00:00and we're seeing people making software
00:01in 30 minute increments someone the
00:04other day put a request for a landing
00:05page up they attached a figma file they
00:08got it back in 30 minutes that's really
00:10never been done before and why shouldn't
00:13have it existed uh like that before it's
00:16hard to find concrete numbers but some
00:18estimates predict that less than one
00:19percent of the world knows how to code
00:21yet it's also widely known that the
00:23minority of people who know how to not
00:25only consume but create software yield
00:27outsized returns with some developers
00:29generating some of the highest paying
00:31salaries out there but even this field
00:33is Shifting quickly especially with the
00:35Advent of wide scale Ai and in this
00:37interview we get to chat with amdad
00:39Mossad founder of replit and integrated
00:41development environment that allows you
00:42to code live in the browser here we chat
00:45about how replit has tackled the
00:46difficult problem of making coding fun
00:48but also how it's now integrating AI
00:50into its platform via Ghostwriter and
00:52the implications of these shifts on both
00:54current and future developers in
00:56addition to the applications that can be
00:57built as a personal anecdote I actually
01:00taught myself to code in 2018 and to
01:02this day it's one of the best decisions
01:04I've ever made and despite screaming
01:06this from the rooftops still many other
01:08people find it extremely daunting my
01:10journey took around 300 hours yes I
01:12tracked it but with the advancements in
01:15tooling and Technology it may actually
01:17be easier than ever to learn to code and
01:19I was surprised to hear from amjad just
01:21how quickly he thinks people can get up
01:22to speed today let's get started as a
01:26reminder the content here is for
01:27informational purposes only should not
01:29be taken as legal business tax or
01:31investment advice or be used to evaluate
01:33any investment or security and is not
01:36directed at any investors or potential
01:37investors in any accz fund for more
01:40details please see a16z.com disclosures
01:53thank you so much for joining the a16z
01:55podcast it's my pleasure I've been
01:57listening to this podcast for a really
01:58long time so it's great to be finally on
02:00yeah well we're happy to have you so
02:02just to set the tone for listeners you
02:06are the founder and CEO of replit I
02:08think you're also the head of
02:09engineering now but we'll get to that uh
02:12can you just share what replit is and
02:14I'm probably getting a little ahead of
02:16myself but how it differs from maybe
02:18other similar products or development
02:21environments whether it's scorm or just
02:23other Ides out there I like to think
02:29the technology that reduces the distance
02:32between an idea and a product the moment
02:36any person in the world gets an idea for
02:39a piece of software and so the distance
02:42between you know having that idea and
02:45making something in the world is
02:47typically very large
02:48the internet has really reduced things
02:51down further and you know we think
02:54replied is the sort of the ultimate uh
02:57answer there where we want to get to a
03:00place where the moment you get an idea
03:02you just reflect you know and that's
03:04sort of where the name came from and we
03:07do that in a number of ways one way is
03:09that we just simplify the development
03:11process crucially we don't make it
03:14stunted like a lot of tools sort of
03:17think that to simplify is to reduce the
03:21power and we actually keep the power
03:23while also making the process a lot more
03:27enjoyable unlearnable we also super
03:29power it with AI we also you know have
03:33various courses and things that can
03:35teach you how to code if you don't think
03:37coding is for you then you can hire
03:40someone from our community to make the
03:41program for you so we have this thing
03:43called reflect bounties where you can
03:46just like put a price on a project you
03:49want to get done describe the project
03:51and then a combination of an AI and a
03:54human being will get that done for you
03:55and so it's really ultimately about that
03:58idea of like getting from idea of
04:00product and will help everyone in the
04:03world kind of have access to that
04:06superpower that we call software I love
04:09the way you put that because you've kind
04:11of represented the different modalities
04:13depending on where someone is and their
04:15proficiency or at least their
04:16hypothetical Proficiency in terms of
04:18being able to code and regardless of
04:20where they sit getting them to that end
04:22product that Envision that they they
04:25want and we'll get to Ghostwriter we'll
04:27get to bounties those are two things
04:28that have been released recently but I
04:30want to give people a real sense of some
04:33of the stories behind replit because
04:35that's something that's really gripped
04:36me as someone who's been following the
04:37company for a period of time and I'll
04:40call out two but then I want to play a
04:42quick game so that we can share more of
04:44these stories with the audience so two
04:46things that I remember one is recent we
04:49shared a 13 year old kid that set up a
04:52sizable crypto mining operation via
04:55replit another example that I saw a
