[00:00] every 3 months things have just kept
[00:02] getting progressively better and now
[00:04] we're at this point where we're talking
[00:05] about full-on vertical AI agents that
[00:07] are going to replace entire teams and
[00:09] functions and Enterprises that
[00:11] progression is still mind-blowing to me
[00:13] a lot of the foundation models are kind
[00:15] of coming head-to-head there used to be
[00:17] only one player in town with open AI but
[00:19] we've been seeing in the last batch this
[00:23] has been changing thank God it's like
[00:26] competition is you know the the soil for
[00:28] a very fertile Market Marketplace
[00:30] ecosystem uh for which consumers will
[00:33] have choice and uh Founders have a shot
[00:36] and that's the world I want to live
[00:39] [Music]
[00:44] in welcome to another episode of the
[00:47] light cone I'm Gary this is Jared Harge
[00:50] and Diana and collectively we funded
[00:52] hundreds of billions of dollars worth of
[00:55] startups right when they were just one
[00:57] or two people starting out and and today
[01:01] Jared is a man on fire and he's going to
[01:04] talk about vertical AI yes I am I am
[01:09] fired up about this because I think
[01:11] people especially startup Founders
[01:15] especially young ones are not fully
[01:17] appreciating just how big vertical AI
[01:20] agents are going to be it's not a new
[01:22] idea some people are talking about
[01:23] vertical AI agents we funded a bunch of
[01:25] them but I think the world has not
[01:27] caught on to just how big it's going to
[01:28] get and so I'm going going to make the
[01:31] case for why I think there are going to
[01:33] be
[01:34] $300 billion plus companies started just
[01:38] in this one category nice I'm going to
[01:40] do it by analogy with SAS and I
[01:43] think in a in a similar fashion people
[01:46] don't understand just how big SAS is
[01:49] because most startup Founders especially
[01:51] young ones tend to see the startup
[01:53] industry through the lens of the
[01:54] products that they use as a consumer and
[01:56] as a consumer you don't tend to use that
[01:58] many sass tools because they're mostly
[01:59] built for companies and so I think a lot
[02:02] of people have missed the basic point
[02:04] that if you just look at what Silicon
[02:06] Valley has been funding for the most for
[02:08] like for the last 20 years like we've
[02:10] mostly been producing SAS companies guys
[02:12] like that's literally been like most of
[02:14] what has been coming out of Silicon
[02:16] Valley it's over 40% of all venture
[02:18] capital dollars in that time period went
[02:21] to SAS companies and we produced over
[02:23] 300 SAS unicorns in that 20-year time
[02:26] period which is way more than every
[02:28] other category software is pretty
[02:30] awesome software is pretty awesome I was
[02:32] thinking back to the history of this
[02:35] because we we always like to talk about
[02:37] the sort of how the how the history of
[02:39] Technology informs the future and um the
[02:42] the real Catalyst for for the SAS boom
[02:45] was a do you guys remember XML HTTP
[02:48] request oh my God like I I'd argue that
[02:51] that was quite literally the Catalyst
[02:53] for the S boom like uh Ajax Ajax yeah in
[02:57] 2004 browsers added this JavaScript
[02:59] function XML HTTP request which was the
[03:01] missing piece that enabled you to build
[03:03] a rich internet application in a web
[03:04] browser so for the first time you could
[03:06] make things in websites that looked like
[03:08] desktop applications and then that
[03:10] created Google Maps and Gmail and set up
[03:12] this whole like SAS boom essentially the
[03:15] the key technology atlock was that
[03:18] software moved from being a thing that
[03:20] you got on a CD ROM and installed on
[03:22] your desktop to being something that you
[03:23] use through a website and on your phone
[03:25] yeah Paul Graham actually uh shares in
[03:28] that lineage in that he was one of the
[03:29] first people to realize that he could
[03:31] take the HTTP request and then actually
[03:34] hook it up to a Unix prompt and you
[03:37] didn't actually have to you know have a
[03:40] separate computer program that would
[03:43] change a website so via web was a online
[03:46] store kind of like Shopify but way back
[03:48] in the day yeah it was basically like
[03:50] the first SAS app ever like like PG
[03:52] actually invented SAS in like 1995 it's
[03:55] just that those first SAS apps kind of
[03:57] suck because they didn't have XML HTTP
[03:59] request and so every time you would like
[04:00] click a button you would have to reload
[04:02] the whole page and so it's just a shitty
[04:04] experience and so it didn't really catch
[04:05] on until 2005 when X XML HTP request
[04:08] white spread anyway I I see this llm
[04:11] thing as like actually very similar um
[04:14] it's like it's a new Computing Paradigm
[04:16] that makes it possible to just like do
[04:17] something fundamentally different and in
[04:20] 2005 when cloud and mobile finally took
[04:23] off there is this sort of like big open
[04:25] question of like okay well this new
[04:27] technology exist what should you do with
[04:29] it where is the value going to acrw
[04:32] where are the good opportunities for
[04:33] startups I was going through the list of
[04:35] like all the billion dollar companies
[04:36] that were created and I kind of had this
[04:38] realization that um you could kind of
[04:41] bucket the the different paths that
[04:43] people took into like three buckets um
[04:46] there's there's a first bucket that
[04:48] people started with which was like I
[04:50] would call them Obviously good ideas
[04:54] that could be Mass consumer products um
[04:57] so that's like docs photos email
[05:00] calendar chat all these things that like
[05:03] we used to do on our desktop with that
[05:05] obviously could be moved to the browser
[05:06] and mobile and the interesting thing is
[05:10] zero startups won in those categories
[05:12] 100% of the value flow to incumbents
[05:15] right like Google Facebook Amazon they
[05:17] own all all those businesses folks
[05:19] forget that like Google Docs wasn't the
[05:21] only company that tried to bring
[05:23] Microsoft Office online there were like
[05:24] 30 companies that tried to bring
[05:26] Microsoft Office online but they all
