00:00 how would you differentiate between an
00:02 idea that could be a great foundation
00:04 for a billion doll company and an idea
00:07 that is likely to get run over by GPT 5
00:10 something that's boring might actually
00:11 be an incredible business but why is
00:14 that yeah let's talk about GPT rappers
00:16 are people worried about giving these
00:18 data sets to open AI all these AI agents
00:21 are passing the touring test I mean this
00:23 is why I think the chat interface is
00:25 wrong you want to do something in AI
00:27 like this is a good place to like look
00:29 into big generational companies are
00:30 getting built as we speak great startup
00:32 ideas just lying on the ground you'd
00:34 like trip over them this might actually
00:35 be like a once- in a lifetime
00:36 opportunity and I I think I actually
00:38 agree what a time to be
00:47 alive welcome to the very first episode
00:50 of the light cone I'm Gary this is Jared
00:52 Harge and Diana and we're group Partners
00:55 at Y combinator and we get to work with
00:56 some of the best Founders in the world
00:59 Jared why are we calling it The Light
01:01 cone well in special relativity the
01:05 light cone is the path that light takes
01:07 from a flash of light you can imagine a
01:09 flash of light and it spreads out in a
01:11 cone shape and in special relativity you
01:14 think about it spreading out in a cone
01:16 both in the future but also in the past
01:20 and in this podcast we are here in the
01:22 present but we are going to talk about
01:24 both the past and future of technology
01:26 so that's how we came up with the name
01:29 and one of the things that we're all
01:31 seeing is the encroachment of AI into
01:34 almost every piece of uh Society at this
01:37 point you know every business
01:39 transaction every uh thing that we sort
01:42 of use with computers uh suddenly a new
01:45 burst of technology is sort of entering
01:48 everything we're doing and we're seeing
01:49 it in the startups that we're funding
01:51 which is why we're so excited about it I
01:53 think you know what what's the
01:55 percentage of companies you've backed
01:57 right now that have large language
01:59 models I think for summer 23 was close
02:01 to 50% of the batch and it's pretty
02:03 interesting like I think a lot of people
02:06 like see that number and they think oh
02:08 YC must have funded so many AI companies
02:10 because we have this thesis about Ai and
02:15 like it's just easier to get into YC if
02:16 you're an AI company because we just
02:18 like love funding AI companies and it's
02:20 funny to us because we know how that's
02:23 not true and yet that's probably what
02:24 like 90 that's probably how 90 plus per
02:27 of people actually think YC Works how
02:29 does Howes how's it actually work can we
02:31 tell people like how it actually works I
02:33 actually think it's interesting the
02:34 smart Founders apply to us with what
02:36 they want to work on and we fund the
02:37 smart Founders like irrespective of what
02:39 they want to work on actually and
02:41 exactly and so the fact that half the
02:43 batch is working on AI says something
02:45 much more interesting than just the YC
02:47 Partners think AI is cool it's an
02:49 emergent phenomenon of what the the
02:52 smart Founders want to work on right now
02:54 is like where do they think there's the
02:55 high beta to build the largest company
02:58 and I think the most ambitious and
03:00 smartest Founders are going after this
03:02 because it's definitely I think the
03:04 exciting thing about right now with AI I
03:06 think it's like real there's been a lot
03:08 of waves for AI and multiple AI Winters
03:11 but this one actually gbt 3.5 and then
03:15 four blew out of the water a lot of task
03:19 and it impressed a lot of smart people
03:21 when a lot of smart people start paying
03:22 attention and building in this current
03:25 idea mace I think big generational
03:27 companies are getting built as we speak
03:29 one thing I'm seeing that's interesting
03:31 is I feel like a lot um a lot more
03:34 Founders are dropping out of college to
03:35 start working on AI because they don't
03:37 there's a f off yeah there's like an
03:39 actual like and usually it's so funny my
03:41 my interview question is almost always
03:42 like what's the rush like why do you
03:44 want to drop out of college like why
03:45 don't you just like graduate because it
03:47 makes a lot more sense to graduate and
03:48 then do a startup um and the reply is
03:50 usually like well like this might
03:51 actually be like a once in A- lifetime
03:53 opportunity and I I think I actually
03:54 agree and and the other cool thing is
