. So it's like
1:56not very high standards of living
1:58unfortunately. So um but both my parents
2:01are professors of mechanical
2:02engineering. Since I remember myself
2:04since I was eight, my parents told me
2:06that I must get to the United States
2:09because this is the place where
2:10technology matters.
2:12So um my mother had to work three jobs
2:17because basically my education was to
2:19compete in programming competitions all
2:21the time and to go to various
2:23educational camps where I could learn
2:25from the best like certain data
2:28structure data structures algorithms and
2:30so on. Can I ask you a question? Did you
2:33feel pressure as a child competing being
2:37pushed into these environments when you
2:40were so young?
2:42>> Absolutely. Uh but and and and I'm very
2:44grateful to my parents that they showed
2:46me the path really from that from that
2:48early on. Um definitely when you come
2:52from this part of the world think about
2:54post Soviet countries uh India China
2:57like getting to the top of the rankings
3:00in any competition in any international
3:02competition is the only way to really
3:05break out. So by the age of 19 I was
3:08able to get to top three in the world in
3:09competitive programming. But then
3:12instead of pursuing like um like
3:14academical career decided to do
3:16startups. [laughter]
3:19I'm sure your parents were thrilled. Uh
3:21can you take me to that decision? Like
3:23this is like the penultimate moment.
3:24You've worked 19 years for your parents
3:27have told you this is like the mother
3:28load. This is the thing and you're like
3:31I'm going to go and do this really risky
3:33thing called a startup at this point
3:35like what happens then?
3:37>> So let me take you back to 2014.
3:40I was very fortunate to work on
3:43pre-transformer architecture neural nets
3:45and I was primarily just doing
3:47optimization make it run faster um
3:50parallel across multiple machines and so
3:52on and um I was and we actually build
3:55state-of-the-art system for language
3:58translation from English to Russian and
3:59Russian to English apparently talent
4:02wars were a real thing even back then a
4:04lot of my teammates were hired by Deep
4:07Minds and Meta and uh but My passion was
4:11actually different. I was very very
4:13surprised to learn when I come to to for
4:16the first time how quickly Uber actually
4:19spread out. And I was thinking if like
4:23this app can take over the world so
4:26quickly and transform the whole
4:28industry, maybe what's going to happen
4:30is that mobile phones are going to
4:32become the most used devices in the
4:34world. Maybe there is going to be a
4:35version of the future where everyone is
4:38going to be spending most of their time
4:39in their life watching AI generated
4:42videos on the phones cuz I mean who else
4:44is going to produce videos for for the
4:46phones? Maybe it's going to happen with
4:47AI.
4:48>> Okay. And so that was the company that
4:50we built before that you sold to Snap.
4:52>> So yeah, so the company was called a
4:54factory. Um was fortunate to meet Mahi
4:572018. He's co-founder of Hicksfield and
5:00he is a like veteran of Silicon Valley
5:03went through ups and downs and um sold
5:06it to Snap for 100 for million for 166
5:10million and um then I was leading Jenny
5:14there pause no offense dude you come
5:17from um you know a family of incredibly
5:21ambitious parents who push you to do
5:23well and you just skipped the moment
5:25where you sell for 166 million It's a
5:28lot of money. Um, how did that feel when
5:31you did it?
5:32>> We both remember these times where the
5:35capital for AI companies was not really
5:38that much available and when and AI
5:41multiples were not like 200 to revenue
5:43as they are today but closer to zero cuz
5:46AI was not a topic. So there was like
5:48severe del dilution which we
5:50experienced. So you [laughter]
5:52just to calibrate. So can you
5:54>> okay what was around?
5:56>> No look I mean back then rounds like
5:58rounds of like$12 million having like$12
6:01million in investments was considered to
6:03be really good. Uh but it but it was
6:05still an opportunity for me to finally
6:07go to the United States. So after the
6:09acquisition I permanently moved to uh
6:12first to LA and then to Silicon Valley
6:13and my dream simply came true.
6:16>> Was it what you thought it would be?
6:18>> That's a good question. So um as San
6:20Francisco is definitely a place where no
6:24one judges by race, nationality and so
6:27on and that's that's um that's truly
6:30phenomenal. There is definitely a
6:32meritocracy in a sense that it's
6:34possible to meet anyone but in the same
6:37time what I see across Silicon Valley
6:39investors it's extremely consensus
6:42driven. So um I mean I think that last
6:44part I expected to be different but then
6:47I read the book about the law of capital
6:49and I realized this is just how the
6:50world works.
6:51>> So then tell me we have sold to Snap
6:54we're now in the US this is the moment
6:57you wanted how does Higsfield come to be
7:01back then like Snapchat 2020 was uh
7:04really growing so so quickly and the
7:06face filters which my team has built was
7:09driving most of daily new users. What
7:11what's important is that um these face
7:14filters we were able to manage to run on
7:17mobile devices. So it was virtually for
7:19free for Snapchat. It's not like current
7:22LLM tokens cost. Um and but but it and
7:26it and it scaled to hundreds of millions
7:28of people throughout the world. And it
7:30was truly phenomenal to me to build a
7:32product which is still probably the most
7:34used consumer media AI product. But then
7:37um but then what I realized is that
7:40there are a lot of unmet needs on
7:44advertising sites. Average company
7:47cannot figure out how to be relevant on
7:50social media. So and this is a major gap
7:53like social media is the main media in
7:55the world. A lot of companies are
7:58actually able to build direct response
8:00advertising so that they can actually
8:03sell more. But in the same time, most of
8:05the companies in the world cannot simply
8:07do that. And basically, no because no
8:10one simply can keep up with the pace of
8:12production for social media as trends
8:14change pretty much every day.
8:16>> Mhm. And so you were like, hang on a
8:17minute, these big brands aren't able to
8:20have media houses and so we need to
8:22create a tool that lets them. That was
8:24the cell.
8:24>> Yeah. Ex. Absolutely. So where it all
8:26really started is that we like there was
8:29a tool like to upload set of images and
8:31transform them into a slideshow with
8:33music.
8:34>> It's kind of better than nothing but
8:36still pretty bad, right? So another
8:38solution was to take long form video and
8:41cut them to short vertically oriented
8:43videos. This was better but still really
8:46not perfect. And it felt to me that um
8:48especially 2023
8:51it was absolutely clear that scaling
8:53loss finally work. It's not just a
8:56concept from science that scaling laws
8:59work. Video just takes couple I mean
9:02maybe two three years longer than LLMs
9:05and coding. Uh but it was clear that uh
9:08actually finally scaling loss should
9:10work in video as well and I decided just
9:13to take a bet. But I just want to go
9:15back. I get that in terms of what we
9:17see, which is, hey, we want to empower
9:18these brands and companies to create
9:20amazing media for social media,
9:23but it wasn't a hit from day one. And I
9:26spoke to Amy at Menllo who mentioned
9:29like a couple of pivots before and the
9:31meandering that we had. So what happened
9:34when we launched? Did we have immediate
9:36product market fit? No, actually we
9:39spent
9:41more than a year in a search of a
9:44product which could work. We burned more
9:47than 10 million out of 16 million raised
9:51in seed fundraising.
9:54So we felt we have just one attempt
9:56left.
9:58And frankly I feel I I'm responsible cuz
10:02I was focusing on the wrong things. I
10:05think I just lost the touch with reality
10:09back then. I was so much optimizing for
10:12what's hype today, what's the right
10:15narrative, how we can hijack the
10:17attention, all these things really
10:20like everything instead of building a
10:22good product. So when we had less than 6
10:26million lefts, I guess it was slightly
10:27less than five actually, I realized that
10:30the only thing which we can be focused
10:31on is to lean into the product PLG and
10:36just finally set belief that the best
10:40product is going to win. And um so and
10:44then we just started to talk to
10:46customers. We spoke to eight creative
10:49directors about their experience with AI
10:52and what's simply missing. Everyone told
10:55us that camera control does not exist in
10:59AI and camera control is so important to
11:01tell a story. So this is a very
11:03important bottleneck to solve. So we
11:06released our products uh March 31st last
11:09year and since then we are really riding
11:12this crazy wave.
11:13>> Was it immediate product market fit
11:15then?
11:15>> Like yeah it was immediate. Is product
11:17market fit like love? When you know, you
11:20know.
11:21>> Um, yes, it's definitely when you know,
11:23you know. Like for example, we don't do
11:25any paid and like we have we have on the
11:28team people who scaled businesses to
11:32over like billion and two billion in
11:34revenue like other businesses um with
11:37paid advertising. Like at Hicksfield, we
11:40decided to really make a bet that
11:42>> we don't do paid.