[05:27] lost Google one then there was a second
[05:30] category which was like Mass consumer
[05:33] ideas that were not obvious that nobody
[05:36] predicted um that's like uber instacart
[05:39] door Dash
[05:41] coinbase th Airbnb those ones those ones
[05:44] came out of left field like the the dot
[05:47] dot dot between XML HTTP request and
[05:49] Airbnb is like very not obvious yeah and
[05:52] so the incumbents didn't even try
[05:54] competing in those spaces until it was
[05:55] like too late and so startups are able
[05:57] to win there and then there's a third
[06:00] category which is all the B2B SAS
[06:02] companies and that's like 300 of them
[06:04] and so like Mo like by by by number of
[06:08] logos way more billion dollar companies
[06:10] were created in that third category than
[06:12] the first two I think one reason why
[06:14] that happened is like there is
[06:16] no like Microsoft of SAS like there is
[06:19] no company that somehow does like SAS
[06:22] for like every vertical and every
[06:23] product like for structural reasons it
[06:25] seems to be the case that like they're
[06:27] all different companies and that's why
[06:29] there so many of them I think Salesforce
[06:31] is probably like the first true SAS
[06:33] company um and i' I remember Mark benov
[06:37] coming to speak at YC and he tells the
[06:40] story as just very early on people just
[06:42] didn't believe you could build
[06:43] sophisticated Enterprise applications
[06:46] like over the cloud or via SAS it was
[06:48] just so um there was just like a
[06:50] perception issue right it was like no
[06:51] like you don't you buy like your box
[06:53] software and that's like the real
[06:54] software that you run the way we always
[06:56] do it it was it was quite contrar cuz
[06:58] the early web app sucked they were like
[07:00] via web where you had to be a Visionary
[07:02] like PG and understand that the browser
[07:03] was going to keep getting better and
[07:04] that eventually it' be good which feels
[07:06] like quite reminisent of today right
[07:08] where it's like the yeah the same thing
[07:10] like oh no like you won't be able to
[07:11] build like sophisticated Enterprise
[07:13] applications that use these llm or AI
[07:16] tools because they hallucinate or
[07:18] they're not perfect or they um they kind
[07:20] of like just toys but yeah that's like
[07:22] the early SAS story exactly the same and
[07:25] so when I think about the parallels with
[07:27] LMS I could easily imagine the same
[07:30] thing happening which is that there's a
[07:31] bunch of categories that are like Mass
[07:33] consumer applications that are obviously
[07:35] huge opportunities but probably the
[07:38] incumbents will win all of those so
[07:39] that's something like a per like a
[07:41] general purpose AI Voice Assistant that
[07:43] you you know you can ask it to do
[07:45] anything and it'll like go do that thing
[07:46] that's an obvious thing that should
[07:47] exist but like all the big players are
[07:49] going to be competing to be that thing
[07:51] right W Apple's a little slow on that
[07:53] one why is Siri so stupid still what
[07:55] year is it it makes no sense I mean it's
[07:58] like a count to that is like the very
[07:59] obvious thing is search and maybe Google
[08:02] will still win um on search but
[08:05] perplexity is definitely give them run
[08:08] for the money right yeah this is the
[08:09] classic innovators dilemma at the end of
[08:11] the day I mean you could argue going
[08:12] back to what you said about Uber or
[08:14] Airbnb these were actually really risky
[08:17] things from a regulatory standpoint so
[08:20] if you're Google and you have basically
[08:22] a guaranteed you know giant a pot of
[08:25] gold that you know sort of comes to you
[08:27] every single month like why would you
[08:29] endanger that pot of gold to sort of
[08:31] pursue these things that uh might be
[08:33] scary or might ruin the pot of gold I
[08:36] think that's I think that's like
[08:37] probably the primary reason why the
[08:38] incumbents didn't end up building those
[08:41] products and didn't even clone them even
[08:42] after they got big and it was obvious
[08:44] that they were going to work who will
[08:45] never launch an an Uber clone they never
[08:47] launch an Airbnb clone um I was
[08:50] listening to this uh talk by Travis and
[08:53] one of the things that he said that
[08:54] really stuck with me is that in the in
[08:56] the first years of uber he was very
[08:58] scared that he would was going to
[08:59] personally go to prison for like a long
[09:01] time like he was actually personally
[09:03] risking going to prison in order to
[09:05] build that company and so yeah no highly
[09:07] paid Google exactly was going to do that
[09:09] what do you think about um why the
[09:12] incumbents didn't go into B2B SAS is it
[09:15] part of the reason is that a lot of the
[09:17] use cases are very there's a very wide
[09:20] distribution I it's a great question I
[09:22] love to hear what you guys think my take
[09:24] is that it's just too hard to do that
[09:28] many things as a company like each B2B
[09:31] SAS company really requires like the
[09:33] people who are running the product in
[09:34] the business to be extremely deep in one
[09:37] domain and care very deeply about a lot
[09:39] of really obscure issues you know like
[09:42] take like Gusto for example like why
[09:44] didn't Google build a Gusto competitor
[09:45] well there's no one to Google who really
[09:47] understands payroll and has the patience
[09:48] to like deal with all the nuances of all
[09:50] these like stupid payroll regulations
[09:52] and like it's just like like it's just
[09:55] not worth it for them it's easier for
[09:57] them to just focus on like a few really
[09:59] huge categories in the B2B SAS world
[10:02] it's it's sort of about the unbundling
[10:04] bundling of software argument that comes
[10:06] up a lot as well I think and why didn't
[10:09] why did all these vertical B2B SAS
[10:11] products evolve versus just like Oracle
[10:14] or sap or um netet yeah netu just owning
[10:18] like everything um and I think it might
[10:21] be Al is another thing that's
[10:23] attributable to the shift to like SASS
[10:25] and the internet is in the old ways of
[10:28] selling software again like you had this
[10:29] box software that was really like