03:56 that this is an opportunity where
03:58 college students are particularly well
04:00 like young Founders are particularly
04:02 well positioned to work in it because
04:04 nobody has like like there's no one
04:06 walking around with like four years of
04:08 LM experience so like everyone is
04:11 starting from the same playing field and
04:12 so if you can learn fast you're going to
04:14 be at the same level as everybody else
04:16 that's right and you know one an area
04:18 I've seen that come to play is like
04:21 developer tools for prompt engineering
04:24 I've been seeing like these sorts of
04:25 tools are getting uptick it's like
04:26 ability to like chain together different
04:28 prompts and test your prompts and see
04:30 like the second order effects um and
04:32 actually a lot of college students are
04:34 the people who are just like playing
04:35 around with like prompting models and
04:37 seeing what comes out and it's a really
04:39 easy startup idea for them to like just
04:40 build the tools that they want and like
04:42 the tools that they want are literally
04:44 setting like the standard for what every
04:45 developer should want like I know a lot
04:48 of the headlines are all around like AGI
04:50 and all of the fancy stuff and then the
04:52 really cool demos of like multimodal AI
04:55 like AI generated video and and this
04:57 kind of stuff the stuff that I've seen
04:58 in the batches actually taking off is a
05:00 little bit more mundane like it's um I
05:04 probably say a lot of it sort like
05:05 workflow automation like um it's finding
05:08 things where there was like a human
05:09 doing some repetitive task usually
05:12 involved like searching for things or
05:13 filling out forms and then using like
05:16 llms to replace that it feels very
05:19 obvious to us the people who work at YC
05:21 that this is an amazing opportunity
05:23 there's so many jobs in the world that
05:25 are basically very mundane information
05:27 processing typically stuff that's hidden
05:29 in some back office somewhere where
05:31 there's somebody who's just like reading
05:33 stuff and summarizing it re-entering it
05:35 from one system into a different system
05:37 and like a slightly different format and
05:39 it's such a perfect fit for llms LMS are
05:41 like perfect for this job and yet we
05:45 actually don't get that many
05:46 applications for people working on this
05:48 and there's a lot of Founders out there
05:49 who are searching for a great idea so if
05:51 you're out there and you're looking for
05:52 a great startup idea and you want to do
05:53 something in AI like this is a good
05:55 place to like look into I give you an
05:57 example so last patch had a company I
05:59 worked with called sweet spot and we
06:01 funded them the idea was something about
06:04 like food ordering from food trucks
06:06 something like random and they pivoted
06:09 immediately looking for a new idea and
06:12 the idea they found was um using llms to
06:16 automate searching for government
06:18 contracts to bid on and God such a good
06:21 idea yeah and submitting the proposals
06:22 that sounds so boring what could be more
06:24 boring than searching through like a
06:26 list of all the government contracts you
06:28 know how they found it is um exploring
06:30 startup ideas and then they realized one
06:32 of their friends his job was to work for
06:35 one of these like government contractors
06:37 and his whole day was just spent like
06:39 refreshing this government website um to
06:42 like find things and then submitted
06:43 proposals and they're like what like
06:44 that's like exactly that that's so
06:46 boring like wouldn't you like a tool
06:48 that did this for you yeah and they
06:51 launched and like pretty much straight
06:52 out of the gate got like um a pretty
06:54 decent amount of traction because
06:56 they're like opening up um the people
06:59 who who would actually do it like it
07:00 becomes easier to like find government
07:02 contracts to bid on when it's all
07:04 automated away and like software does it
07:06 for you you know obviously we all know
07:10 that you know something that's boring is
07:12 actually kind of awesome but why is that
07:15 that's like you know just off the bat
07:17 you know we have a sense that something
07:18 that's boring might actually be an
07:20 incredible business there's an old PG
07:23 essay where he talks about this and he
07:24 says um he he quotes a phrase where
07:27 there's muck there's brass it's like
07:30 it's as it's almost like Old English you
07:32 want to explain it har just means like
07:34 you can find treasure in surprising
07:35 places yeah and I think the cool thing