11:43>> We don't do paid. Is influencers not
11:46paid?
11:46>> That's a good point. So, um with
11:49influencers, there is typically there
11:51are different types of influencers, but
11:54typically there is um some fee for just
11:57video production and then like some cost
12:00per click like attribution which is like
12:02works really well on YouTube. You you
12:04guys got into some controversy
12:07[laughter] for like I can't remember
12:09what it was. you were like pay paying
12:12people to promote for you or doing
12:15something rogue with influencers.
12:18Was that completely unfair? Was it kind
12:20of my bad we did do that? How do how do
12:24you respond to that?
12:25>> The main takeaway from like our
12:27experience is that it's very important
12:29to own own distribution. Distribution
12:32now more important than ever. And like
12:34we basically did outsource we had just a
12:38team of like two people on creator and
12:40customer success sides and we just did
12:42outsource to the agency and this was not
12:44uh that was not a good experience but uh
12:47we are still but but we are still trying
12:50to
12:52find interesting opportunities to tell
12:55about new media formats. Some of them
12:58are rather controversial. So, for
13:00example, recently we partnered with
13:02Neon, one of the largest streamers in
13:04the world, and launched like his own
13:06sort of AI generated stream. Um, like no
13:09one else did this before cuz this is
13:11like real creator making a replica of
13:14themselves. A lot of people start to
13:16question uh start to question their um
13:20like is it really authentic content or
13:23not? But in the same time, those
13:25creators are under immense pressure. We
13:28all know about the story for about from
13:30Mr. Beast about like really how much
13:32like there is just pressure to
13:33constantly perform. So um and we also
13:36know through conversations with many
13:38talent agencies a lot of top stars
13:41actually want to be able to do more if
13:45they could create digital replica. But
13:48so what's happening today very
13:49frequently is that um those
13:53a tier celebrities they simply come up
13:56for a recording on like let's say green
13:59screen and then there is just a lot of
14:01post-prouction which goes on top of it
14:03and it feels to me that uh we are we we
14:06naturally going to come to the point of
14:08time where a AI digital replicas are
14:11going to become just one of the ways how
14:13creators can monetize.
14:14>> Totally get that. I do just want to go
14:16back to part of the story. Where are you
14:19at revenue-wise today?
14:21>> Uh so today is actually exciting day
14:24like when we record just Bloomberg
14:26article went out so that we cross 1
14:29billion in annualized revenue. Um if I
14:32had a gong here I'd be like hitting the
14:34gong. A billion in revenue.
14:36>> Yes. Um actually it took us 18 months
14:41from 1 million to 1 billion for Corsor
14:45it took 24 months. Um so we are probably
14:50uh probably like the thirds after open
14:52the anthropic
14:5418 months from a million to a billion.
14:57>> Yes. How do you calculate revenue? Like
15:01it's a controversial topic. Um, how do
15:05you help calculate revenue?
15:07>> Absolutely. Uh, by the way, your um,
15:09co-host uh, Jason also asked this
15:11question in May. [laughter]
15:13Luckily, answer didn't change. So, we
15:15are at least consistent. So, but let me
15:17be transparent on that. What we do is we
15:19look um, revenue over the last four
15:23weeks and multiply it by 13 from what I
15:27know openable all of them use the same
15:29methodology.
15:31What's very important is that we are we
15:36take revenue not sales. So if that's
15:38like annual subscription or annual
15:40enterprise contract we prorate this
15:43across 12 months and take only this uh
15:46and only take like a piece which
15:48corresponds to one month to 28 days to
15:51be uh to be precise. That's the first
15:53piece and second it's only live revenue.
15:56It's only live revenue. We are not
15:58taking like three year enterprise deals
16:00and baking into like 1 billion figure.
16:02No, we don't do that.
16:03>> If you were to break that billion up
16:05today into annual contracts, monthly
16:09subscriptions and then token spend, what
16:12would that be?
16:13>> So, um, videoi is still relatively early
16:16in my opinion. Uh, it is still probably
16:20two years behind coding in terms of
16:23adoption. So on demand usage for leading
16:27to coding companies could be over 50%.
16:30And I would be honest for video it's
16:32substantially less than that. Um in the
16:35same time what's very interesting for us
16:38to observe in the business is that there
16:41is sub significant revenue expansion. I
16:45always love to study stories of the
16:47largest customers on the platform. So,
16:50one customer started um 6 months ago
16:53spending just subscription $99
16:57a month. $99 a month. And now we just
17:01signed a deal over 6 million.
17:04>> 6 million.
17:04>> 6 million a year. Right. So, yeah. Like
17:07this level of acceleration is something
17:10which really like mind-blowing to me.
17:13Dude, what are they getting for 6
17:15million a year? that's like a Hollywood
17:17content team almost.
17:19>> So there are multiple trends um as and
17:22all of them frankly coming from Asia.
17:25>> So first we're seeing a lot of um direct
17:27to consumer e-commerce companies
17:30rebuilding their whole go to market to
17:33be AI native where they make uh where
17:36they just make hundreds of ads if not
17:39thousands a week where they can AB test
17:42what performs well. But we all know
17:45about like short form dramas, right?
17:47Like most like short form dramas today
17:49is an industry over 10 billion owned
17:52primarily by Chinese companies having
17:55huge impact both in China, United
17:57States, in Europe, everywhere in the
17:59world and most of new shows there are
18:02made with AI end to end. So look, I
18:04think uh like the this adoption
18:07obviously is uh coming like bottom up,
18:10but um that that's very difficult to
18:13refute this new reality.
18:15>> What percent of revenue is consumer
18:17versus enterprise?
18:18>> So that that that's a great question. So
18:21um
18:23so B business revenue is slightly over
18:2650%.
18:26>> Wow.
18:27>> Yeah,
18:27>> that's impressive.
18:28>> Thank you. Um on the consumer side, it's
18:31also very important to break it down. So
18:34on the consumer sides out of these 50 is
18:37around like 10% is pure consumer use
18:40cases pure consumer and that's roughly
18:42people who use it on mobile. So share of
18:44our revenue from mobile is less than
18:4610%. That's why we are we are very
18:49different from many other companies and
18:51but there are lots of aspiring creators
18:54like basically those people who are
18:56freelancers doing social media marketing
18:58projects and so on who try to learn
19:00video AI so that they can make more
19:03money. It's true that their behavior is
19:06a little churny. uh within a year most
19:10of them actually come back to try again
19:12and we do believe that over the time
19:16most of them are going to figure stuff
19:18out and they're just going to become
19:20this new AI native workforce. So it's
19:23still important for us to educate them
19:26and uh that's why we invest so much in
19:28like Hicksfield Academy, YouTube channel
19:30and so on. But we also are f fully
19:34cognizant that we will never be able to
19:38win in a market of subscriptions of $20
19:42a month.
19:43>> So why? Because like I think like today
19:46Google and Open AI they pursue like ads
19:49so much but fundamentally I think they
19:53are going to completely demolish all the
19:56consumer subscription markets which is
19:59uh $20 a month subscriptions.
20:01>> Oh, so you saying that because they
20:04provide a horizontal product that's very
20:05good, you're just going to not pay for a
20:08lot of the verticalized products that
20:10you used to pay $ 20 $30 a month for.
20:12>> Yeah, I do believe that. That's
20:14essentially what's going to happen over
20:15the time. Um, I know this is a very
20:18contrarian bets, but um, at least we can
20:20see some of that.
20:22>> I think it cannibalize Canvas growth if
20:24you're honest. A lot of the lowhanging
20:25fruit on the consumer design side that
20:28Canva used to serve can now be done in
20:30open AI in particular.
20:33Is that what you're talking about?
20:34>> Yeah. And I do believe this is just the
20:36most apparent example, but there are
20:38couple more which is which is already
20:39happening. And I do believe that uh
20:41that's why for at Hicksfield what what
20:44really matters for us is how we even if
20:46we get someone on like $20 a month
20:48subscription like how can we show them
20:51value how can we make them to upgrade to
20:54over to spend over um to over $1,000 a
20:58year with us. I can't believe that's 6
21:00million a year from 99 bucks. That's the
21:03best ever slide on a fundraising deck.
21:06[laughter]
21:06>> And all of our customers are going to do
21:08the same. Exactly.
21:09>> Can I ask you mentioned there kind of
21:10churn rates when you look at 30-day
21:13retention rates for consumers and 90-day
21:16retention rates. What are yours and what
21:19is good? So, there is um quite massive
21:24drop within the first month
21:27>> just simply because people don't fully
21:28realize the value and that's a that's a
21:31core priority for us to actually get
21:33better in that. So, showcasing the
21:34value. Is it like half or like
21:36>> No, it's uh it's maybe like 30% drop.