[10:31] expensive to install and you had like a
[10:34] whole ecosystem around it and anytime
[10:35] you wanted something custom like the
[10:38] integrators would just say oh no like we
[10:40] can like just build a UA custom like
[10:41] payroll feature or something like that
[10:44] and then Salesforce comes along with
[10:46] like a SAS solution and it just seems
[10:48] like it could never be as powerful or
[10:50] sophisticated as like the expensive
[10:52] Enterprise installation you just paid
[10:54] for but they Prov that it totally was
[10:56] the case and I think that just like
[10:58] opened the gates for all all of these
[11:00] like vertical sash solutions to emerge
[11:02] doing exactly what you're saying the
[11:04] other problem is that with a lot of this
[11:05] enterprise software if you're a user of
[11:08] Oracle and a netw suite because they're
[11:10] they have to cover so much ground the
[11:13] user experience is actually pretty bad
[11:15] they're trying to be jack of all trades
[11:17] but master of none yeah so ends up being
[11:19] a bit of a kitchen sync type of
[11:21] experience and this is where if you go
[11:24] and build a B2B SAS vertical company you
[11:27] could do literally a 10x better
[11:29] experience and more delightful because
[11:31] there's this Stark difference between
[11:33] consumer products and Enterprise user
[11:35] experience yeah well there's only uh
[11:37] what three price points in software it's
[11:40] uh $5 per seat $500 per seat or $55,000
[11:45] per seat and uh that Maps directly to
[11:48] Consumer SMB or Enterprise sales and
[11:51] then I think time in Memorial has taught
[11:54] us that in the past and this is less and
[11:56] less true uh with new software
[11:58] thankfully
[12:00] but Enterprise is terrible software
[12:02] because it's not the user buying it you
[12:05] know some high up Mucky muuk inside
[12:08] Fortune 1000 is the person who's getting
[12:10] whin and D for this you know Mega seven
[12:13] figure contract and you know they're
[12:16] going to choose something that maybe
[12:18] isn't that good actually for the end
[12:20] user the person who has to actually use
[12:22] the software day-to-day and um I'm sort
[12:25] of curious to see how this changes with
[12:28] llms actually I mean to date one of the
[12:31] more Salient things that we've seen for
[12:33] both SMB and enterprise software
[12:35] companies is that or all software
[12:37] companies all startups period is like
[12:40] you know there's a sense that as Revenue
[12:42] scales the number of people you have to
[12:44] hire scales with it and so when you look
[12:47] at unicorns uh even in today's YC
[12:51] portfolio uh it's quite routine to see a
[12:54] company that reached a hundred or $200
[12:56] million a year in Revenue but they have
[12:58] like 500 a th000 2,000 employees already
[13:02] and I'm just going to be very curious
[13:04] like uh even the advice that I'm
[13:06] starting to give companies that are you
[13:08] know a month or two out of the batch uh
[13:11] it's a it's feeling a little bit
[13:13] different than the kind of advice I
[13:14] would give last year or two years ago in
[13:17] the past you might say you know let me
[13:20] find the absolute smartest person uh in
[13:23] all of these other parts of the org like
[13:26] customer success or sales or different
[13:28] things like that
[13:29] and uh I want to find someone who I've
[13:31] worked with who is I know is great and
[13:33] then I'm going to go sit on their you
[13:35] know uh on their doorstep until they
[13:37] quit their jobs and come work for me and
[13:39] I want them to be someone who can you
[13:41] know build a team for me hire a lot of
[13:43] people that might still be true but I'm
[13:45] starting to sense that uh the Met is
[13:48] shifting a little bit like you actually
[13:50] might want to hire more really good
[13:53] software Engineers who understand large
[13:56] language models uh who can actually
[13:59] automate the specific things that you
[14:01] need that are the bottlenecks to your
[14:02] growth and so it might result in you
[14:06] know a very subtle but you know
[14:08] significant change in the way startups
[14:10] grow their businesses sort of post-
[14:11] product Market fit it means that I'm
[14:14] going to build llm systems that bring
[14:16] down my costs that cause me not to have
[14:19] to hire a thousand people I think we're
[14:21] right at the beginning of that
[14:22] Revolution right now I mean we talked
[14:24] about this in a previous episode we
[14:26] talked about there will be a future
[14:29] unicorn company that's only run if we
[14:31] take it to the limit with only 10
[14:33] employees that's completely plausible
[14:36] and they're writing the evals and the
[14:37] prompts does it I think what you're
[14:39] saying is like a trend that was already
[14:40] underway pre llms like I remember when I
[14:43] was running triple bite for example we
[14:46] needed to like build marketing or cust
[14:49] like user acquisition basically um and
[14:51] especially after we raised a series B
[14:54] the like traditional way you were
[14:55] supposed to do that is to like hire a
[14:57] marketing executive and build out like
[14:59] marketing team and um and just like
[15:01] basically spin up this machine to do
[15:03] like sales and marketing but I'd
[15:06] actually met like a y founder um Mike
[15:10] who was his company was basically
[15:11] building like a smart frying pan sounds
[15:13] like bizarre but like he was a MIT
[15:15] engineer yeah you remember this um he's
[15:17] an MIT engineer and to sell the smart
[15:19] frying pan he had to get really really
[15:21] good at understanding like paid
[15:22] advertising and um uh Google ads and
[15:25] just a whole bunch of stuff and so he he
[15:26] taken this Engineers mindset approach to
[15:28] it and remember just talking to him
[15:29] about it and realizing this would be so
[15:32] much better to have an MIT engineer
[15:34] working on like our marketing efforts
[15:37] than any of the marketing candidates
[15:39] I've spoken to and he was able to like
[15:41] scale us up to like we were spending
[15:43] like 1. like a million dollars a month
[15:45] on just marketing and various like
[15:47] campaigns and triple bite had great
[15:49] marketing like I remember like the Cal
[15:51] trained station takeover that you did
[15:53] all the like out of home stuff that you