07:37 is you have to go deep and vertical and
07:41 solve a very concrete problem like some
07:43 of the problems with let's maybe talk
07:46 about AI tarpits what a tarpet idea is
07:48 is it's an idea that from the outside
07:52 looks really shiny and attractive it
07:54 looks like a great startup idea and so
07:56 lots of Founders go and they start
07:57 working on it and then you realize once
07:59 you're in it that it's actually not a
08:00 good startup idea but but by the time
08:02 you're there you're like stuck in it and
08:04 so it just attracts founder after
08:06 founder and they just get stuck in the
08:08 tarpet idea and we see this a lot at YC
08:10 because we see all these applications
08:12 and so it's really obvious to us when
08:14 like 500 people apply to a YC bat for
08:16 the same idea but they don't know that
08:18 499 other Founders are also stuck in the
08:20 same tarpet what's tricky I think about
08:22 topet ideas for AI is like we know
08:25 something's that top it idea in
08:27 hindsight once like enough people have
08:28 been stuck in it so with AI it's so new
08:31 we don't know yet so I have a couple
08:33 that I'm actually like Keen to get
08:34 your's thoughts on um a very common one
08:37 is AI co-pilot so it's like hey I'm
08:39 going to make it easy for um people to
08:43 like build an AI co-pilot for their
08:46 product or or service it's it's really
08:49 unusual type of phenomenon where there's
08:52 so much interest from potential
08:54 customers to like want a co-pilot that
08:57 it's actually quite easy to start
08:58 getting getting like inbound leads if
09:00 you pitch this and if it's even easy to
09:02 get people to pay you money up front but
09:04 what's really hard is to get them to
09:07 actually like use the co-pilot because
09:10 they don't actually know what they want
09:11 it for like they just heard that AI
09:13 co-pilots might be changing the feature
09:14 of software so we should have an AI
09:16 co-pilot but they don't actually know
09:18 what their customers will use it for I
09:20 think for me and maybe I just have a uh
09:23 a mental block around chat interfaces
09:26 but I've never been that big a fan of
09:28 chat because it puts so much of the
09:30 emphasis on the user knowing how to
09:33 speak to a computer and you know while
09:36 in the next five or 10 years I think we
09:38 will all get far more used to using it
09:41 that way um I think the the lwh hanging
09:44 fruit right now is just using the large
09:46 language model to actually do the sort
09:49 of knowledge work that a human being
09:51 could do and then package it into the UI
09:55 that you know whether it's a mobile app
09:57 or a web app that is just familiar like
10:00 sort of what people use to do their work
10:02 right now and it's you know basically
10:05 the llm is better used as sort of this
10:08 like I I mean it's almost like you know
10:11 this thing that's sprinkled in that you
10:13 know the software suddenly does
10:15 something really powerful but you don't
10:17 have to change the way you would want to
10:19 use the software as it is sort of like a
10:22 an example of a phenomenon that like I I
10:25 think we have seen in the past when like
10:27 some technology gets really hot and all
10:29 of a sudden like all these companies are
10:30 like they're being asked by people like
10:33 what's our AI strategy they're like oh
10:35 well we better get an AI strategy or
10:37 like with crypto there was like oh
10:38 everybody needed a blockchain strategy
10:40 and even before that it was like
10:41 everybody needed a mobile strategy for a
10:43 moment in time it's like easy to sell
10:46 them something that like placates their
10:47 desire to check some box but in the end
10:50 you've got to actually make it
10:53 successful for them like otherwise it's
10:55 not going to stick I agree and so like
10:58 perhaps with this AI co-pilot thing like
11:00 maybe it's too early to call like
11:01 perhaps they actually will find product
11:02 Market fit maybe with something that's
11:04 not a chap out UI like they'll like keep
11:06 iterating on the UI until they find
11:08 something that's an AI co-pilot people
11:09 actually want or maybe it's just going
11:10 to like fizzle it just like turns out
11:12 most people don't need an AI co-pilot
11:13 some of the advice I've been giving
11:15 those those specific companies is the
11:18 another old PG essay about if you if
11:21 you're trying to sell technology to
11:22 someone and they're not buying like see
11:24 if you can just build a competitor and
11:26 so it's like hey if you're trying to
11:29 uh fintech company a co-pilot and
11:32 they're not buying it well like if you