21:39Okay.
21:39>> But then it's it's really flat after
21:41that. It's we look obviously at like
21:43logo retention.
21:44>> Mhm.
21:45>> I wouldn't say it's great but because
21:47like we all remember like B2B SAS era
21:49like uh retention was expected to be
21:52logo retention month one was expected to
21:54be over 80%. M
21:56>> um so clearly we have uh we have a lots
21:59of work to do on uh user education to
22:01get there but some things are truly
22:05phenomenal like when I look at the cor
22:07at the business segments and NRR at
22:11month 12 obviously like you're going to
22:13argue it's like 18 months old company
22:15like what are you talking about but
22:17still when I look at the numbers which I
22:18have today NR at month 12 is over 300%.
22:22just it just never happens in B2B SAS
22:25right so um that's why I'm saying that
22:28while there is substantial churn in
22:30month zero and we have to do better job
22:34with user education to address that
22:36expansion is unprecedented can we
22:38actually just unpack the two different
22:40go to markets cuz you got consumer and
22:42you got enterprise and I spoke to quite
22:44a few of your competitors in all honesty
22:46before this show [snorts] and I said hey
22:49you we've got Alex coming on what should
22:50we ask him everyone said the same thing
22:53which was an admission of their respect
22:55for this particular kind of GTM. They
22:58said you've ex executed the most
23:01impressive influencer campaign in tech
23:04and what I wanted to understand was when
23:07you look at the consumer growth
23:10what worked what didn't work and how do
23:13you reflect on that first and foremost
23:16like the goal is to make sure that the
23:18best commercial video content is
23:21generated on Hicksfield and we show all
23:24the workflows of how to make such uh
23:27professionallook videos and we have an
23:29in-house team of over 150 creative
23:33professionals.
23:34150. It's it's almost half of the whole
23:37work workforce frankly. And um they th
23:41those people they make product launch
23:43videos, they make tutorials like for
23:47example we made the first generated
23:49movie which is also like obviously um a
23:53very um a very sensitive topic but
23:56what's important we open sourced all of
23:58it and what we learned is that for 90
24:02minutes of uh of like let's say TV
24:05quality content it was over 100 hours of
24:09for yet generated contents. So creative
24:11decisioning like picking the right piece
24:14is still very important. Um so that's
24:17really what's what we are focused on and
24:18that's what's driving most of the most
24:21of the revenue.
24:22>> So you're saying the the growth in
24:24consumer subscription is through own
24:26content and distribution.
24:28>> Yes. We don't do any paids. Early on you
24:30made an interesting architectural
24:31decision to have your own models and
24:35then you since walked that back. Can you
24:38talk me through why did you choose own
24:40models and why the walk back?
24:43>> Oh, um yeah, obviously this was
24:45[laughter] uh obviously this was my
24:46mistake. I'm going to be I'm going to do
24:48my best to be um transparent. What I
24:51need to admit, we really tried we I at
24:54some point of time I really was thinking
24:56that chasing benchmarks
24:59um is valuable but I don't believe this
25:01is just sort of corporate scops frankly.
25:04So um and I was part of the large
25:07organization so I know what happens.
25:08What happens is that everyone just
25:10thinks like we need to show some
25:12progress. So we need to have some
25:14benchmark but then when I talk to the
25:16top researchers from these labs
25:18especially larger companies what happens
25:22is that they start to put test data into
25:24the training.
25:26They start to kind of use uh leverage
25:29test data to use LLM as a judge for
25:32training of the models. use all the
25:34various tricks to basically gain
25:35benchmarks, get get like quarterly
25:37bonuses and so on because like who
25:39cares, right? So if I make my couple
25:41million dollars a year in inside in one
25:43of these labs, I can move to another lab
25:46easily. So that's unfortunately what's
25:48happening in larger organizations. Um
25:51and
25:51>> can I just stay on that?
25:52>> Yeah.
25:53>> What do you mean? You're saying that
25:55they are incentivized by benchmarks and
25:58so because of that they are doing
26:02artificial things to improve their
26:04scoring in benchmarks which actually
26:06don't increase output efficiently. Yeah.
26:08Look, I think let's just look at the
26:10outcomes which we have today. Out of all
26:13the incumbents in the United States,
26:15when I look at open router data, the
26:17only company which is relevant is
26:20Google.
26:22out of all the incumbents when I look in
26:24China where probably obsession with
26:26benchmarks probably is less we have 10
26:29cent shyomi Alibaba
26:33uh like three incumbents being
26:35completely relevant and obviously like
26:38by dance obviously trying to catch up as
26:40well what's your takeaway from that
26:41>> I just do believe that uh the there is
26:44just obviously in the in tech bubble
26:46there is a strong obsession over the
26:47benchmarks uh which do not uh
26:50necessarily
26:51represent the reality. But I can talk
26:54specific specifically in the for video.
26:56>> A lot of benchmarks today for video is
26:58really text to video which does not
27:01represent actual workflows at all. Um
27:04the way to think about video models
27:06today, it's just modern rendering
27:09engine. It think about this as like
27:11Unreal Engine or Unity but just
27:14different types of inputs.
27:17And it's virtually impossible to really
27:20define a visual output and and direct
27:23the execution just through text. If you
27:26just go and to our open source projects
27:28like this movie which I mentioned
27:30average prompt length is over 3,000
27:32words. That's the first thing and like
27:35look all these benchmarks which we are
27:37talking about they are not like as
27:38comprehensive in terms of the details of
27:41prompts and people who are labeling they
27:44obviously don't cannot read like 3,000
27:46long word long prompts but also on
27:49average there are at least 10 image
27:52references
27:54for every for every scene. The reason
27:56why it's important because it's
27:58important to define how the characters
27:59look like, how the background looks
28:01like, like how actually characters are
28:05located to each other in the scene and
28:07so on. And so that's why like prompting
28:09and like just the workflow is so
28:11complex. Benchmarks just don't rep don't
28:13represent that.
28:14>> So going back to the model selection,
28:17why did we decide we're going to do our
28:19own and then why walk it back? It's true
28:22that like with VFX and camera control,
28:25we got very very quickly from like maybe
28:291 million to 20 million in AR within
28:33maybe the first 3 months. Then we
28:35released own image model which is really
28:39good at um aesthetic photo shoots and
28:43product consistency. This is what
28:45allowed us to scale then from 20 to 100
28:48million. So help me understand, Alex,
28:51why did you decide that you were going
28:53to do your own models and why did you
28:55abandon them? [snorts]
28:56>> We still do them whenever we see like
28:58specific use case like these photo
29:00shoots.
29:01>> Uh but but as soon as this is what our
29:03customers want. So it's all driven based
29:05on the customer feedback, not just by
29:08ambition to conquer the worlds and
29:11[clears throat] build the best model in
29:13the world. Do you think every company
29:15will have their own models like we're
29:17seeing Harvey, we're seeing Cognition,
29:20we're seeing Mccor, Ramp build their own
29:23models and we'll see every company have
29:25their own models with their own data or
29:28we actually all use a series of
29:30providers. So um first of all whenever
29:34just to be honest whenever someone says
29:36we build our own models very likely what
29:38they mean is something what see what's
29:40happened with Corsor. We we do remember
29:42right a lot of companies they actually
29:44take open weights model and just post
29:47train on own data.
29:48>> Mhm.
29:49>> Um and post training can happen in two
29:51ways.
29:53Most importance is whenever you have um
29:56customer data around like decisions they
29:58make like sequence of decisions and you
30:01can teach the model to actually take
30:03like learn how to compress these 10
30:06steps into one step. like this type of
30:09reinforcement learning is the most
30:11valuable. So and I think like
30:13increasingly more and more companies
30:14will have to do that frankly just we see
30:18this in the market as well. So the most
30:22most of the companies in the world today
30:25most of the businesses they don't
30:27necessarily need Astra specifically they
30:30don't necessarily need the newest fable
30:33model and that's why like open router
30:36reports that uh share of open source
30:38models went from below 30 to over 60
30:41within within this year.