[15:55] did it was like really high quality
[15:57] stuff it stuck with it you could tell
[15:59] was not being done by some like VP
[16:00] marketing person um and that was all mik
[16:03] and like the comment I would often get
[16:05] when people would ask me around that
[16:07] time like how big is triple bite and we
[16:08] were like 50 people and so much yeah
[16:11] yeah people be like I thought this's
[16:12] like hundreds of people I was like no
[16:14] it's all because if you put a really
[16:15] smart engineer on some of these like
[16:17] tasks they just find ways to make they
[16:20] find leverage and now like llms can go
[16:22] even Way Beyond like The Leverage you
[16:23] had which is pure software okay so
[16:25] here's my pitch for 300 vertical AI
[16:28] agent unicorns literally every company
[16:31] that is a SAS unicorn you could imagine
[16:33] there's a vertical AI unicorn equivalent
[16:36] in like some new universe cuz like most
[16:39] of these SAS unicorns beforehand there
[16:42] were some like box software company that
[16:44] was making the same thing that got
[16:45] disrupted by a SAS company and you could
[16:47] easily imagine the same thing happening
[16:49] again where now basically every every
[16:51] SAS company builds some software that
[16:53] some group of people use the vertical AI
[16:55] equivalent is just going to be the
[16:57] software plus the people in one product
[17:00] one thing might be just Enterprises in
[17:01] general right now are a little unsure
[17:03] about what exactly they like what agents
[17:05] they need and one approach I've seen
[17:07] from especially more experienced
[17:08] Founders like um Brett Taylor the CTO of
[17:11] Facebook started his company Sierra I
[17:13] don't know all the details but as far as
[17:15] I can tell it's essentially more like
[17:17] broadly about letting Enterprises like
[17:20] deploy these AI agents and spinning them
[17:23] up like custom for the Enterprise versus
[17:26] like oh hey we have like this specific
[17:27] agent to do this it's something I've
[17:29] seen from one of my companies called um
[17:31] uh Vector shift that funded about a year
[17:33] ago they're two really smart like
[17:36] Harvard computer scientists and it's a
[17:39] that what they found is that they're
[17:41] trying to build a platform to make it
[17:42] easy for Enterprises to build their own
[17:44] like use like no code or sdks to build
[17:47] their own like um internal llm powered
[17:50] agents but like Enterprises often don't
[17:53] know exactly what they want to use these
[17:54] things for and so bring it back I wonder
[17:57] if like in like the box software world
[17:59] you started off with just like a few
[18:01] vendors who just basically were trying
[18:02] to convince people to use software at
[18:04] all and it was just like it does
[18:06] everything um and then it gets more
[18:08] sophisticated and higher resolution and
[18:09] you get lots of like vertical SS players
[18:12] we go through that same period with llms
[18:14] where the early winners might just be
[18:16] these like general purpose hey like we
[18:18] like make it easy for you to do llm
[18:20] stuff and then it the vertical agents
[18:22] will come in over time or do you think
[18:25] there's reasons it's different now and
[18:26] the vertical agents will take off on day
[18:28] one
[18:29] yeah that's interesting because if you
[18:30] think about the history of SAS the
[18:32] consumer things worked first like 2005
[18:35] to 2010 was mostly consumer applications
[18:38] like email and chat and maps and people
[18:41] got people as individuals got used to
[18:43] using these tools themselves and I think
[18:45] that made it easier to sell SAS tools to
[18:48] companies because you know the same
[18:49] people are both employees and consumers
[18:51] yeah I I think the answer might just be
[18:53] like this is this is all just a
[18:54] continuation of software and just
[18:58] there's no reason it has to reset back
[19:00] like llms don't have to reset back to a
[19:02] few general purpose like Enterprise llm
[19:05] platforms doing everything because
[19:07] Enterprises have already been trained on
[19:08] like the value of Point Solutions and
[19:11] vertical Solutions um and like the user
[19:14] experience not going to be that
[19:15] different these things will just be a
[19:16] lot more powerful and so if Enterprises
[19:19] have already built the muscle of
[19:21] believing that like startups or vertical
[19:23] Solutions can be better than like Legacy
[19:25] broad platforms they are probably going
[19:28] to be willing to take a bet on a startup
[19:31] promising a very good vertical AI agent
[19:33] solution today and I feel like we're all
[19:34] seeing that in the batch now with some
[19:36] of our companies are getting faster
[19:39] Traction in Enterprises for these
[19:41] vertical AI agents than like we've ever
[19:44] seen before I think we're just early in
[19:45] the game right like all software sort of
[19:47] starts quite vertical and then as the
[19:50] industries actually get much more
[19:52] developed um then I mean I just answered
[19:56] my earlier question it's like you know
[19:58] why does company end up having a
[19:59] thousand employees it's actually that uh
[20:02] you know early early in the game
[20:04] everyone's making these specific point
[20:06] Solutions and then at some point you've
[20:08] got to go horizontal like you're already
[20:11] doing this crazy spend on sales and
[20:13] marketing and then the only way you can
[20:15] actually continue to grow once you sort
[20:18] of get 100% or you know some large
[20:20] majority of the market is you actually
[20:22] have to do like not just a point
[20:25] solution but things that sort of work
[20:27] together the other point of why the bull
[20:32] case for vertical AI agents could be
[20:33] even bigger than SAS is that SAS you
[20:36] still needed a operations team or set of
[20:39] people to operate the software in order
[20:41] to get all the workflows to be done I
[20:43] don't know approval workflows or you
[20:44] have to input the data the argument here
[20:47] is that you will get not only replacing
[20:50] all that set of SAS software so that
[20:52] would be like one to one mapping but is
[20:54] also going to eat all of the a lot of
[20:56] the payroll because we look a lot of the