11:34 are convinced they should have a
11:35 co-pilot like why don't you just like
11:37 build the company with the co-pilot as
11:40 the main experience and see if you can
11:41 out compete them or not I like that that
11:43 I like that I think getting people to
11:45 focus on the use case I think the
11:46 problem is the whole thing with um kind
11:49 of the Gold Rush people selling more the
11:51 shovels and the tools and even then in
11:54 this case it is a bit of that but a lot
11:56 of people aren't digging gold yet like
11:59 the reality is this is such a new
12:00 technology and even the end applications
12:03 that apply AI the reality is there so
12:07 early they don't have product Market
12:09 fits so it's sort of bit of a the blind
12:10 leading the blind in here it's like what
12:13 do I even know what the pattern is for
12:15 copilot I mean it sounds cool just to
12:18 join the cool kid Club of we're doing Ai
12:20 and we're going to check mark So I think
12:22 that's the danger for a lot of these uh
12:24 startup it's like it seems that they're
12:26 getting traction as you mentioned but
12:28 then when you we poke them closer is
12:30 anyone actually using you what are the
12:32 actual use case and then the founders
12:34 come back and they startare a blank at
12:35 us oh but look at all the sign up look
12:38 at the revenue but then they're not
12:40 really using your product I mean we're
12:42 seeing even the second order effects
12:44 right so a bunch of us are funding uh
12:46 Dev tools companies that sell to AI
12:49 companies and they're selling tooling
12:52 but then they might you know they might
12:54 sell an Enterprise contract to someone
12:56 who also Upstream has a Fortune 00 that
12:59 said that they'd pay $100,000 a year for
13:01 that contract and then 6 to n months
13:04 later that you know Fortune 100 went
13:07 back to the incumbent uh you know some
13:09 other leading you know IBM Salesforce
13:13 like something like that um because they
13:16 ended up adding large language model
13:17 technology to what they they were doing
13:19 and people just switched back and
13:21 suddenly the dev Tool Company suddenly
13:23 realizes oh I had five contracts but
13:25 three of them went away because my
13:27 customer actually their customer so it's
13:30 actually like sort of remarkable how
13:32 fast this is evolving you know right now
13:34 in 2024 a specific type of idea I'm
13:37 curious to get thoughts on here as well
13:39 is um offering like fine-tuning open
13:43 source models sort of as a as a service
13:45 broadly like that's a very popular idea
13:49 I think over the course of 2023 here's
13:52 what I've seen so like why do people
13:54 want like why is there any demand for a
13:55 fine-tuned like open source model at all
13:58 um it tends to be initially I think the
14:01 Big Driver was cost like open AI like
14:04 chat GPT was expensive and people wanted
14:06 a um cheaper version of it and so I
14:09 think it was very easy to get customers
14:11 with the pitch of hey like we can f tune
14:13 an open source model and it's just going
14:15 to be much cheaper what I think a bunch
14:18 of the companies in space are seeing is
14:19 that like that's not enough to keep the
14:22 customers especially because like open a
14:24 like the cost of all of the models just
14:26 going down and that's going to keep
14:29 open AI has a plan for all of those so
14:32 there's something more that all these
14:33 fine-tuning companies need to do yeah it
14:35 has be better not just cheaper I think
14:37 where is exactly that where I think is
14:39 having more legs is when these companies
14:42 need to customize it to private data
14:45 sets so you have the open General big
14:48 foundation model but then you have to
14:52 to specific data sets that for example a
14:55 healthcare or fintech can't give out can
14:58 give out and they don't have the team of
15:00 um experts to do it so I think the one
15:03 company that I think Brad worked with
15:05 was credle that kind of was doing that
15:06 what are you seeing about like so the
15:09 concern around data privacy is another
15:11 big reason like are you seeing that as
15:14 being enough like are people worried
15:16 about giving these data sets to open AI
15:19 it's really interesting I mean whenever
15:20 you have something so new like this it's
15:23 actually um sort of resets the clock on
15:26 the competitive landscape again so
15:29 you know you almost can expect all the
15:31 same things will happen again um you
15:33 know just as 10 15 years ago Cloud was
15:36 brand new and then you had Cloud cyber
15:38 security and Cloud strike and all these
15:39 companies sort of come out um you know