04:58while back was these two kids they
05:01looked maybe six years old maybe seven
05:03years old crashing a computer science
05:05teacher conference and sharing their
05:08startup pitch to The Listener so I just
05:11thought those were incredible examples
05:12but getting to this idea of a game I've
05:14come up with five or so different
05:16scenarios I have no idea if these
05:19scenarios have actually happened on
05:20replied I haven't shared them with you
05:22prior to this but I think they might
05:24stimulate either other stories that have
05:27happened on replit or you can tell me if
05:29these things have indeed happened so
05:31let's give it a shot the first one is
05:35someone has started learning to code via
05:38replit they haven't you know gotten a
05:40computer science degree they haven't
05:41done a boot camp they've learned within
05:42replit and since been hired as a
05:45developer at a Fang company has this
05:48happened yet I would say it happened yes
05:52um we've had someone who learned a code
05:55on droplet primarily on their own became
05:57a very productive Community member and
06:00applied to work at Ed ruplet
06:03and we thought they're very talented by
06:05the time there wasn't like a really good
06:07fit for them and so they started
06:08applying for other companies I helped
06:10them apply for other companies and they
06:13ended up uh at Google actually that's
06:16awesome let's move on to the next one
06:17someone has used replit to build over 50
06:20projects now you can use your own
06:21definition of project here these don't
06:24need to be full-blown startups but have
06:26you seen someone build that magnitude of
06:28projects on replit yeah I think that's
06:32um that's fairly reasonable number of
06:36projects so there's a 16 year old
06:38developer his name is reihan he's one of
06:41our most prolific programmers he
06:45built a bunch of interesting projects
06:48where he reverse engineered how replic
06:52works and he built he built an
06:56unofficial API one of his unofficial
06:58apis is a security program that searches
07:02people's Rebels to find Discord tokens
07:07so a lot of people build Discord
07:08bathroom and they copy and paste the
07:11tokens and clear text as opposed to
07:13putting them in our encrypted service
07:14the secrets manager and so he would find
07:18those tokens he would invalidate them
07:19because Discord has service to
07:21invalidate those and then he would send
07:23them a notification he would say like
07:25hey like we found that you've exposed
07:27your token so that's one example of a
07:30project he made he he was and Cilla is
07:33one of the most prolific bounty hunters
07:36and he built one of our earliest
07:41a startup that wanted to build like a
07:44stable diffusion based t-shirt generator
07:46so he would generate a t-shirt based on
07:49prompt and then get it printed and sent
07:51to you like there isn't one week where I
07:54don't see Rayhan producing like a new
07:56piece of software that's amazing so I
07:59would say like 50 is probably on the low
08:02side of uh things here yeah I mean I I
08:05undershot it I guess well let's rapid
08:07fire through the last three the third
08:10one is someone has built an app on
08:11replit and through that app has since
08:13made a million dollars or more from it a
08:16million dollars I don't think that
08:18happened uh yes there are people that
08:21are on trajectory to doing that right
08:24now and so there are like Bounty Hunters
08:27that are making thousands and thousands
08:29of dollars a week and I think that'll
08:31continue to grow there are startups that
08:33are starting entirely on replica today
08:35there are some a couple of AI startups
08:38that have their entire stock built on
08:42um and I I presumably one of those
08:46projects will maybe get to a million
08:49dollars there are a couple of earlier
08:51examples of startups that sort of
08:53prototype their projects for example fig
08:56is this command line autocomplete tool
09:00that this YC startup I think it was like
09:03YC 20 or something like that or was he
09:0521 and it's used by tens of thousands of
09:08Engineers and the entire the original
09:10product was entirely built on on raplet
09:12uh they're starting to sell I don't
09:15think they're making a million dollars
09:16but maybe that's another example of
09:19maybe on the path to a million dollars
09:21but I would say it would be a success if
09:24someone like an individual person just
09:26made a million dollars I think that's
09:28really my dream and I hope we can get to
09:30that this or next year that's awesome
09:32and I should mention there's a time
09:33element to this I believe replit was it
09:35founded in 2016. that's right so we're
09:39around six to seven years later so let's
09:41see how things progress next one is
09:43someone over the age of 90. I went a
09:48little crazy with this one someone over
09:50the age of 90 has used replit has this
09:52happened someone who uh ended up uh
09:55working at a rufflet their grandpa in
09:58India is like a physics uh PhD they do a
10:02lot of like physics uh sort of writing
10:05and in one recent paper they published
10:09they used uh replica to write physics
10:12simulation and droplet was published as
10:15part of that paper I don't believe
10:16they're 90 but they're they're
10:18definitely 70 plus okay final one