30:43>> What do you think share of open models
30:45will be in two years time? Look, I do
30:47believe that just because the cap
30:49capitalism works, I mean openly
30:51ananthropics still are going to have
30:53more than 50% of the markets
30:54>> in terms of the dollars generally
30:56>> in terms of the dollars, right? And
30:58especially because uh for coding still
31:00remains to be very very prolific use
31:02case where coders are always jumping to
31:05to to to the recent model
31:08over but for our markets we're seeing
31:10completely different dynamics. what's
31:12actually happening in social media
31:13marketing as companies start to print
31:16hundreds of create ad creatives um a
31:19week they want to have maybe cheapest
31:23more steable models cuz like PhD level
31:26intelligence is not necessarily needed
31:30for to make viral social media video. So
31:35um and that's where we actually have
31:36seen that um we get like 80% plus margin
31:40whenever we run open-source models like
31:43post-trained open source models. Uh but
31:47it can be way more cost efficient for
31:49our end customer compared to the
31:51proprietary models.
31:53>> What's the comparison on margins between
31:55open versus closed for you?
31:57>> The margin on own models and open
32:00weights models is over 80%.
32:03Um, and then it almost doesn't matter.
32:05And for closed source models, it's
32:07probably between 20 and 30%. And then
32:09what becomes important is can we
32:12actually steer the traffic. What makes
32:15me excited about Hicksfield is that umic
32:19grows so quickly and actually for us as
32:24companies start to actually create those
32:27agentic workflows to make more ads we
32:31choose which model we can use. So like
32:34we choose what model to use in over 40%
32:39cases.
32:39>> In a way model routting becomes a core
32:42feature of the business. No.
32:43>> Yeah. We call we call it tokconomics
32:46essentially right as like there is
32:48certain amount of work customers want to
32:50do um how can we optimize number of
32:54tokens which requires and how we can
32:56pick the most efficient tokens for them
32:58there are actually two incumbents in the
33:00United States who figured out models
33:02it's not just Google it's also Nvidia
33:03why do you think that is what I'm
33:05constantly seeing is that um the there
33:09is the versions of models so there are
33:11these state-of-the-art models
33:14which have to be really good in computer
33:16use like Astra or in coding. Um but they
33:21can be prohibitively expensive and we
33:24we're chatting about that like on
33:25average at Hicksfield person on the team
33:27spends over 10,000 over $10,000 a month
33:32on various models and remember like we
33:34are split across United States and Asia
33:37across
33:37>> so how much do you spend on models per
33:39month? So internal usage of models a
33:44month is over four million.
33:47>> Wow. How many people do you have?
33:49>> We have close to 400 people and just
33:53want to make sure that the math adds up.
33:55Yes, it's um it's definitely over it's
33:58definitely over $10,000 per person.
34:01>> How has that changed over time?
34:03>> That's the best question of the whole
34:04show, by the way. Um that's the best
34:06question.
34:08What actually started to happen is the
34:12creative team started to do VIP coding
34:16like the like like this month I was I
34:19just caught a guy who spent over 30k in
34:23a week on Astra model
34:28cuz he was frankly frustrated that some
34:30like asset organization workflow and as
34:33you said like basically auto editing is
34:36still not very good in production and he
34:39said, "Oh, I'm just going to do this
34:40myself." And just went like five nights,
34:43five nights straight on Astra
34:46>> and it works.
34:48>> We learned a lot. I wouldn't say it was
34:50production ready, but we learned a lot.
34:52>> 30,000 in a week.
34:54>> Yeah. Yeah. Many people spend over
34:5610,000 in a week.
34:57>> Do you mind?
34:59>> Yeah. My finance team will probably say,
35:02I don't know if if you know if you ask
35:04any of them, but they will probably say
35:05that I'm like being too stubborn, too
35:08relentless to control the spend cuz
35:10sometimes I feel it goes like [laughter]
35:12it it really goes out of control like
35:1430k in a week is quite a lot. But we
35:17learned this. So this was actually net
35:18positive experience.
35:20>> Okay. So the internal spend 4 million
35:22about 10,000 per head. What will that be
35:25in 12 months time do you reckon?
35:27>> So that that's very interesting. So
35:29across uh the top uh the top engineers
35:32and across top creatives I think it's
35:35going to keep growing and I do believe
35:38we are going to get to to to spend um
35:41close to 50k and 100k a month for those
35:44who can call 10x engineers 10x creatives
35:47unfortunately I also expect that these
35:50people will ask for comparable salary
35:53raise as well so I think that's just
35:55going to correlate at some points um but
35:57also for a lot of other jobs. Let's say
35:59we to take legal finance and so on. I
36:02think it it really stabilizes around
36:04like um $500,000
36:08a month very very quickly. With those
36:1110x engineers, the idea is they have
36:14thousands of agents running below them
36:16doing a lot of the difficult execution
36:18work that took time. Do we just have
36:21dramatically smaller teams with those
36:2210x engineers, 10x designers, 10x
36:26finance leaders? I can definitely say
36:29that
36:31the I I I had sort of a feeling that
36:35legal
36:37customer support
36:39um is going to be mostly replaced and
36:41that's obviously one of the main u
36:43mistakes operation which we have done in
36:45the company that we didn't ramp these
36:47teams quickly. Um what we are seeing
36:50today is that like let's say our legal
36:52team is like over 10 people our customer
36:55success team is over 40 people all of
36:58them use AI heavily we like at at these
37:01professions where I say quite close
37:03today I definitely can say that uh there
37:06is I don't see any elimination it's true
37:09that probably over 60% of customer
37:12support requests especially the first
37:13line of defense can be handled with AI
37:16but when it especially comes to B2B
37:18It doesn't like it like AI just doesn't
37:20work.
37:21>> Revolute has now over 92%
37:24resolution rate on customer support for
37:26consumers.
37:28>> Pretty good.
37:29>> It's it's it's pretty good. But
37:30obviously they did invest a lot into
37:32that
37:33>> [ __ ] ton. A [ __ ] ton.
37:35>> And and but also very important the way
37:36how Nick thinks about that uh is in
37:39terms of the playbooks. We launch
37:41products, new products pretty much every
37:43week. So um we have to we have to keep
37:48update agents with all the information
37:50and so on and just due to the high
37:53velocity having um extremely smart
37:55coordinated team is is very important.
37:57That's really interesting how product
37:59velocity increases leads to harder
38:02customer support for agents.
38:05>> Of course, cuz uh the agents are as good
38:07as context and rules which they have.
38:09And if context and rules change pretty
38:11much twice a week, it gets a little
38:14difficult.
38:14>> When you look at your engineering team
38:16today, what are they on? Are they on
38:18cursor? Are they on codeex? Are they on
38:21core code? So from a period from March
38:24to June, everyone really moved to claude
38:28um including the creative team and
38:31that's where we actually started to see
38:33creative team vibe coding functionality
38:36which we don't have in production. But
38:39then we started to see that all the
38:42coders quickly moved from claude to
38:46codeex um as of mid June and um over the
38:52time especially
38:5410x creatives moves to codex as well but
38:57look I do believe that there it's it's
38:59it's cyclical so
39:01>> it's so cyclical my question to you is
39:03will we continue to see the velocity of
39:05model release that we're seeing now you
39:07in 3 years time will It be like, "Oh,
39:10Gemini this week. Oh, CL Anthropic this
39:13week, OpenAI this week." Or will we see
39:15a a reduction in model release rate? I
39:20don't think that's going to happen
39:22anytime soon. So, I believe like for
39:25example, recently OpenAI announced that
39:26they basically build OpenAI for law,
39:29>> right? But that's only V0. So over the
39:32time they also are going to try to print
39:34smaller specialized models for like not
39:37like exactly smaller uh but really
39:39specialized model for certain use cases.
39:42Um clearly like Astra excels in
39:45long-term horizon.
39:46>> Do you buy that? Like I look at that GPT
39:48for law from Astra and I'm like I'm
39:51sorry I think it's complete [ __ ]
39:53with the greatest of respects. It is a
39:55very deep functionality required to
39:58serve some of the biggest law firms in
39:59the world. like very very deep and
40:01specific functionality. It's very
40:03specific according to the different
40:04types of law as well. Plus, if you want
40:06to sell into these law firms, it's a
40:09multi-year sales cycle with some of the
40:12stodgy old lawyers and partnerships.
40:15You can't just say, "I'm open AI. Yep.
40:18We've just hacked into the Australian
40:19government, by the way, but we're here
40:21to serve your law firm."
40:24Uh, okay. Yeah. So, first of all, I
40:28think uh just uh definitely
40:31the ability to switch internal use just
40:35for internal teams outside of law firms.
40:37I think that's I think that's definitely
40:39happening. Oh, I think we both investors
40:41in company called solve intelligence.
40:43>> Love it. Yeah. Very specific. Very
40:45specific. And let me try to maybe bring
40:48couple examples
40:50>> why like solve intelligence is so
40:52special and like where like for example
40:55how we learn from this.