15:42 we're seeing the first wave of cyber
15:43 security companies you're like prompt
15:45 armor so they sort of wrap your API
15:48 calls and uh what they actually have
15:50 figured out is that for a lot of large
15:52 language models if you do any sort of
15:54 fine-tuning or training with private
15:56 data you can actually just speak to the
15:59 and get it to spit out your private data
16:01 again and they have a solution that
16:03 stops IT so it's so interesting because
16:06 you know it's entirely possible you know
16:08 they're basically creating a new
16:09 industry again um of cyber security for
16:12 llms sort of in the same way that cloud
16:15 opened up that space and created cyber
16:17 security for the cloud yeah I definitely
16:19 think that whole world of controlling
16:21 within an Enterprise in particular like
16:23 controlling who has access to like which
16:25 llm has access to like what data and who
16:27 has permission is like a really ripe
16:29 space for building interesting software
16:31 I think the other exciting area that a
16:34 lot of the tools are getting built is
16:36 getting more this is like a step further
16:39 fine-tuning but more purpose
16:42 trained models that are smaller so take
16:45 a for instance a llama and getting those
16:48 to run locally in machines for inference
16:51 and when you customize some train on a
16:53 specific domain and Target data is going
16:55 to perform better than the general model
16:58 The General model was kind of trained on
17:00 all of the human language for all of the
17:03 task but if you wanted to build like the
17:05 best let's say um language model for
17:09 parsing SQL queries you would then parse
17:13 very specifically just a set for SQL
17:15 quer and I think some of those that are
17:18 interesting companies that we funded is
17:19 like AMA that you funded that's trying
17:21 to make the development process for
17:23 running all of these locally a lot
17:24 faster and I think we're also funding
17:26 some of these that are custom for coding
17:29 the thing that was surprised learning
17:31 from some of the startups that are
17:32 building um coder type of uh co- Pilots
17:37 which I think is is a use case that's
17:39 working out making a lot of the workflow
17:40 for programming a lot faster it's kind
17:43 of like autocomplete and co-pilot type
17:45 of thing they're training on older
17:48 models of a GPT they don't even need the
17:51 newest one and then I asked like why is
17:52 that and even for like one of the
17:54 companies who funded last batch
17:55 metalware for Hardware they're not using
17:57 the stateof the AR model like the older
18:00 GPT I forget which one was like the
18:02 older 2.5 or three was sufficient and
18:04 actually creating good enough results
18:06 because the vocabulary for a specific
18:10 domain for Hardware or software is a lot
18:12 smaller than the human language so this
18:14 is other world where the open model
18:17 that's customized I think is going to
18:20 win and compete versus the big one for
18:22 specific domains so there lots of
18:24 companies with this yeah that's what uh
18:26 Toby loty from uh shop actually still
18:29 dabbles with the stuff I think he
18:30 actually built the uh internal co-pilot
18:33 for Shopify and what he was saying is
18:36 the best way to use whatever gp4 or the
18:39 you know latest Clos Source models that
18:41 are most expensive and have the most
18:43 parameters uh just think of it as a
18:45 prototyping tool anything you do with
18:48 those prompts you can get your own model
18:50 to do with a little bit more training
18:53 it's kind of like uh when people build
18:54 Hardware you have the analogy of uh
18:56 prototyping with fpga
18:58 which are very expensive right and then
19:01 when you have the right architecture for
19:02 Hardware then you do the circuit path
19:05 and actually do the custom s so so right
19:08 now for some of these tasks the large
19:11 language model is sort of like your
19:14 fpga whatever GPT 4 and then when you
19:16 customize it you do like the super
19:18 efficient one coding path for I don't
19:20 know Shopify for coding assistance and
19:22 Hardware software Etc that becomes your
19:25 so that you train and customize which is
19:27 cool I think that patterns emerging it's
19:29 like as I hear you talk about that
19:30 what's I just think it's just like so
19:33 many different startups that could be
19:35 built it just feels like we've never had
19:38 this moment at least I didn't feel like
19:39 I've never experienced a moment where
19:41 there's just so many potential startup
19:42 ideas to be built like all that ones
19:45 yeah there there absolutely hasn't in we