10:20strangers have met via replet maybe it's
10:23the Forum maybe working on a project
10:25together and have since gotten married
10:29man that that would be awesome to know I
10:32haven't heard of it if that happened I
10:34know there are some community members
10:36that like recently got married I don't
10:38think they met on on the community but
10:42you know I've certainly like seen some
10:44people date and things like that but uh
10:46married I haven't heard that yet so well
10:49I had to ask it like I said I hadn't
10:50shared these with you beforehand so I
10:52went from what I think are maybe more
10:54ordinary scenarios and maybe more
10:56extraordinaries one story though let me
10:58mention is that you know there's a story
11:00in the news about the world's youngest
11:03Microsoft Azure AI certified programmer
11:07and naturally someone looked them up and
11:09they found them that they're unreplet
11:11this kid is six years old oh my goodness
11:15and if you go to the replit profile they
11:18have a thousand followers on droplet
11:21and they just like they have games with
11:25thousands and thousands of runs and
11:28they're just like completely prolific
11:29and I just can't trap my head especially
11:32now that I have kids
11:33I can't drop my head around like a
11:35six-year-old like doing this like every
11:37day and it's like pretty freaking
11:39amazing that's incredible because just
11:45in grade one I believe I was learning to
11:48like trace letters so that I could write
11:50the alphabet wow that's incredible
11:54um are there any other stories worth
11:55mentioning here any top of Mind
11:57scenarios like a six-year-old who's
12:00coding on replit one of my favorite
12:04stories is one day I wake up and I see
12:07my Twitter blowing up and I go on
12:09Twitter I find that this Indian mom and
12:12dad uh are tagging me and saying that I
12:15corrupted their child their kid is
12:16totally addicted to raplet they're
12:18supposed to go to IIT as you know IIT
12:21requires a lot of preparation but that
12:24kid is not interested in it so not
12:26studying at all and just programming all
12:29the time and I actually felt pretty
12:32um because it started going viral and a
12:35lot of people were like pretty negative
12:39um apparently it hit a nerve and some
12:40Discord uh discourse in India about how
12:43parents push their children really hard
12:45on an IIT and so I reached out and try
12:49to help I try to talk to the kid and
12:51tell them like hey you should listen to
12:52your parents and then um a few months
12:57um India was hit pretty hard with kovid
12:59if you remember the the issue there
13:02and that kid wrote the application that
13:06First Responders are using to find
13:09equipment to find oxygen to find all
13:12these things and it went totally viral
13:15it brought it brought down our website
13:18that's how viral it went and in my mind
13:21I was like that his parents should be
13:23really proud now and then something even
13:25more fantastic happened out of that
13:27experience and him going viral using his
13:30skills he got a job and now he gets paid
13:34more than his entire family and and so
13:37you go is sort of this hero's journey
13:39you go from getting blamed for someone
13:41ruining their future it actually turns
13:43out that there actually did a huge favor
13:45to their future and I think that's the
13:48sort of power of of software and the
13:50internet and things like repl that's
13:53beautiful and something that also stands
13:55out to me from that story which we'll
13:57touch on throughout this interview is
13:59this idea of coding being fun like this
14:02kid really just wanted to code all the
14:04time and as someone who taught myself
14:06several years ago prior to learning I
14:09thought coding was really dull just
14:13something more prescriptive instead of
14:14creative and artistic and again using
14:17this word fun and after I learned to
14:19code I saw it in a different light and I
14:22think that's really important as this
14:23idea where it's like how do you make
14:25coding fun so what are your thoughts
14:27there how have you been able to design
14:28replet to actually enhance people's
14:32creativity and make them want to come
14:33back and make them see this skill in a
14:37new light so I think most things that
14:39are fun tend to be devoid of a lot of
14:46drudgerous routine work right you know
14:49when you're playing video game you're
14:51not like building the video game every
14:53time or setting up the TV or or doing
14:56some like wrote it task right when you
15:00are doing a sports hobby you're in the
15:04flow you're doing the thing you're
15:06excited about doing the problem with
15:08coding is that a lot of the maintenance
15:10around the development environment and
15:12the packages and the integration of all
15:15the different components was the thing
15:18that Engineers were spending most of
15:20their time doing the moments they were
15:22coding they were an absolute Bliss but
15:24those moments were actually
15:26very little in terms of the if you think
15:29about the pie chart of what it meant to
15:31work as a programmer so the first thing
15:34that replica does is remove the need to
15:38do all this setup that in itself made
15:40programming a lot more uh fun
15:43and then add the collaborative aspect