[20:58] spend for companies big chunk is still a
[21:00] payroll and software's Tiny exactly they
[21:02] spend way more on employees than they do
[21:04] on software so it'll be these smaller
[21:06] companies that way more efficient that
[21:07] need way less humans to do random data
[21:11] entry or approvals or click the software
[21:14] I agree I think it's very possible the
[21:16] vertical equivalence will be 10 times as
[21:18] large as the SAS company that they are
[21:20] disrupting I mean there there's two case
[21:22] it could be that the vertical Point
[21:24] solution could be just big enough and
[21:26] you don't need to do that bro breath
[21:28] thing right it that could be a nice
[21:30] scenario should we give some examples I
[21:32] feel like we've all been working with so
[21:34] many vertical AI agent companies we've
[21:37] got like news from the
[21:38] front how it's actually going well your
[21:41] former uh head of product Aaron Cannon
[21:44] is working on a YC company called outset
[21:46] that I worked with and uh basically
[21:48] they're taking llms uh to the surveys
[21:51] and qual Trix space so qual Trix is
[21:54] almost certainly not really going to
[21:56] build the best of breed uh large
[21:58] language model with reasoning and then
[22:00] the funny thing about surveys is you
[22:02] know who's it actually for it's for
[22:04] people who run products for marketing
[22:06] teams it's for people who are trying to
[22:08] make sense of like what do our customers
[22:10] actually want and what are surveys like
[22:12] guess what that's language so um and
[22:15] then I feel like these types of
[22:18] businesses um actually have to thread
[22:20] this needle um because Enterprise and
[22:23] SMB software often is sold based on a
[22:27] particular person who who is the key
[22:29] decision maker and um you have to go
[22:32] high enough in the organization so that
[22:34] the people you're selling to are not
[22:36] afraid that their whole their job Andor
[22:38] their whole team's job is going to go
[22:40] away totally that's kind of the move
[22:42] that I seen that a lot of companies that
[22:45] sell need to do because if you're going
[22:46] to go and sell to the team that's going
[22:48] to get replaced by AI they're going to
[22:50] sabotage it man it just does not work so
[22:53] I think this is an interesting way that
[22:56] a lot of these are top down and you have
[22:58] to go through at some point even get the
[22:59] CEO to sign off on it a company I'm
[23:02] working with u MCH that's sort of
[23:04] essentially an AI agent but for at least
[23:06] where they're starting is like QA
[23:08] testing um they're getting really great
[23:11] traction right now and it's interesting
[23:13] because you remember a decade ago um why
[23:15] can't we worked with rainforest QA like
[23:17] rainforest was a QA as a service company
[23:21] and that they had this exact tension of
[23:24] where they couldn't actually replace
[23:26] your QA team and so they needed to build
[23:29] software that made the QA te more like
[23:31] efficient but really that obviously
[23:33] meant trying to replace as many of them
[23:34] as possible they couldn't replace the
[23:35] whole team and so they were always on
[23:39] the sort of like tight rope between
[23:40] trying to sell the software to like the
[23:42] head of engineering as like this will
[23:43] mean you'll need less QA people and
[23:47] great but then you also have to go sell
[23:48] that to the QA team who don't want to be
[23:49] replaced and so I think that was always
[23:51] like a friction for that business for
[23:53] how it could like scale and grow but now
[23:56] like mtic with AI can actually just
[23:59] replace the QA people so their pitch is
[24:01] not oh this like makes your QA people
[24:02] faster it's like this just means you
[24:04] don't need a QA team at all so they can
[24:06] just focus the sell onto like
[24:07] engineering and Engineering doesn't need
[24:09] buying from QA at this point and you can
[24:12] also go in I mean to start with you can
[24:14] go and sell to companies that don't even
[24:16] have big QA teams at the moment they
[24:18] just use something like mtic and then it
[24:19] will just like keep scaling with them
[24:21] scaling and they'll just never build a
[24:22] QA team ever yes that is a real life
[24:25] case study of what Diana was saying
[24:26] about why these vertical AI agent
[24:28] companies going to be 10 times as big as
[24:30] the SAS companies yeah I'm seeing this
[24:31] interesting now um like in recruiting
[24:33] too I had this exact same issue with
[24:35] triple vet where to build the software
[24:38] um to build software that makes it easy
[24:40] to like screen and hire software
[24:41] Engineers you need buying from both the
[24:43] engineering team that they're joining
[24:45] but also the recruiting team and
[24:47] effectively the software we were
[24:48] building was trying to replace the
[24:49] recruiters but we couldn't completely
[24:51] replace the recruiters but now with NYC
[24:54] and so the recruiters were always like
[24:57] oppos
[24:59] opposing it CU it was a threat to them
[25:01] yeah so it just always like friction on
[25:03] like how um on like how far you can get
[25:06] when the customer you're trying to sell
[25:08] to is worried about being replaced um
[25:11] but yeah I think now it's still early
[25:13] days but now with AI you can build
[25:15] things that do the whole stack like of
[25:18] recruiting we have a company we worked
[25:20] with last batch like Nico work with them
[25:22] a priora which is actually just doing
[25:23] like the full like technical screen the
[25:25] full initial recruiter screen and
[25:27] getting great traction so I think as
[25:29] those things keep going like they won't
[25:31] have they have the same thing you won't
[25:32] have the friction of oh I need to
[25:33] convince recruiters to use this you're
[25:35] probably just like not build a
[25:37] recruiting team in the same way that you
[25:39] used to I mean other example is even for
[25:43] de tool companies they have to do a lot
[25:46] of a developer support and I work with
[25:49] this company called cap. AI that
[25:51] basically buildt one of the best chat
[25:54] Bots that responds to a lot of the a lot
[25:58] of the technical details that are hard
[26:01] to answer and I think a lot of the