19:46 we definitely saw this in the last batch
19:48 with all the pivoting companies oh yes
19:51 people don't always realize this but
19:53 like many of the companies get into YC
19:54 within a month after we fund them
19:56 they're looking for a new idea cuz the
19:57 old thing didn't didn't work or they
19:58 lost interest in it or something and
20:00 it's normally like not actually that
20:02 easy to find a great startup idea for a
20:03 team to work on but man was it easy last
20:05 summer God it was just just like great
20:07 startup ideas just lying on the ground
20:09 you'd like trip over them yeah that was
20:11 a fast I think you actually had a tweet
20:12 about it that was one pretty uh viral
20:15 that talked about this is the batch the
20:18 batch ever in your whole career working
20:20 at YC where Founders got to good ideas
20:22 the fastest ever and hard has been here
20:25 even even longer yeah know it definitely
20:27 feels unique I've never had so many
20:28 successful pivots yeah and Gary to your
20:31 point about the chat gbt rapper I think
20:34 back like I feel like that Meme really
20:36 came out like just about a year ago yeah
20:38 let's talk about GPT rappers yeah like
20:40 like I feel like the first sort of group
20:42 of ideas I saw in the batch were all
20:44 generative AI ideas built on Chop top of
20:47 chat gbt so was stuff like hey like
20:49 automate your marketing copy or automate
20:51 like your creative content or something
20:54 like that and that term got thrown out
20:56 oh these things are all just like
20:57 rappers on top of chat GPT and um open
21:00 AI is going to like take all of like
21:02 it's just going to build all of these
21:03 things and they were going to release
21:04 their App Store and like it's just going
21:06 to take all the value and these things
21:07 will die of the mem all of all of SAS
21:09 software is just my sequel rappers
21:12 exactly I think this is a great analogy
21:14 you can think about any SAS product as
21:17 basically a database rapper like you
21:19 could imagine like negging any SAS
21:21 product CU like the first version of a
21:23 sass prod it's basically just a crud app
21:25 and just like you took like my SQL then
21:28 you like built like a website on top of
21:30 it and I think people are going to look
21:32 back on this term GPT GPT rapper like
21:35 similarly how we think of like how we
21:37 would look at the term database rapper
21:39 which just seems like silly I mean this
21:41 is why I think the chat interface is
21:42 wrong like I actually think there is
21:44 value acur to really great ux like good
21:47 copy good um you know interaction design
21:51 information hierarchy uh you know being
21:54 able to approach a product and say like
21:56 this is the job to be done and for for
21:58 users to come in just sort of naturally
22:00 understand what to do like there is a
22:02 craft to building software that is
22:05 timeless and that sort of transcends
22:07 whether or not you're using a large
22:08 language model and so you know that that
22:12 I think is what I mean by you know these
22:15 things are not you know SAS software is
22:17 not uh a MySQL rapper well here'd be a
22:20 question I'd be interested in in in
22:22 everyone's thoughts on suppose you're a
22:25 new founder and you really want to build
22:28 company and you want to do something on
22:31 LMS how would you differentiate between
22:34 an idea that could be a great foundation
22:37 for a billion dollar company and an idea
22:39 that is likely to get run over by gbt 5
22:42 and is probably like not a good starting
22:44 point I think if a Founder is working on
22:47 something too General and not solving a
22:51 specific need for a user they can
22:52 actually go talk to another use case so
22:55 I I worry about the ones that are too
22:57 generic generic and building going
23:03 abstract it will solve all the things
23:06 yeah if it's like hey like throw your
23:08 data in here and we'll do like
23:10 automations on top of it like for
23:12 everything that's probably hard to
23:15 compete with whatever one of the
23:17 foundation models might offer but if
23:19 it's like hey we are give us like your
23:23 sales log data and will like um spit
23:28 back like suggested next actions like
23:30 you can like for sales people to make
23:32 them better at sales that's probably
23:33 going to work better or give us all your
23:35 compliance checklist to pass Hippa
23:37 compliance and process that it's like
23:39 that's very specific and lots of
23:41 business logic or give us all of your
23:45 for processing government forms right
23:48 yeah so a lot of custom business logic