15:46like a lot of what we find fun in life
15:49has to do with other people we're just
15:51social animals right and so like if I
15:54can share my program with you with just
15:57a link that's really fun that's what
15:59makes figma fun that's what makes any
16:01other collaborative tool fun is that I
16:03can just like send you a link and you're
16:05in there with me or you can play it and
16:07try it out and then finally there's like
16:09a lot of like explicit gamified elements
16:13of ruplet for example like we have a
16:15built-in currency called Cycles Cycles
16:18allow you to like earn you know from
16:20bounties you can spend it on AI or
16:22compute or you can cash it out and it's
16:25built with a with a mindset of like it
16:28feels like a feels like Roblox it feels
16:32like one of those in-game currencies
16:34well I like your analogy of almost like
16:37the setup or the admin that goes into
16:40coding at times being a lot more
16:43overhead where if you compare it to
16:45other activities out there like I love
16:48playing soccer for example if playing
16:50soccer required me to like put on my
16:52shoes and shin pads for an hour before
16:54every hour long game I'd be like man
16:56this is really frustrating like I just
16:57want to get on the field and so that
17:00ratio is actually important let's take a
17:03left turn and go straight to AI which is
17:05you know the theme I feel like of the
17:08last six months with many companies but
17:10why don't you give listeners a quick
17:13overview of what Ghostwriter is and how
17:18um coastwriter is like a pair programmer
17:20that is an AI we think about it as this
17:24embodied thing that helps you achieve
17:27your goals of making software
17:29um and so it helps you in every aspect
17:32of software creation whether that's
17:34typing code it provides this type ahead
17:38components similar to Gmail so either
17:40complete lines of code or entire
17:42functions or higher classes so you can
17:45see that happen as you're typing you
17:47also can generate sort of entire
17:50programs so you can do right click
17:51generate and you can give it a prompt in
17:54the same way you could stable diffusion
17:55Dolly a prompts and generate entire
17:57images you can give it a prompt and
17:59generate entire programs entire programs
18:02that run entire like applications
18:04and then you can highlight a piece of
18:06code and you can right click and you can
18:08transform it finally like the ability to
18:13talk to Ghost Rider is the final piece
18:16here and we're actively working on it
18:18and it's uh it's our number one priority
18:21it'll be like a chat box inside wraplet
18:24and it'll both chime in when things
18:27happen that it could help so for example
18:30if you get an error it would like I can
18:33help you with that like a clippy like
18:34thing hopefully not as annoying
18:36but clippy was way ahead of its time so
18:39I'll be like you know you got an error
18:41like if you want I can you know try to
18:43help you with that or you can just sort
18:44of ask it like hey like I wrote this
18:47code can you write write tests for it
18:49but it's really remaking the IDE as we
18:53know it I would also say it's less
18:56binary in a way something I've been
18:58thinking a lot about as I've learned
18:59more about how AI is being implemented
19:01with code is when I learned one of the
19:04most frustrating aspects of learning was
19:06just the binary nature of code when you
19:08get into a bad error it's just like no
19:10no no no no no no and then eventually
19:12you hit a yes and sometimes that takes
19:15way longer than is comfortable and so
19:18what I like about this idea of having
19:20this Sidekick is just the Nuance to that
19:24of someone being or someone the AI being
19:26able to say oh have you tried this yet
19:28oh did you mean this by your code like
19:31this is what your code is trying to do
19:32is that what you intended getting
19:35suggestions and again having the a
19:37learning curve be less like a no no no
19:38no yes and more like a gray that evolves
19:42into a yes like a gradient descent sort
19:45of how you how you train an AI actually
19:47so yeah I I 100 agree with that I think
19:51it humanizes the process as opposed to
19:53this like mechanistics of trial and
19:55error thing it allows us to iterate a
19:57lot faster on tools like traditional
20:00Ides are very hard to build they're
20:03super complex classic algorithms I'm
20:06sure you remember stuff but like some of
20:08these Ides are super big like you
20:11download uh IntelliJ that's like a
20:13couple gigabytes of stuff and they're
20:16clunky they take up a ton of ram like
20:20try starting X code you know it's it's
20:22kind of like uh it just consumes your
20:25computer and the cool thing about the AI
20:27Revolution is that the eye is going to
20:30be running in the cloud you're going to
20:31give it a prompt like you know what
20:33you're talking about what your code is
20:34about and it'll be able to implement
20:36suggestions and Implement features and
20:40tools without having this heavy
20:44algorithmic hard to maintain sort of
20:48um so I'm really excited about that and
20:50as a startup you always want to catch
20:53a new platform shift and with replet I
20:57feel like we're catching this platform
20:59shift and the older IDs will not be able
21:01to adapt as fast as we can well