[26:03] companies that started using them they
[26:05] actually ended up having Dev rail teams
[26:08] that are a lot smaller because it
[26:10] ingests a lot of uh the developer
[26:12] documentations even the YouTube videos
[26:14] that Dev tools put up and even a lot of
[26:17] the chat history so it just keeps
[26:19] getting better and better and it's like
[26:22] gives really good answers actually it's
[26:23] one one of the best I've seen yeah I
[26:26] also worked with a customer support like
[26:28] an AI customer support agent company
[26:30] called Power help well actually we both
[26:32] did um last batch and I learned a couple
[26:36] interesting things from parel um the
[26:39] first is customer like AI agents for
[26:42] customer support was like the category
[26:45] that's like famously crowded where
[26:46] there's like supposedly like you know a
[26:48] 100 of them and if you go and you Google
[26:49] like AI customer support agent you'll
[26:52] get like a 100 results on Google um but
[26:54] what I learned through working with
[26:55] parel is like it's actually kind of
[26:57] like like almost all of those
[27:00] companies are doing very simple like
[27:03] zero shot llm prompting that can't
[27:05] actually replace a real customer support
[27:07] team that does a lot of really
[27:08] complicated workflows it just kind of
[27:10] makes for like a nice demo like to
[27:12] actually replace a customer support team
[27:14] for like an at Scale company that has
[27:16] like 100 customer support reps that do
[27:18] lots of complicated things every day you
[27:19] need like really complicated software
[27:21] that does all the stuff that like Jake
[27:22] heler was talking about and there's
[27:24] there were only like three or four
[27:26] companies that were even attempting to
[27:27] do that
[27:28] and cumulative they had cumulatively
[27:30] they had like less than 1% Market
[27:32] penetration and so the market was just
[27:33] completely open I could also see that
[27:35] being another case of um hyper
[27:37] specialization or hyper verticalization
[27:40] like there's not going to be I mean
[27:42] maybe eventually there could be a single
[27:45] general purpose customer support agent
[27:47] software company but we're like in in
[27:50] you know that that'll be like a eighth
[27:52] or ninth inning kind of thing and we're
[27:54] literally in the first inning so you
[27:56] know instead you know you're going to
[27:58] have companies like gig ml that you know
[28:00] it's doing it for zepto doing 30,000
[28:03] tickets uh every single day and
[28:06] replacing a team of a thousand people
[28:08] and but it's very specific and it has
[28:11] you know it's not a general purpose demo
[28:14] Weare kind of thing like it's 10,000
[28:16] test cases in a very detailed uh eval
[28:19] set that you know is basically just for
[28:22] zepto and things like zepto yeah uh but
[28:25] if you are you know any of the other
[28:28] Marketplace companies you're probably
[28:30] going to use it cuz like that's a very
[28:32] well-defined kind of marketplace that's
[28:34] you know instant delivery Marketplace I
[28:36] think this is the kind of dynamic that
[28:38] led there to be like $300 billion do SAS
[28:40] companies rather than like one like1
[28:42] trillion do like meta SAS thing that
[28:44] provides all the software for the world
[28:45] it's just like the customers just
[28:47] require really heavily like tailored
[28:50] Solutions and it's hard to build one
[28:51] that like works for every everyone
[28:53] exactly I mean we already gave three
[28:54] examples of customer support but there
[28:55] are very different verticals it's like
[28:57] de tool comp need very different kind of
[28:59] support that you need and the training
[29:01] set to marketplaces very different right
[29:04] yeah I guess whether you have agents or
[29:06] real human beings working for you you
[29:08] end up with the same problem which is
[29:10] every company bumps up against CO's
[29:13] theory of the firm which says that any
[29:15] given firm will grow only so much to the
[29:18] point where it uh becomes inefficient to
[29:21] be larger than that and then that's why
[29:23] they sort of networks and ecosystems and
[29:27] you know a full blown economy you know
[29:29] like every firm will sort of specialize
[29:31] to do what it is particularly good at
[29:34] and then the limits the outer limits of
[29:36] what those firms can be it's actually
[29:38] based on uh your ability as a manager so
[29:42] yeah that that part a little bit breaks
[29:44] my brain because you know when we spend
[29:46] time with Parker Conrad at ripling uh
[29:49] one of his favorite points is actually
[29:50] well you know everyone's very obsessed
[29:52] with with the fact that the rocks can
[29:55] talk and you know maybe they can draw
[29:57] but the more interesting thing for him
[29:59] you know running HR It software that uh
[30:02] you know he spends a lot of time
[30:04] thinking about HR like actually the
[30:05] coolest thing about the LMS is that the
[30:07] rocks can read and from his perspective
[30:11] like you he's I think he has 3,000
[30:13] employees he still runs payroll for all
[30:16] 3,000 employees through Rippling so I
[30:19] think he spends a lot of time thinking
[30:20] about like how can one person extend
[30:22] their ability as a manager and uh I
[30:26] think we're going to see a lot more
[30:27] there
[30:28] that would be an a reverse argument that
[30:31] if we're at this moment where uh tools
[30:34] for managers and CEOs are going to get
[30:37] much more powerful um oh it could it
[30:40] could it could increase the scale of the
[30:42] firm that you can run right and that's
[30:44] certainly what ripling is trying to do
[30:45] like he's attempting to build this like
[30:47] Suite of HR tools where if he wins he's
[30:49] going to eat a whole bunch of billion
[30:51] dollar SAS companies and like one one
[30:53] giant company it's very interesting
[30:54] point Gary I think what made me think
[30:57] about this is that with having all these
[30:59] AI SAS tools it's going to give the
[31:03] ability to all these leaders and all
[31:06] these Orcs to basically open the
[31:08] aperture of the context window of how
[31:10] much information they can parse because
[31:12] there a limit of how much uh humans we
[31:14] can have meaningful relationship there's