23:50 so the same thing with the SAS era a lot
23:53 of the applications and how you build
23:55 applications in there there's always the
23:57 separation business logic and they crow
23:59 in a lot of architectures for these app
24:02 and a lot of the value of the company is
24:04 accured on that business logic that is
24:06 so custom per company and there's a
24:09 whole pattern of uh programming patterns
24:10 on how people separate those yeah gu as
24:13 this all goes multimodal this is going
24:15 to get really interesting so early days
24:17 but yeah we've seen companies work on
24:19 voice AI apps to be like a sales rep and
24:22 I think um it's an interesting example
24:25 of the kinds of ideas that might be
24:26 possible now with AI is where you take
24:28 something like a Salesforce and you try
24:31 and reimagine like what would Salesforce
24:33 do if it were started today with all the
24:36 power of AI what it almost certainly do
24:38 more than just be like a CRM right like
24:41 it would make like it would find who
24:43 your leads might be like maybe now it
24:45 can make the calls for you it could like
24:48 set them up like maybe it goes all the
24:50 way to start like implementing like the
24:53 first version of the product for them
24:55 like I think it's just like the scope of
24:57 software you can build with AI now is so
24:59 big I think that's another good way to
25:01 find ideas like look at software today
25:03 and reimagine it with the power of AI
25:06 today which you funded a number of
25:07 companies that effectively are AI voice
25:10 agents for small businesses because they
25:12 receive I don't know if you're like a
25:14 flower shop or a AC repair man in the
25:18 middle of U the US there's a lot of
25:19 calls for you to schedule and you don't
25:22 have a lot of stuff automated and
25:24 there's these YC companies that are
25:25 using that building these AI voice
25:28 agents to basically be the
25:30 receptionist I know one of our partners
25:32 Paul buight is quite worried about this
25:34 actually he's worried about there's
25:35 going to be a world of just s like all
25:37 these AI agents that are out trying to
25:39 do malicious things and that we're going
25:42 to need like our own like good defensive
25:45 AI agents out there making sure we don't
25:47 get scammed out of all of our money I
25:50 mean this is actually why I'm so uh an
25:52 advocate for open source AI because
25:54 these things are sort of real
25:56 considerations um you know can you
25:58 imagine there only being one hyperd
26:01 dominant AGI and it's totally close
26:03 Source it's owned by one company and uh
26:07 you know it's only available to the
26:09 highest bidder and uh you know imagine
26:12 you being uh you know someone who just
26:14 had to go to the doctor and uh on the
26:17 other end of it is uh some health
26:19 insurance company that uh you know
26:21 bought the bought access and blocked it
26:23 out from everyone else and you know you
26:25 getting on the phone you're not able to
26:27 sort of navigate or go against the sort
26:29 of you know impenetrable AGI that is
26:32 able to sort of get around anything that
26:34 you know your side might throw at it
26:36 like we actually want you know some form
26:38 of actually Equity at the AI level like
26:40 we actually want uh you know not merely
26:43 the biggest companies to own the most
26:46 capable AIS we want all consumers to be
26:48 able to have from the bottom up uh the
26:50 same access to that same technology and
26:53 that's uh you know the best insurance
26:58 certain that's actually what a lot of uh
27:00 also not just Founders but smartest
27:03 researchers who are really at The
27:04 Cutting Edge is I went to near IPS this
27:07 past December which was incredible to
27:10 see the energy in there the conference
27:12 has grown so much I think it like over
27:14 10,000 attendees there were 3,000 papers
27:17 more than 3,000 papers accepted and I
27:20 think um back in 2017 there was only
27:22 around 600 papers when I went back in
27:26 2010 it was was just in a ski lodge and
27:29 maybe like a 100 papers it's crazy the
27:32 kind of exponential growth and one of
27:35 the big topics of Interest was a lot
27:36 around AI ethics and Regulation and how
27:39 do we measure that so that that was
27:42 interesting um but the thing that's
27:44 different about typically that was
27:45 interesting in this conference is the
27:47 amount of interest from researchers
27:49 wanting to start companies too one
27:51 interesting data point is um a lot of
27:54 this era with GPT came about from from
27:57 One Foundation paper is all attention
28:00 you can need it was this paper that got