21:03something that I've heard you talk about
21:05is also a decision that you made which
21:07was to build Ghostwriter on top of your
21:10own models so something like a co-pilot
21:13is built on top of gpt3 to my knowledge
21:16and that's a decision to be built off
21:18another platform but you went a
21:21different route so can you speak a
21:22little bit more to how you made that
21:23decision and what kind of inputs led to
21:26that output well first of all how crazy
21:29it is that uh that Microsoft had to
21:31depend on another company whereas
21:33rappler built our own thing there are
21:36like a multiple ways to chance with this
21:38one of them is ux ux is inherently
21:42Inseparable from the infrastructure for
21:46how a product works I think most people
21:49think as their separate things but if
21:52you're serious about making products you
21:54know uh you know famous
21:56um Allen K quote to Steve Jobs he told
21:59Steve Jobs if you're serious about
22:01making software you have to make
22:03um and that's why like Apple's this full
22:05stack company is because they think
22:08about everything from the transistor to
22:10the touch right and so I think for us it
22:13like if this is gonna be a core
22:16interaction with our platform we have to
22:19be able to optimize it and we have to
22:22get the latency down to the point that
22:24we we like we feel it's going to be a
22:26really great user experience and we
22:28weren't able to really get that when
22:30we're hitting something of our n API
22:32because the latency will be all over the
22:34place we can get the caching ride we
22:38um you know location right we didn't
22:40have control about any of these things
22:42that's a huge downside of being a
22:44consumer of a mirror API and then uh the
22:47other part is a strategic part which is
22:49like if you think if you believe that
22:52this is a primary platform shift and
22:55this is going to be a core part of your
22:58technology then you have to build it if
23:01you call yourself a technology company
23:03that means you build technology right it
23:06just doesn't mean you're just like
23:07building glue code on top of like
23:09existing technology finally we think
23:12that we have a bit of a data advantage
23:15and that data Advantage will compound
23:17over time and so it will allow us to to
23:20train more advanced AIS over time
23:24so all these three reasons just made
23:28sense for us to bring at least part of
23:30it in-house I should say that we still
23:32use openai for a lot of the sort of like
23:35um you know bigger uh workloads that
23:39require really large models something
23:41that I found really interesting was
23:43Daniel gross and Matt Friedman were on
23:45the strategory podcast and they talked
23:48about how they ultimately their
23:50investors are not the creators of
23:51copilot but how co-pilot ultimately got
23:54to the interface that it now is and
23:57originally they actually wanted to
23:58create it as a chat bot they thought oh
24:00people will run into an error they're
24:02going to want to talk to someone and ask
24:04hey how do I fix this and they're going
24:06to get a response and implement it but
24:07ultimately they ended up kind of
24:10pivoting to what you might imagine as
24:12like a robot on your shoulder that only
24:14speaks up when it has confidence and so
24:16I know it's the early days of Ghost
24:18Writer but thoughts on how you got to
24:21the specific interface that is
24:22Ghostwriter today and how you as you
24:25said kind of linked the ux that people
24:27see to what's happening in the back end
24:29as you're building these models
24:31so so I think I think there's
24:33um there's two modalities one is pull
24:36and one is push right so pull is the
24:39human knows what they want and they're
24:41gonna ask for it you write a prompt
24:43you're gonna wait a little bit and
24:44you've got to get it and then there's
24:46push which is the robot on your shoulder
24:49that is like continuously suggesting
24:52improvements and their trade-offs to
24:54both the push model is actually fairly
24:57expensive because you're Computing all
25:00the time it needs to actually be fairly
25:02low latencies but that commits you to a
25:04smaller model so you're going to use a
25:06super large model uh and and you know on
25:08the other hand on the sort of pool model
25:11like I'm asking something I'm actually
25:13going to formulate my question in a way
25:15that the AI could understand it better
25:18um I'm going to be able to wait to get a
25:20much uh to give the AI time to think or
25:23to compute and so I would say it's it's
25:25not either or you have to do both and I
25:28think that's the fundamental ux
25:29Innovation we brought to the space is
25:31that we called Society of models copilot
25:34uses a single model an applet uses like
25:36three different models of different
25:37sizes so the smallest models is the
25:40model you know over your shoulder
25:42continuously giving you suggestions then
25:44a medium-sized model to do the
25:46Transformations and things like that
25:48then a super large model the kind of
25:50model that you would want to talk to
25:51chat gbt style model that's great and I
25:53guess that shows a concrete example of
25:57the capabilities you get from developing
25:59your own models and really being able to