40:57What can happen very often is that a
41:01company want to just control the patent
41:04workflow even if they outsource the work
41:09and that's very valuable just to have
41:10one system of records. So whoever can
41:13create AI native system of records is
41:16going to win. And but going back to
41:18Hixel why it's so important for
41:19Hicksfield
41:21there are so many systems today which
41:23are used for just to store assets
41:25>> like some people use Dropbox
41:27>> some people use Google Drive
41:29>> some people are going to try to use Miro
41:32some people are going to try to use
41:33frame.io like there are many solutions
41:36but let's think about what people need.
41:38What people need, they want to be able
41:40to search contents and and marketers
41:43especially want to make sure that
41:45content is on brands in terms of the
41:47visual identity, but also like if that
41:51sort of adheres to certain brand
41:53guidelines
41:54and that's where like semantic
41:57understanding and semantic controls
42:01become finally possible. It never
42:03existed before. So in our space there
42:05are definitely other companies like
42:06Adobe and Canva who builds the best
42:10software for the pixel first era where
42:13everything was defined with pixels but
42:15that's clearly not how the world is
42:18going to work in the future. What we're
42:20envisioning and that's what everyone
42:22wants. They want to just be able to
42:24search and um like really work through
42:27the library of assets and all the
42:29knowledge through natural interfaces. So
42:33being able to own this interface and
42:36build the analytics uh like this system
42:38of records is important. That's why at
42:41Hicksfield we invests it so much in
42:43harness so that it improves over the
42:45time. And this harness also um allows it
42:51basically learns visual style over the
42:54time which let's say cloud and open AI
42:56cannot necessarily do.
42:58>> Do you believe in moes anymore? you
43:01you've been around startups for a long
43:02time. We always talked about moes and
43:04defensibility. I largely think they're
43:06[ __ ] You know, we we saw lovable
43:08when I invested. Everyone was like, "Oh,
43:10it's a rapper. It's a rapper, you idiot,
43:12Harry." And actually, it was a rapper,
43:15[laughter] but it's about speed of
43:17decision making, product execution, and
43:20building value over time very, very
43:24fast. Instinct is a rapper. Of course,
43:27it is. It's not that difficult to do an
43:28AI assistant today which why there's so
43:30many but they're building incredibly
43:33quickly very valuable features and you
43:36build it over time. Do you believe that
43:38moes actually exist really?
43:41[sighs and gasps]
43:41>> I know like you ask this everyone um so
43:44um and this is cuz this is on top of
43:46everyone minds like how to think about
43:48the metrics which matter today and how
43:50to think about the modes. So um I think
43:55um it's very difficult to figure out
43:57where the value occurs in the supply
44:00chain. Um we do believe that there are
44:04only two like ways of uh modern value
44:08creation or modes today. First is when
44:11you deliver the outcome and for us it's
44:14allowing businesses to sell more through
44:15AI ads. So that's the first thing and
44:18the second thing is network effects.
44:21Unfortunately, AI does not replace
44:23network effects. And when people talk
44:25about swarm of AI agents talking to each
44:27other, I'm not sure this is happening in
44:29the next five years. So, um, that's why
44:32it's so exciting that within Hicksfield,
44:34like we really wanted to empower
44:36community to create more projects, open
44:40source, open source them to really build
44:42a snowball where people can capitalize
44:45on each other output. This is the reason
44:47why software grows so quickly cuz it's
44:50so easy just to go and fork someone's
44:51project on GitHub. So, and like we were
44:54able to scale from basically like I
44:56don't know 10 seeded projects, open
44:59source projects like 8 weeks ago to over
45:0210,000 today like seeing these type of
45:04network effects I believe can become a
45:07mode over the time. When we look at your
45:09growth, fundraising is a big part of it.
45:12It costs a lot of money to be able to
45:14spend four million on, you know,
45:16different aspects of, you know, uh,
45:18inference band.
45:20What was the best VC meeting you've ever
45:22had?
45:23>> Obviously, um, Yuri Milner gets gets it.
45:27>> How was that meeting? Like, was it was
45:28it in person?
45:29>> Yeah, definitely in person. And
45:31definitely Yuri stays on top of all the
45:33trends. And
45:33>> how was it? Were you nervous?
45:36>> I I wouldn't say nervous. It was just
45:38more uh to see how much of the uh if we
45:43see the market the same way and I was
45:47truly surprised that Yuri deeply
45:50understands this transformation of
45:51content first and foremost. Obviously it
45:54starts with this direct to consumer AI
45:56ads. It starts with short form dramas.
45:58All these trends come from Asia to the
46:00west. And um also fundamentally
46:06we believe that most of contents on
46:09social and in the world is going to be
46:11AI assisted or AI generated
46:14and uh the and like this multi- trillion
46:19advertisement industry and you know like
46:22contextual
46:23advertisement is the main business model
46:25of the internet. It's all going to be
46:28substantially disrupted with video AI.
46:31This industry still going to be very
46:33valuable, but it's never going to be the
46:35same.
46:35>> Did you know when you left the meeting
46:37with Yuri that he was going to write the
46:38check?
46:39>> You know, sophisticated investors, they
46:41can play games. I had like so many scars
46:43like people really shook hands said we
46:45do at this price
46:48and next day what I learned is that they
46:50called other investors and they pulled
46:52the syndicates and to invest in 30%
46:54lower valuation compared to what we
46:56discussed. So like look these things
46:57just happen so you never can be sure but
47:00it didn't happen with Yuri.
47:01>> I think there's a discount placed on
47:04Higsfield because you're not Silicon
47:06Valley insider. Like let's be clear
47:07you're at a billion in revenue now.
47:10>> Yeah. If you were a Silicon Valley
47:12company, that would easily be a $25
47:15billion company growing at the rate that
47:17you're growing in 18 months.
47:18>> Yeah, you could also argue that's what
47:20cognition was valid at 50, right? So
47:22there is definitely an upside
47:23>> upper band even more. Yeah, 100%.
47:26>> So a couple things which I believe are
47:27very important. So first we build for
47:30long term. We have seen that direct to
47:33consumer space like e-commerce can be
47:35disrupted like Shopify is a great
47:36example how they become they have become
47:39infrastructure to build like direct to
47:42consumer businesses and we become
47:43infrastructure to essentially
47:46build distribution for direct to
47:48consumer businesses. That's one
47:50aspiration and second aspiration is
47:52obviously a plain like companies worth
47:54over $200 billion. It's insane. So look
47:58and as we think long term just this you
48:01know like these multiples don't don't
48:03matter that much as we know we're
48:05building long term we're going to be
48:06over 100 billion it's true that most of
48:09the people don't get the opportunity
48:11that we are going after the biggest
48:13industry in the world but I wanted to
48:15drop another another number so when I
48:18and I asked the team to double check so
48:19it's at least four people on the team
48:21who prove so it's not like random fact
48:24so I asked um When we look at public
48:28companies
48:30and we exclude pharma and big tech,
48:33spend on sales and marketing is higher
48:36than spend on R&D. Like what when it
48:39comes to sales and marketing, the goal
48:40is to deliver personalized offering
48:44which converts the best. A lot of that
48:46is human work of course, but a lot of
48:49that is going to be personalized videos
48:50in one in some shape or form. So that's
48:53why I'm saying that um many people just
48:56and that's good for us that many people
48:57don't understand the opportunity this
48:59large market which we go after.
49:01>> Can I ask you you've mentioned Asia
49:04short form dramas a lot. What percent of
49:06revenue is from Asia versus the west?
49:09>> Um so oh the west makes well over 70% of
49:13revenue. well over
49:14>> but just important to say that we learn
49:17a lot from trends coming from Asia like
49:19Hicksfield does not exist in China for
49:21example which is massive market for AI
49:24um Hicksfield uh but the largest city
49:29by usage is soul in South Korea while
49:33the largest country is obviously the
49:35United States
49:35>> what's the biggest lesson from Asia that
49:37you've learned
49:38>> there is so much IP
49:41so many products coming from Asia and
49:45they all try to figure out distribution
49:47direct to consumer. That's why they lean
49:50into the new tooling like video AI which
49:53actually helps to achieve that. That's
49:54just very different mindset. They feel
49:56that they could do they could do way
49:59better if they could establish direct
50:02relationship with customer instead of
50:03having like some other layer. That's why
50:06they go so many so much direct to
50:08consumer rather than using some resale
50:11platforms and so on. I sacrifice a lot
50:13of life for for the life that I have and
50:16the career that I have and I love it. Do
50:19you think you will one day regret
50:21spending a day with your son in 3 and
50:231/2 months? Look, this is goes even
50:25beyond that because my um
50:30from the age of 7 to 12,
50:34my mother had to work um three jobs. So,
50:38I didn't see her. My father was spending
50:40all the time with me going to all and it
50:42was I was basically minor so he had to
50:45go to all these camps with me. Um I also
50:48play checkers. I was top three in the
50:49world. So we went we travel throughout
50:51the world and um then I did programming.