28:03 released got launched in a New York IPS
28:06 back in 2017 it was a team at Google who
28:10 was trying to figure out how to make a
28:11 machine translation between
28:14 languages more cheap because the English
28:17 translation to any language was actually
28:18 pretty good but if you wanted to do I
28:21 don't know German to Japanese there was
28:22 not enough data so they figur out this
28:25 way to compress data which became the
28:27 Transformer models for GPT and it was
28:29 like groundbreaking and this is the
28:31 foundation for llms that paper came out
28:35 2017 and the fun fact I was just looking
28:37 this up out of all those author eight
28:40 authors seven of them start at different
28:43 companies and all of the companies in
28:46 total their rate their worth valuation
28:50 billion and now people are seeing oh
28:53 these like industry Pioneers did this
28:56 and it's creating this new crop of I
28:58 think Founders that I don't think would
28:59 have started because I talked to a lot
29:01 of AI researchers and I don't think they
29:03 wanted to be Founders and I got a l this
29:05 question how can I turn my paper into a
29:06 company which I think is cool because
29:09 this is like going back to the root of
29:10 um why I F funding hardcore technical
29:14 Founders and I think it's cool to see
29:16 that energy there so when we went and
29:18 host our event we uh I didn't plan and
29:22 it was like 3x over subscribed nice
29:24 standing room only huh yeah yeah it's
29:27 that sounds like really the new Homebrew
29:29 Computer Club so NPS in December yeah we
29:32 got to mark it on the calendar we'll
29:34 come back yep Diana I love your point
29:36 about how this is sort of like returning
29:38 YC to its roots it definitely felt that
29:40 way last summer because when YC got
29:45 started the internet was really new and
29:48 the people who were building stuff on
29:49 the internet were mostly technologist
29:51 because actually like pretty hard to
29:52 build websites back then and pretty hard
29:53 to build like good software and like as
29:57 building software and building websites
29:58 got commoditized a lot more people came
30:01 space and this is a cool reversion back
30:04 to the like Origins where like the
30:05 people who are building the most
30:07 interesting stuff are like mostly really
30:08 hardcore like researchers and
30:11 technologists because there's actually
30:13 real new technology being invented it's
30:15 not just like innovating on business
30:17 models with like commoditized technology
30:20 and again just like every great
30:21 technology it's being dismissed right so
30:24 going back to like the chat gbt rapper
30:26 meme I actually think that was great for
30:28 YC because it meant we only got the
30:31 people who are like tune who could tune
30:34 that out and we just like hey like
30:35 either I'm just so interested in this
30:36 technology I don't care like what the
30:38 memes are or I'm just too busy building
30:39 it to pay attention to the meme on
30:41 Twitter which is also great but like I
30:43 feel like this has always been the case
30:45 right like Homebrew Computer Club like
30:47 PCS are like dismissed as like toys like
30:50 the internet is dismissed as a toy like
30:52 all all of these things so feels like
30:54 that moment again yeah there is a a
30:58 essay that I love that I saw off Hacker
31:00 News do you guys remember this it's
31:03 sociopaths in a subculture Evolution and
31:07 you know I think that that actually is
31:09 the one thing that's quite durable and
31:10 like keeps returning right it's always
31:12 the Geeks Who are going to be into the
31:14 tech no matter what they're on The
31:16 Cutting Edge you know uh I always think
31:18 of Steve wnc talking about like you know
31:22 we started Apple computer with no idea
31:24 that it would ever be a company like we
31:26 just wanted computers for ourselves and
31:28 our friends and so you know at some
31:30 point the you know sociopaths come along
31:34 and they start sort of uh monetizing the
31:37 people who you know come to the scene
31:39 and then the cycle returns and repeats
31:42 so that's why I like being at the
31:44 beginning of a new cycle and clearly AI
31:47 is exactly that so don't don't count it
31:50 out don't write it off it's one of the
31:52 most interesting things that are is
31:54 happening out there um but you know
31:56 there are clearly things to be careful
31:58 of like don't be uh attracted to the new
32:01 shiny thing uh instead look for the muck
32:04 because where there's muuk there's brass
32:06 so that might be a great place to call
32:08 it for the very first episode of the
32:11 light cone we'll see you next