26:01fine-tune them to specific use cases
26:03let's zoom out a little bit and talk
26:06about you know right now we're at the
26:07beginning of this phase where AI is
26:09being applied to code copilot came out
26:12what was it like a year ago or so Ghost
26:14Writer came out recently but we're in
26:17you might say the first inning so if we
26:19extrapolate to maybe say Ghostwriter V4
26:22or co-pilot V6 many years from now I
26:25want to think about how you see the
26:28overall environment for developers uh
26:31charging or evolving so on one hand I
26:34could see how people might argue you
26:36know what having this technology being
26:38able to just tell a computer hey I want
26:41to create this app go go basically build
26:43it for me or at least build the
26:45fundamentals for me is going to create
26:46this wave of really low quality
26:50developers who don't really know what
26:51they're doing we're just relying on this
26:53AI crutch to be able to do a lot for
26:56them I could also see an argument though
26:58if we're talking about some of these
27:00suggestions within their development
27:02environment this actually creating
27:04better developers they understand a
27:05little bit more about how their code is
27:07being executed maybe they more quickly
27:09pick up new skills new languages new
27:11Frameworks because they have this
27:14assistant and so just curious to know
27:16how you see this all evolving and again
27:19this kind of ecosystem of Engineers
27:22changing yeah I mean anytime you make
27:25something more accessible you just get
27:27the entire gamut of things like
27:31Instagram made photography way more
27:33accessible and you get like a like a
27:36long tail of crappy photographs but you
27:39also discover people who wouldn't have
27:40never been photographers right you know
27:43everything in the world is like that
27:44like YouTube most of the creators don't
27:47get any views but then a few creators
27:51pre-youtube would have been just a kid
27:54in Sweden not a super celebrity right
27:56and so making things accessible
28:00gets you a lot more people involved and
28:03the trade-off you're making is that
28:04you're going to get a lot of noise
28:06but you're gonna discover tan that
28:08wouldn't otherwise be discovered and I
28:11think that's good for the world
28:12so with software engineering becoming
28:15more accessible because of AI I do think
28:18we're going to get a lot more developers
28:20I think we're going to get way way more
28:22devops probably 10x so right now there's
28:25like 30 million developers in the world
28:26there's probably 300 million developers
28:28by the end of the decade and the way I
28:31see so the developer markets evolving to
28:34accommodate this new technology is that
28:37there's going to be this bit of a
28:39bimodal distribution so bimonial
28:41distribution meaning there's no middle
28:42end there's a large tail on both sides
28:46right so the middle end is the sort of
28:50um the glue code plumber type developer
28:53that we have today I think that'll go
28:55away and the reason that'll go away is
28:57because platforms are going to be a lot
28:59more expressive they're going to be able
29:01to be programmed using natural language
29:03a lot of the cloud platforms are just
29:05building better abstractions things like
29:07replica will just like make back ends a
29:09lot more accessible and so that the
29:11middle end I think will probably really
29:13disappear because of that because of
29:16pressure from both sides the front
29:17engineer is just going to get way way
29:20so front engineer will be able to build
29:23full stack products just because they
29:24have access to all this really powerful
29:26platforms and they're going to be able
29:28to just produce a lot more code and be
29:30able to use AI in every part of the
29:32coding process whether it's testing cicd
29:36design everything is going to be sort of
29:39powered by AI it just made a lot better
29:41including quality control by the way so
29:44that's the sort of on the front end side
29:45and then on the sort of back end low
29:48levels of platform engineering I think
29:50those people are just gonna get a lot
29:52more powerful like Imagine John Carmack
29:55right Jonah Carmack is what we call 10x
29:57developer today imagine giving him a
30:00army of AI developers that he could
30:04delegate work to that he could you know
30:09ask questions of that
30:12um you're just going to make him 100x
30:141000x more proactive and and so you can
30:18have maybe fewer of those sort of low
30:22level 10x Engineers but they're going to
30:24be a thousand X engineers and so maybe a
30:27single company would need two or three
30:29like Elite engineers
30:32and then maybe dozens of front-end
30:34Engineers kind of building all these
30:35products and maintaining what the
30:37customers see but those the core group
30:40of elite Engineers their impact is just
30:43going to be tremendous they're going to
30:45be demanding a lot more money they're
30:46going to be making a lot more money so
30:48if if engineering is really your craft
30:51it's not going away and it's actually
30:53you're going to be able to actually be
30:55accentuate your power I think it's funny
30:58that you mentioned the 10x engineer
30:59because a lot of people make fun of this
31:01concept of a 10x engineer because you
31:04don't see for example 10x plumber