50:54He spent all the time with me like
50:56really dedicated his life to me like he
50:59did sacrifice
51:01and uh since 21st he has Parkinson
51:04disease. So um
51:08even like having some ability to capital
51:10and exits cannot fully change things and
51:13um this is something which is um deeply
51:16personal obviously.
51:17>> Totally.
51:18But you don't need to do what you're
51:20doing now. Alex,
51:22>> I didn't [snorts] need to anymore
51:24either. [laughter] I still am. I still
51:26miss family birthdays. I still miss
51:28weddings
51:29cuz like mine's about a deep insecurity
51:32rooted in me being a fat kid.
51:35um why are you doing it?
51:37>> So I think Mark and Jason actually
51:39described it really well. There are like
51:40five archetypes. So obviously for me
51:42it's just huge conviction about the
51:44technology, about the market, about the
51:46opportunity and just huge fear of
51:49missing that huge fear of missing that.
51:52But remember that um my parents really
51:55taught me that um there is a place in
51:57the world where technology like good
51:59technology products matter. I remember
52:01like when I was six there was um like
52:04this I guess
52:06magazine about Bill Gates like building
52:09Microsoft and not being like very like
52:12socially accepted everywhere back then
52:15and like my mother just told me oh like
52:16these examples basically happened in the
52:18world. I think she didn't fully
52:20understand like San Francisco and
52:21Seattle are different cities but still
52:23uh this that's still deeply rooted in
52:25me.
52:25>> Childhood shape us a lot.
52:27>> Yeah. What did your parents teach you?
52:30>> For them, what was important is to
52:35just be in merit-based environment sort
52:37of um and um that's why getting to uh
52:41California felt so important.
52:44>> What's your biggest lesson on hiring?
52:46Speaking of a merit-based environment,
52:48we see a lot of uh focus on your
52:50cognitions of the world who hire mass
52:52Olympiads.
52:54>> Yeah.
52:55What's your biggest lessons on hiring
52:57effectively?
52:58>> I think one of the things why Europe
53:01thrives so much like I know that you
53:06typically say otherwise but let me just
53:08challenge you like who are the most
53:11relevant NeoClouds today? It's Nscale,
53:14iron and Nobus and Cruso.
53:18>> Mhm. Brusso okay Silicon Valley story I
53:22am from Australia and scale from the UK
53:25and anobus is UK and Netherlands let's
53:28talk about the companies on application
53:30layer they that matter I know that you
53:32mentioned Merore and you mentioned
53:36Harvey but Legora 11 Labs lovable they
53:40all deeply matter so if we just go
53:42outside of the model layer h cuz then I
53:46don't want to go into the mistral topic
53:48right but if we go cuz I think like by
53:50usage they have the numbers are very
53:53strong but people for some reason don't
53:54don't believe in that so I don't know
53:56why but public public data shows that
53:58the usage is there but on every other
54:01layer Europe is extremely competitive
54:04like ASML like without ASML this whole
54:07thing just wouldn't happen so I I think
54:09fundamentally what's matter is if if
54:10like Europe is going to figure out
54:12energy but that's goes outside of that's
54:14above my pay grade right um so very
54:16important to say here is that um now
54:20there are more opportunities to create
54:22company from um different kind of cities
54:26from different parts of the world while
54:28before it all felt extremely centralized
54:32um and we are we are excited uh we are
54:34obviously excited about that and um
54:38another thing about hiring um is that in
54:41Silicon Valley unfortunately what I'm
54:43seeing is that people just jump between
54:45jobs every two years that's why um I
54:48think Europe can be so competitive
54:51because the sense of loyalty matters a
54:54lot and that goes sort of a little bit
54:56to the childhood. We just discussed that
54:58like let's say if you're a Fulham fan
55:01you're not going to root for Arsenal
55:02just because they won or played in the U
55:06Champions League final. But in the
55:08United States, uh if uh Lakers are on
55:11the top, people are going to say, "Yeah,
55:12I'm I'm fan of Lakers because it's just
55:15makes it easier to start conversation."
55:17You know,
55:18>> when you think about your own CEO style,
55:22what's changed most
55:23>> in AI? It's so important to look at
55:27actual
55:29signals
55:30and actual adoption and having access to
55:34raw information. Um I was obviously
55:37taught the corporate school of
55:39management in the United States. Um and
55:42when I look at the CEOs whom um whom I'm
55:45learn from is obviously Jensen, Elon and
55:48Nick. Um Nick was on the show. So like
55:52obviously like those three are those
55:55three they completely abandon all the
55:57management principles. They don't
55:59necessarily are like fans of like
56:01one-on-one and like soft feedback. All
56:04of them I think are encouraged like
56:06being down to the points knowing the
56:08details while it would be called in like
56:11corporate America something like
56:13micromanagement.
56:14>> What management principle do you
56:16disregard that many people think is
56:18important?
56:19>> I do believe that it's as simple as hire
56:22the best people to do the best work and
56:24figure out how to retain them.
56:27Everything else is frankly secondary and
56:30people just create so much theory around
56:33that and and essentially there is just
56:35so many like fake rules uh which are
56:38disconnect from reality. It's it's
56:39really as simple as hire the best
56:41people, empower them to do the best work
56:43and just figure out how to establish
56:45relationship and retain them.
56:47>> A lot of them bluntly are do see dollar
56:50signs. We mentioned the transactional
56:53nature of America and secondaries are a
56:56part of that. How do you think about
56:58doing annual tenders to retain people
57:00>> across our team? Um roughly 50 are in um
57:05California. Uh maybe we're going to get
57:08to roughly 50 remotes and um over 300s
57:12in Kazakhstan. So look, I just hope
57:14we're going to print uh more dollar
57:17millionaires in Kazakhstan, in Central
57:19Asia, in this part of the world uh than
57:22any other company.
57:24>> I I I do too. Um what's the labor
57:28arbitrage on cost between Kazakhstan and
57:31the US?
57:32>> I I know that a lot of people when they
57:34look at Hixel, they think about the
57:35arbitrage first and foremost like the
57:38way
57:38>> is that not true? Look like Kazakhstan
57:41is top five in the world in physics.
57:44Like you look at the recent
57:46international physics olympiad for high
57:47schoolers like they're top five in the
57:49world on par with like the United
57:51States, China, India and this is also
57:53like the core of our team are people who
57:56won international competitions in math
57:58and physics. Um that's the first part.
58:02The second part is that about Kazakhstan
58:05is that they actually took this Soviet
58:06school of math but really upgraded with
58:09Singaporean principles and Singaporean
58:11system of education is considered to be
58:13probably the best in the world. At least
58:15many people in Silicon Valley believe
58:17that. Um and they and the government
58:19basically subsidizes for thousand of
58:22high schoolers to study abroad and many
58:25of these people come back. Um and there
58:28is strong desire just and so the just
58:30the density of talents uh definitely got
58:33there. It's uh like top 10 largest
58:35countries in the world. So over 20
58:37million population and we are also
58:39actively hiring bringing their talents
58:41from Europe from other countries in Asia
58:44and people just enjoy like some benefits
58:46like 15% personal income tax. Yeah man,
58:50it's like
58:52>> don't even get me started in [ __ ] UK
58:54will tax you to breathe. Uh, seriously,
58:57it's in the UK, you get your, you know,
59:00paycheck and then it's like, I don't
59:02100,000 and then you get the end and
59:05it's kind of like 3,500.
59:07>> But it's also English common law, so
59:09it's not like that bad as people think.
59:11Uh,
59:14you move it. Let's swap places. Do you
59:16have a mega pad in Kazakhstan?
59:19>> No, I don't. I don't own any property.
59:21>> What? Why?
59:23>> Remember that I come from Asian family.
59:25So um whenever we sold the company, I
59:29made over a million dollars and I spent
59:31all this money buying apartments for my
59:35parents, relatives, my wife parents cuz
59:39it's just part of the culture and the f
59:40like extended family is not small by any
59:42means. Uh but look, it's just part of
59:44the culture to give back. And then um
59:48when it comes to the family, especially
59:51to my parents, they obviously sacrificed
59:53a lot. So I I felt like I had to give
59:54back at least at least like things like
59:57monetary things which I which I could do
1:00:00but I drive like Tesla Model 3 like and
1:00:03I le so like I I'm not like a guy who's
1:00:06going to just show up with Lamborghini
1:00:08or Porsche.