[31:16] like the whole thing with the D Mar
[31:17] number it's about 300 people that you
[31:20] 150 that you can have a meaningful
[31:22] relationship with but with AI because
[31:25] all of these rocks now can read I think
[31:27] think we will be able to extend that
[31:30] dumbar limit that we have yeah I think
[31:32] uh Flo crell had this interesting post
[31:35] on Twitter that went viral around um I
[31:38] think someone had made a voice chat like
[31:41] just weekend project as a CEO but it
[31:44] would call uh all
[31:46] 1,500 of their employees yeah and uh you
[31:50] know it was you know very short call
[31:52] like kind of sounded like it was from
[31:53] the CEO just asking kind of personally I
[31:56] mean it sort of reminds of um that scene
[31:59] in her where it zooms out and uh
[32:02] actually you know you're following the
[32:04] experience of one person using the her
[32:06] OS but actually that her OSS is actually
[32:09] speaking to 15 you know thousands or
[32:11] tens of thousands of people all at one
[32:13] time how many others
[32:16] 8,316 yeah I mean large language models
[32:19] can talk and can have conversations and
[32:21] then to what extent can uh you know this
[32:25] power actually extend the capability of
[32:28] one or a few people to uh understand
[32:31] what's going on I I heard about that yuk
[32:33] it got it definitely got me thinking
[32:34] because as understood the product is
[32:36] something like it just it will call up
[32:37] all your employees and then your
[32:39] employees can just like ramble about
[32:41] what they've been doing and it will just
[32:43] extract the meaning out of it and give
[32:45] the CEO of like a like bullet point
[32:46] summary of here the most important stuff
[32:48] and there were a bunch of like SAS
[32:50] companies that attempted to do these
[32:51] sort of like weekly pulse Pulses from
[32:54] employees using like traditional SAS
[32:56] software but like that version is is
[32:58] literally a 100 times better than the
[33:00] pre- elm version of this idea but I
[33:02] wonder with like that particular tool um
[33:07] just like it's not it's going Beyond
[33:08] just like reading and summarizing like
[33:11] this this is the argument of like if
[33:12] writing is thinking then like there's
[33:14] actually just a huge amount
[33:16] of work that's involved in the effort of
[33:19] figuring out like who's an effective
[33:20] communicator and like what are the most
[33:22] important things to be like what what
[33:24] are the key things to be focused on as a
[33:25] company and I just wonder if that at
[33:28] some point do the llms do like they go
[33:30] beyond just like summarizing and reading
[33:31] and doing actual thinking at which point
[33:33] like who's actually running the the
[33:37] organization an interesting
[33:39] thought I guess the other thing that's
[33:41] kind of interesting about how Parker
[33:43] Conrad's thinking about it is um I found
[33:45] out about this recently off a an
[33:47] interview with Matt mcginness his coo
[33:49] that uh there are more than a hundred
[33:51] Founders who work at ripling now as sort
[33:54] of specific people who run like an
[33:57] entire SAS vertical inside Rippling it's
[34:00] super cool the way he's built the team
[34:02] har probably knows a lot about it
[34:03] because you've done a bunch of
[34:04] interviews with him um yeah I mean it's
[34:06] definitely very focused on uh recruiting
[34:10] Founders and I mean Parker like Rippling
[34:14] is essentially the the case against
[34:17] vertical like verticalization trying to
[34:20] uh
[34:26] horizontalization like lots of value and
[34:29] he wants to recruit Founders and teams
[34:31] that build on top of the platform like
[34:33] it's almost a little bit more sort of
[34:34] like amazones whereas like shared
[34:36] infrastructure um yeah I think every
[34:39] product that they've released I mean
[34:40] things like time tracking and whatnot I
[34:42] mean basically they launch a thing and
[34:45] it hits like multi-millions of dollars
[34:46] in ARR on day one of launching and
[34:49] that's exactly what we were talking
[34:50] about earlier like once you once you
[34:53] have a vertical once you have a tow hold
[34:55] what you're saying is Well I have to
[34:56] spend this money on sales and marketing
[34:58] anyway can I uh you know basically get
[35:01] higher LTV and hold my CAC constant and
[35:05] uh that's sort of what you if you look
[35:06] at all the top uh software companies
[35:09] today it's like that's what Oracle is
[35:10] that's what Microsoft is that's what
[35:12] Salesforce is ripling knock on wood
[35:14] going to be the next but um it's it's an
[35:17] interesting alternative to uh going from
[35:20] zero to one totally on your own do you
[35:22] guys want to talk about some of the
[35:23] voice companies that we have I think
[35:25] that's like an interesting like sub
[35:27] category of this of this stuff is like
[35:30] really blowing up now I have a company
[35:32] that I work with called Salient that
[35:35] basically
[35:37] does AI voice calling to automate a lot
[35:40] of that collection in the auto Ling
[35:42] space which tradition so they like call
[35:44] up people and they're like hey you owe
[35:46] $1,000 on your car yeah which actually
[35:50] up with that actually this kind of job
[35:52] is one of those butter passing job it
[35:54] kind of sucks because a lot of uh these
[35:57] low wage workers work in all these call
[35:59] centers and it's like a terrible boring
[36:01] job so very high churn and giant
[36:04] headcount to run these because there's
[36:05] just so many accounts with these banks
[36:07] that have to do that and this is a
[36:10] perfect task that AI could automate and
[36:14] what Salient has done is has been able
[36:17] to actually get very very accurate and
[36:19] it has been going live with a lot of big
[36:21] Banks which is super exciting and this
[36:23] was a company from last year and
[36:25] demonstrating that that part of it that
[36:28] they were able to get in because they
[36:29] sold through top down I guess the space
[36:32] feels like it's moving very quickly and
[36:33] that we have incredible companies that
[36:35] are voice infra companies like vapy and
[36:38] then people can sort of get started
[36:40] right away and Retail also I mean these
[36:43] companies that have reached pretty fast