31:06because you're a little bit Limited in
31:07The Leverage that you can get with your
31:09time in a scenario like that but we know
31:12that one of the reasons they're there
31:14truly can be 10x Engineers is because
31:16software can actually give you leverage
31:18and so if you enhance the power that
31:20software can provide it's actually not
31:23crazy to your point so imagine a 100x
31:25engineer or a thousand X engineer if you
31:28basically have these like robot
31:30developers that you can write in and
31:33apply in a specific Direction it's scale
31:36right yeah exactly it's the concept of
31:38scale it's concept of like this is what
31:41technology has allowed us to do since
31:43the dawn of humanity
31:45you know you go from collecting crops
31:48manually by hand and doing everything to
31:51using animals to then using robots and
31:55now a single farmer can maintain entire
31:57Acres of farms just because of all the
32:00technology they're using that's
32:01technology right I mean the 10x
32:04denialism is is kind of funny to watch
32:08like every one of those people that say
32:1110x Engineers don't exist knowing their
32:14heart of hearts that 10x Engineers exist
32:16and they probably worked with someone
32:17they highly respected admire the reason
32:20they say that is just politically
32:21motivated they just don't like that the
32:24fact that some people can be better than
32:26other people and that's just a fact in
32:28life it's uncomfortable to accept that
32:31some people actually do bring more value
32:32to different organizations
32:34all right so with the bimodal
32:37description that you've shared I have to
32:40ask the question which is just simply is
32:42it still worth learning to code to me it
32:45feels like it's more worth it today
32:46because it's a tie that uh raises old
32:50boats right like we said the 10x becomes
32:52a thousand Acts but like a 0.1x becomes
32:56a you know 10x right so if you're
33:00someone who's previously maybe whose
33:04impact with Coatings can be very minimal
33:07now it's going to be meaningful because
33:10there's this Rising tide previously like
33:13you learn a bit of coding you learn how
33:15to plug together some Frameworks and
33:17create some UI going from that to like
33:20doing a little bit of parsing for
33:21example parsing some texts on manonado
33:26right now you can just use gpt3 to do
33:29right so you're doing the basic coding
33:32and then anytime you find something
33:34that's like a little difficult you can
33:36plug in gpt3 in that place
33:38and so I actually think that programming
33:41becomes more fun and impactful because
33:44of that so it's like totally worth doing
33:46even more worth doing than than before
33:49because you're going to get more done
33:50you're going to get not going to get as
33:55probably learn I wonder if you have any
33:57idea it on this but when I learned I
34:00don't know why I did this but I decided
34:02to track how many hours I was spending
34:04from when I first started to when I
34:06actually felt proficient and for me that
34:09number ended up being 300 hours
34:12um maybe I was slow maybe as fast I
34:14don't know but it kind of shocked me in
34:16a way because if you actually distill
34:19300 hours down to if you were learning
34:23full time which I know not everyone can
34:24do that's less than two months and now I
34:27think to the tools we have today and I'm
34:28like gosh like we can probably do it way
34:30faster than 300 hours so any thoughts on
34:33how quickly someone can actually get to
34:35proficiency with the tools that we have
34:38so I know you're you're a fan of uh
34:40bulging he was on your podcast he you
34:43know he has this uh beef with media and
34:45so he wanted to show that The New York
34:48Times is a bunch of bots right it's
34:50and so he he put in a bounty on replica
34:54he was one of our early adopters of
34:56bounties and he wrote it's like I want
34:59uh uh someone to build uh the GPT times
35:02instead of like New York Times
35:05but um just totally based on tweets you
35:09give it a tweet and it goes and generate
35:12an article written in the style of the
35:15New York Times so actually uh it is it
35:18should be up now so if you go to the
35:22so that's basically like the site that
35:25was built using uh the bounties program
35:27and the the person who built it was on
35:31day 80 I think of 100 days of python so
35:34100 days python is one of our programs
35:37to learn how to code and we say you can
35:40do a day of python in basically 20
35:43minutes so if you was on day 80 spending
35:4620 minutes how many hours is that that's
35:4926 hours so in 26 hours they were able
35:52to do uh to build an entire website
35:57and earn 1800 from from biology
36:00right so that's that's I think your
36:02answer I mean maybe that uh guy was an
36:05outlier but we're seeing it all the time
36:08it's taking weeks for someone to get to
36:11a place where they can build things and
36:13I think that's really what matters I
36:15think that's incredible because I think
36:17one of the best decisions I made was
36:19learning to code because you just to me
36:21it really was like a foundational shift
36:23in understanding the world around me
36:25because again so much of what we do is
36:27digital it is online and so even if you
36:30don't want to go and be the developer