1:00:09>> Do you invest? We mentioned solve
1:00:11intelligence. um when before I did that
1:00:14but now I spend roughly 90 hours a week
1:00:1980 90 hours a week on Hicksfield. I try
1:00:23to spend ideally
1:00:26um at least um 3 hours a week with my
1:00:29wife at least 5 hours a week with my
1:00:33son. Um sometimes I do the catch up
1:00:36because when I travel um for a week, for
1:00:39two weeks, for three weeks, then I try
1:00:40to take Sunday off to spend the whole
1:00:43day with my son. And over the last 3
1:00:46months, yes, I was able to find one day
1:00:47when I spent like end to end with my son
1:00:49without emails, without talking to
1:00:53without talking to the team members. I
1:00:55get in trouble for this, but I think
1:00:57there's no um shortcut to hard work. The
1:01:00harder I work, the luckier I get. I meet
1:01:02more founders. I find more great
1:01:04companies. I do more shows. I have more
1:01:06success. Do you buy the [ __ ] of the
1:01:09balance and uh oh, it's okay. You can
1:01:12leave at 5 and be home for bath time and
1:01:15crush it. This is a good question. So,
1:01:16look, obviously um being an immigrant, I
1:01:18always have that I have to prove like
1:01:20that I belong, right? So, I hope that I
1:01:23feel like now people accept people
1:01:24recognize that Hicksfield is probably a
1:01:27top 10 um application AI companies by
1:01:29revenue, probably number one. But I
1:01:32think when it comes to um hard work like
1:01:35the people whom we know in common like
1:01:38we we talked about like let's say Peter
1:01:40Salis like legend in the cons in
1:01:43consumer space obviously Jack look I I
1:01:48spent decent amount of time with them
1:01:49and other product leaders at stamp like
1:01:51the density of product talent and stamp
1:01:53was unprecedented all of them work
1:01:55really hard all of them are smart I I
1:01:59like none of them just uh checks emails
1:02:03for five hours a day and and calls it
1:02:05work. Each of them is deeply rooted into
1:02:08the recent trends in product product
1:02:10design activation. They know data really
1:02:13well. So yeah, I don't believe that
1:02:15there is any shortcut to hard work.
1:02:18>> 3 hours a week with your wife. Yeah,
1:02:22I don't know about you, dude. Mine would
1:02:24dump me for 3 hours a week. How do you
1:02:27make marriage [clears throat]
1:02:28work [laughter and gasps] on three hours
1:02:31a week?
1:02:32>> Yeah, look, I'm I'm I'm I'm very um I'm
1:02:34I'm very grateful for my wife for being
1:02:36patient, you know. It's also very
1:02:38different if that's like Asian culture.
1:02:41Uh it's just kind of more natural to try
1:02:45to do sacrifices for each other sort of.
1:02:48Um, and I'm deeply I'm obviously deeply
1:02:51grateful for her for supporting me. But
1:02:53like sometimes at this scale I get
1:02:55invited to parties. I always send her
1:02:58some and don't show up myself. I don't
1:02:59know if I piece people off, but this
1:03:02happens um very frequently.
1:03:05>> So wait, you say yes and then she goes,
1:03:07"Yeah, I say maybe we both can some come
1:03:09together." Then there is always some
1:03:11urgent fire last minutes and my wife
1:03:13just goes. [laughter]
1:03:15>> What fire was most urgent? What was the
1:03:19Oh [ __ ]
1:03:22Yeah. Look, I think obviously for all
1:03:23the things which we touch base earlier
1:03:25whenever we are not very good in
1:03:28communicating the features or we felt
1:03:30like I mean now it's like team of 40 so
1:03:33now the life is way better but early
1:03:35days obviously I was involved in all the
1:03:36fires. Um I think recently um all the
1:03:40types of like attacks on AI companies.
1:03:43It's crazy. It's like it's like LLMs are
1:03:47being used to hack companies. It's like
1:03:51new types of LLMs to do some frauds, you
1:03:54know, like basically bots using credits
1:03:57and then doing auto refunds. All of
1:03:59that. Look, I like since I have like
1:04:02kind of machine learning background
1:04:03myself, data science backgrounds, I
1:04:05still can move a needle substantially
1:04:07when it comes to statistics and data. So
1:04:09yeah, I have to be involved somehow. But
1:04:11like these LLMs, they they amplify many
1:04:14types of behaviors including various
1:04:16types of attacks and fraud, but and we
1:04:19have to fight against that. Uh we're
1:04:21going to do a quick fire around. So I
1:04:23say a short statement, you give me your
1:04:25immediate thoughts. What have you
1:04:26changed your mind on most in the last 12
1:04:29months?
1:04:30>> Oh, I was thinking that HubSpot is going
1:04:33to get obsolete. Everyone is going to
1:04:35build their own CRM and but when
1:04:37especially when as we hire and scale B2B
1:04:39go to market team just having familiar
1:04:41interface matters a lot.
1:04:44Wow. I would still say they're going to
1:04:46get [ __ ] You think that just
1:04:48stickiness is there with SMBs?
1:04:51>> Yeah, I I I do think so. And especially
1:04:53I see that when I hire go to market
1:04:55talents.
1:04:55>> Wow. Why? Like what is it about hiring
1:04:57them that makes you think that just
1:04:59they're so used to it?
1:05:00>> I mean like people who are very good in
1:05:01understanding customers and talking to
1:05:03customers they may not just simply
1:05:05accept new interface so quickly and just
1:05:07having HubSpot as a system of records
1:05:10being able if if there is any mismatch
1:05:11going able to just understand where the
1:05:14data flow went wrong. I think that's
1:05:16just still very valuable like the
1:05:17familiarity. What do you believe today
1:05:20that everyone else thinks is [ __ ]
1:05:23crazy? I mean, look, I think uh people
1:05:25just still don't fully appreciate that
1:05:27most of the content on social media is
1:05:28going to be AI generated. There are
1:05:31going to be some shows like obviously
1:05:32yours where it's like authentic
1:05:34contents. It's going to be 10 15x higher
1:05:37CPM whatsoever than AI generated
1:05:38contents. So it's going to be it's going
1:05:40to be way less in terms of like content
1:05:43create by but it's going to create way
1:05:45more value uh than a generated content.
1:05:48But even when I look into your content
1:05:50specifically like you made multiple very
1:05:54successful shorts millions of views
1:05:58better than anyone else in this space
1:06:00and you do a lot of overlay. While the
1:06:04content is authentic, I think we should
1:06:06do better job so that you use Hicksfield
1:06:08at least for the overlay on top of
1:06:11existing videos. Dude, I would love
1:06:13that. I mean, again, they take 3 hours.
1:06:16So, people don't know this. I spend 2
1:06:18hours a day just doing Instagram. Now,
1:06:20we decided that Instagram short form is
1:06:22going to be a big new push for us. Um, 2
1:06:24hours a day just for me. I write the
1:06:26scripts and then I record them. And then
1:06:28it's two people, six hours per one for
1:06:32those three.
1:06:33>> And that's extremely smart of you. Like
1:06:35you know like going back to some of the
1:06:38topics is like clipping is like a huge
1:06:41topic and that's like has its own
1:06:44upsides and downsides. But obviously
1:06:46everyone sees this opportunity to win to
1:06:49build massive top of funnel like
1:06:50hundreds of millions of views with short
1:06:52form content. as long as you can have
1:06:55downstream monetization like or value
1:06:57creation like you do.
1:06:58>> Totally agree with you. What job today
1:07:01does not exist that will be big in 5
1:07:04years? Okay. So in 5 years people
1:07:07especially in our space creative
1:07:10directors they are going to be talking
1:07:12to computers and generating stories real
1:07:15time and video and AI is going to help
1:07:17to create multiple variations. Today
1:07:20there is no word to really describe that
1:07:22because there is also there are script
1:07:23writers um there are then uh
1:07:27screenwriters like those who are going
1:07:29to break it down shot by shot then there
1:07:31are people who do that storyboarding
1:07:34then there is like people who person who
1:07:36oversees all of that like movie director
1:07:39and so on. There are so many there are
1:07:42so many there are so many parts of that
1:07:45but eventually taste is going to matter
1:07:47a lot and just having stories to tell
1:07:50there is no word to describe it today.