[36:45] scale just because it's one of the more
[36:48] exciting like mindblowing things that
[36:50] you can get up and running within I mean
[36:53] literally the course of hours um and
[36:56] then some of the question that you know
[36:58] remains unanswered and we hope they
[37:00] figure it out is how do you hold on to
[37:02] them especially as you uh run into
[37:04] things like the new open AI voice apis
[37:08] um you know do you go direct like you
[37:11] Pro it's probably way more work to try
[37:13] to use the underlying apis off the bat
[37:16] but these uh platforms are clearly low
[37:19] bar and then the question is can you
[37:21] keep raising the ceiling so that you can
[37:23] hold on to customers forever har you
[37:25] were making an interesting point earlier
[37:27] about like how the apps that people have
[37:29] built on top of LMS has changed from
[37:31] like early 2023 when it started until
[37:34] now voice which we were just talking
[37:35] about as a great example of this I think
[37:36] even if you went 6 months back it felt
[37:38] like the voices were not realistic
[37:41] enough yet the latency was too high like
[37:43] there was it felt like we were probably
[37:45] a ways off having AI voice apps that
[37:48] could meaningfully like replace like
[37:51] humans calling people up and like here
[37:53] we are and yeah I was just zooming out
[37:57] thinking back to the first YC batch
[38:00] where llm powered apps first came in was
[38:03] probably winter
[38:05] 2023 you know almost 2 years ago now and
[38:08] the apps were essentially just things
[38:11] that spat out some text and not even
[38:13] like perfect T Rockit talk that's about
[38:15] it yeah sort of more like copy editing
[38:17] marketing edit email edits it was just a
[38:20] kind of more like just like incremental
[38:22] yeah like I I had a company I mean the
[38:24] one that sticks in my head is a company
[38:26] Speedy brand and all what they did is
[38:28] make it very easy for like a small
[38:30] business to just generate a Blog and
[38:32] spit out content marketing um it's like
[38:35] very obvious idea and it wasn't perfect
[38:37] but it was pretty cool at the time and
[38:40] that's what we've talked about a bunch
[38:41] of the show but that's like the chat gbt
[38:43] raer turned out around that time hey
[38:45] like this is what an llm app looks like
[38:47] it's just a chat GPT rapper it does very
[38:49] basic spits out some text like it's
[38:52] going to get crushed by openi in the
[38:53] next release like and it did yeah well I
[38:56] I I don't know if that one did but but
[38:58] the that that first that first wave of
[39:01] llm apps mostly did get crushed by the
[39:02] next wave of GPT I feel like we've had
[39:06] this sort of boiling of the Frog effect
[39:07] where from our perspect it's sort of
[39:09] like every three months things have just
[39:11] kept getting progressively better and
[39:13] now we're at this point where we're
[39:14] talking about like full-on vertical AI
[39:17] agents that are going to replace entire
[39:19] teams and functions and Enterprises um
[39:22] and just that progression is still
[39:23] mindblowing to me like with two years in
[39:26] which is still relatively early and the
[39:28] rate of progress is just like unlike
[39:30] anything we've seen before and I think
[39:33] what's interesting to see is we
[39:34] discussed this in the last episode is a
[39:37] lot of the foundation models are kind of
[39:39] coming head-to-head there used to be
[39:41] only one player in town with open AI but
[39:43] we've been seeing in the last batch this
[39:46] has been changing Claude is a huge
[39:48] Contender thank God it's like
[39:51] competition is you know the the soil for
[39:53] a very fertile Marketplace ecosystem uh
[39:57] for which consumers will have choice and
[40:00] uh Founders have a shot and that's the
[40:02] world I want to live in so people are
[40:04] watching and thinking about starting a
[40:06] startup or maybe have already started
[40:08] and uh they're hearing all of this how
[40:11] do you know what the right vertical is
[40:13] for you you got to find
[40:16] some boring repeative admin work
[40:19] somewhere and that seems to be like the
[40:21] common threat across all of the stuff is
[40:23] if you can find a boring repetitive
[40:26] admin task um there is likely going to
[40:29] be a billion dollar AI agent startup if
[40:33] you keep digging deep enough into it but
[40:35] it sounds like you should go after
[40:37] something that you directly have some
[40:39] sort of experience or relationship to
[40:42] that is a common like there there's
[40:44] definitely a Common Thread I've seen in
[40:45] the companies that are that I'm seeing
[40:47] promis with and another one just pops
[40:49] into my head sweet spot I think I
[40:50] mentioned on this before like they're
[40:52] basically building an AI agent to bid on
[40:54] government contracts and the way they
[40:56] found that idea and a year ago was they
[40:58] just had a friend whose full-time job
[40:59] was to sit there on like a government
[41:01] website like refreshing the page like
[41:03] looking for new proposals to bid on and
[41:05] they they were pivoting they're like ah
[41:07] like that seems like something an llm
[41:08] could do um a company from a recent
[41:10] batch which pivoted into a new idea
[41:12] that's getting great traction like
[41:13] they're basically building an AI agent
[41:15] to do um process like medical billing
[41:17] for dental clinics and the way they
[41:18] found the idea was um one of the
[41:21] founders mother is a dentist and so he
[41:22] just decided to go to work with her for
[41:24] a day and just sit there seeing what she
[41:25] did and she's like oh like all of that
[41:27] like processing claim seems like really
[41:29] boring like an llm should totally be
[41:31] able to do that and he just started
[41:32] writing software for like his mother's
[41:34] dental clinic so I guess I mean in
[41:36] robotics the classic Maxim is uh you the
[41:40] robots that are going to be profitable
[41:41] and that are going to work are going to
[41:42] be um dirty and dangerous jobs and in
[41:47] this case for vertical SAS look for
[41:50] boring butter passing
[41:53] jobs well with that we're out of time
[41:56] for today we'll catch you on the light
[41:58] cone next time
[42:01] [Music]