1:07:52>> Who do you not have on your board that
1:07:55you would most like to have on your
1:07:57board? Maybe out of like more
1:07:59professional CEOs, I'm definitely Frank
1:08:01Slutman because going back to the point
1:08:04I was just curious all the time, does no
1:08:06[ __ ] culture exist in California or
1:08:10not? Can it allow to scale companies so
1:08:13quickly? Can it is it possible to build
1:08:16successful enterprise go to market
1:08:18motion with no [ __ ] culture? And
1:08:21when I read his ampitab book like book
1:08:23called amp it up, I realized it's
1:08:25possible. So like I'm a huge fan. I
1:08:26watched all his interviews.
1:08:28>> The challenge with him, he's amazing.
1:08:30He's the best leader by far. But the
1:08:32challenge is you can sometimes do it at
1:08:35the sacrifice of product advancement.
1:08:38And so he built a GTM machine at
1:08:40Snowflake, but data bricks wiped the
1:08:43floor because they move product as the
1:08:46priority, not GTM. And that was
1:08:48dangerous. I preferred Chad Pets. Do you
1:08:52know Chad Pet?
1:08:53>> Oh, dude. This guy is no [ __ ] I'll
1:08:55introduce you afterwards. He's the best
1:08:56sales leader in the world. Um, and he is
1:09:00no [ __ ] [ __ ] Unbelievable.
1:09:02>> And we probably should have him on
1:09:03board. [laughter]
1:09:04>> Oh my god. I find any way to have him on
1:09:06board. He is terrifyingly good. Um, so
1:09:09what's the biggest lesson from Snap?
1:09:12>> The momentum doesn't last forever. Um,
1:09:14like today, Snap market cap is is below
1:09:1715 billion. There are a lot of mimis on
1:09:19the internet, but this is a great
1:09:20company. cares so much about trust and
1:09:22safety and experience and it puts it
1:09:24first.
1:09:24>> Do you think it is a great company? No
1:09:26offense. It's like it's been mismanaged
1:09:29as [ __ ] Its SBC is through the roof.
1:09:33It's tough to say it's a good company.
1:09:35>> That's why I say that momentum doesn't
1:09:36last forever. When Snapchat was worth
1:09:39eight $80 billion and the gap with Meta
1:09:43was less than 10x, then it felt, oh, we
1:09:46just go explore. we we just we just
1:09:49really must lean in. Um but momentum
1:09:51doesn't last forever and that's my core
1:09:53learning. So that's why like while we do
1:09:56have the positive momentum, we don't we
1:09:58do not take this for granted. Clearly um
1:10:00like the nature of uh capitalism is
1:10:03there are ups and downs and since we're
1:10:05building long-term, we just should
1:10:06capitalize on the opportunity like with
1:10:08the fundraising and just keep pushing
1:10:10progress every day. What is the reason
1:10:13why the divergence between Meta's market
1:10:16cap and Snap's market cap has increased
1:10:18so significantly if there was one
1:10:20reason? Just maybe saying this trait, a
1:10:24lot of public companies
1:10:26did not figure out their AI story.
1:10:30Um, Snap unfortunately is part of that.
1:10:33We have seen other great companies like
1:10:35Figma trying to tell their story. You
1:10:38mentioned Canva. It's not necessarily
1:10:41easy to be successful in private markets
1:10:43and public markets. And Zach is one of
1:10:47the best CEOs of all time cuz he managed
1:10:50that. He's such a [ __ ] beast. He's
1:10:53such a beast. You watch him last night
1:10:54with the event and you're just like,
1:10:56"Ah, [sighs]
1:10:57now I get it." Like that totally makes
1:11:00sense. And you know what? Scale with
1:11:02Alex Wang. I was one who was like really
1:11:06like what's gonna
1:11:08he he he basically acquired a second CEO
1:11:12you know Alex is now the CEO of Muse and
1:11:15he's crushed it crushed it. What an
1:11:18effective buy for 0.5% of your market
1:11:22cap. Do you know what I mean?
1:11:24>> Yeah. Look, but this happens with
1:11:25Instagram with WhatsApp. That's why I'm
1:11:27saying that we just maybe should put
1:11:30Meta a little bit in its own league.
1:11:33Yeah, but he got rid of Cyrum and Kger.
1:11:36Here he's been like, "No, no, no. You,
1:11:39Alex Wang, are my guy."
1:11:41>> Do you see what I mean?
1:11:43>> Yeah. The best talent hire.
1:11:45>> Look, I I do believe that it's a little
1:11:47bit early to look at whole Meta AI
1:11:49initiatives. We probably need to see
1:11:51like year of like successful launches
1:11:54and so on to and then we can look back
1:11:56and see what was good, what was not
1:11:57good. But at least the consistency of
1:12:00storytelling and explaining what he is
1:12:02doing to public investors being able to
1:12:05articulate why Muse is so different
1:12:09um is phenomenal.
1:12:11>> Okay.
1:12:13Revenues are a billion. What are the
1:12:16revenues in 12 months time?
1:12:19>> Our current business model uh projects
1:12:23uh 4.5.
1:12:26It says by the end of the next year but
1:12:28this basically involves substantial
1:12:31deceleration and that's what like just
1:12:33my finance team like there are couple
1:12:35strong quant people they told me that's
1:12:37just how the business works but look we
1:12:39are still pushing to grow at least 30%
1:12:42month over month what do you think it is
1:12:44they said 4.5 what do you think it is
1:12:47this is me to you not me to your finance
1:12:48team
1:12:49>> over 10
1:12:50>> over 10
1:12:52>> let me tell you why like in a lot of
1:12:56adoption in creative AI space is driven
1:12:59by monetization like all these direct to
1:13:03consumer brands making more ads and also
1:13:07having like the aspirational
1:13:11um cinematic AI content as this inspires
1:13:14creatives to explore the tooling.
1:13:18I it feels to me that Hollywood starts
1:13:23to embrace AI
1:13:27mostly today as a way to as a tool for
1:13:32hybrid production as a just new form of
1:13:37CGI.
1:13:40But the sentiment really shifted from
1:13:42like strictly negative
1:13:45to neutral to slightly negative. And in
1:13:48private conversations, yes, there are
1:13:51maybe more than half of sier talents who
1:13:53is going to say we're anti-AI forever.
1:13:56But increasingly there are more and more
1:13:58people who are actually asking a
1:14:01question. Can we tell more stories with
1:14:03AI? Can we overcome certain budget
1:14:06limitations which existed before? And
1:14:08maybe AI can help to tell new stories
1:14:10which we couldn't tell before. And I do
1:14:12believe this just change in perception
1:14:14that's at least comes from my
1:14:16conversations is extremely is extremely
1:14:18is extremely positive. If you are in a
1:14:21billion today, 10 billion in 12 months.
1:14:25Where do you peg the next fund raise?
1:14:27You know, if you're in a billion, say
1:14:29conservative multiple, you'd be like,
1:14:31you know, 15.
1:14:33Um but if you're hitting 10 next year
1:14:35you're like paying end of year it's like
1:14:3980. Look we are not chasing just the
1:14:41valuation cuz again the goal is just to
1:14:43make sure that the company can be uh
1:14:46sustainable over the time in public
1:14:48market. So there is a lot of company
1:14:49building to be done beyond just uh
1:14:51chasing the revenue. But I just do
1:14:53believe
1:14:53>> do you want to be public? Huh?
1:14:55>> Do you want to be public at some point?
1:14:56Yeah, I do believe that Hixfield has
1:14:58great potential to be bigger than
1:14:59Applain and Shopify because
1:15:01fundamentally like building is one part
1:15:03of that Shopify one layer of
1:15:05infrastructure. Then for coding there is
1:15:07obviously like a cloud,
1:15:10there is codeex but what matters is
1:15:12distribution over the time but
1:15:13distribution matters. You know that this
1:15:15better than any other
1:15:17>> business that's why we do what we do.
1:15:19>> Yeah,
1:15:20>> exactly.
1:15:21>> Dude, I cannot thank you enough for
1:15:22being so amazing on the show. You've
1:15:24been fantastic. I've loved doing it, you
1:15:26can tell. And you've been an amazing
1:15:28guest. So, I really appreciate you
1:15:29joining me today.
1:15:30>> Thank you so much. It's a pleasure.

About this transcript

This page contains the full transcript of The Untold Story of Higgsfield | Burning $4M a Month on AI Models | CEO, Alex Mashrabov by 20VC with Harry Stebbings, generated from the public captions YouTube serves with the video. The transcript has 12,584 words across 1,855 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.