36:32that's hired to build products just
36:34having that understanding and it's kind
36:36of crazy when you think about it if you
36:37were to actually position that like
36:39before and after to people and say well
36:41now it only takes 30 hours
36:43that's like you could do that in a week
36:45you could take four days off from your
36:49go through this and I know there's a
36:51difference between the 20 minutes a day
36:52and you know a straight shot but that's
36:55pretty incredible so the final thing I
36:57want to ask you about when it comes to
36:58AI is how this might shift
37:03how quickly we can code or how quickly
37:05we can learn to code but how this may
37:07fundamentally change what we can do with
37:09code so let me give you two examples to
37:11kind of shape up the question the first
37:13is when we learned that computers were
37:16better at chess than humans we didn't
37:19just learn that fact we also learned
37:22that there were all these other moves or
37:23modalities to chess that we had never
37:25considered right it kind of reframed the
37:29and then another example is one that our
37:33games team shared recently where
37:34basically if you've heard of the Flight
37:35Simulator game you can basically like
37:38you know in the digital world fly and
37:40land an airplane and fly it around the
37:42earth and it's got this 3D model of the
37:45Earth that was built off of the 2D model
37:47from Google Earth and that 3D model was
37:51built with AI and it only could have
37:52been built with AI so those are two
37:55examples of how this technology not only
37:58you know maybe made things faster but
38:00actually made things possible if that
38:03makes sense things that weren't prior
38:05available so have you seen any glimpses
38:07of that or any thoughts around just like
38:09having you know we could position it as
38:11somewhat of a superpower having that
38:13superpower accessible what does that
38:15change man this is such a brilliant
38:17question like I actually haven't thought
38:20about it as much and I think I should
38:22like you know your example with chess
38:25also happened with go you know with the
38:28leastidol game do you remember that move
38:30that basically what happened is the AI
38:33made a strange move that made it give it
38:37a disadvantage in the near term and a
38:39huge advantage in the long term and they
38:41were so confused about it because it's
38:42almost like alien Intelligence intruding
38:45on this thousand-year-old game right and
38:49and producing this fundamentally novel
38:51move so I don't think we've seen that
38:53entirely in in programming yet but I'd
38:57be definitely on the lookout for that
38:58what I would say we've seen is that
39:02tasks that were previously would require
39:06a ton of work like a ton of insane
39:09amount of laborious work getting done
39:11like that you know for example the
39:12parsing uh question gpt3 is incredibly
39:15good at parsing if you give it a
39:18malformed Json it will still parse it
39:20writing Parcels is one of the hardest
39:23things you can do in programming writing
39:25Parcels in gpt3 is one of the easiest
39:28thing you could do you could just spend
39:2915 minutes in the open AI playground and
39:32so really that goes from a task that
39:34requires hours and maybe days and weeks
39:37of building and test thing to something
39:39that takes a 15 minute and so that's a
39:42phase shift in how we do things that's
39:45clear you know I I think there are ways
39:50in which like we haven't totally
39:52explored how to use LMS and programming
39:55like can you create back-end as a
39:58service using llms so basically
40:01Firebase but entirely using natural
40:04language so Firebase is this great
40:07project that Google Cloud acquired in
40:11Firebase allows you to have a back end
40:13without any back-end knowledge you just
40:15start storing data and it'll just
40:16retrieving data and it just works
40:19can you have a back end that's
40:21completely programmable using natural
40:23language can I describe my application
40:25and just write the front end for it and
40:27just have the back end taken care of I
40:29think that if that's possible that'll
40:31cut down on like how many Engineers you
40:33need on your team they'll come down on
40:35time to create a prototype and so I
40:37think that will be incredibly exciting
40:40but your question was more like what is
40:43something fundamentally was
40:46not possible that became possible I
40:50don't think we've seen that yet but I do
40:52want to think a little harder about it
40:54and really be opening my eyes right
40:57about it maybe one thing I'll mention is
40:59like One Direction that could happen is
41:02the action model have you looked at
41:04action models at all no no it'd be great
41:07if you could describe what they are yeah
41:09so Transformer models are the models
41:12underpinning large language models right
41:15so that's what everyone knows GPT GPT
41:17stands for generative pre-trained
41:19Transformer so you take a Transformer
41:21that's a type of model architecture and
41:24you threw a huge Corpus at it and it
41:27learns a ton of things and then you can
41:29program it using prompting right that's
41:31basically what GPT is now you can do
41:33a different type of Transformer where
41:35you take a Transformer instead of
41:38throwing a ton of text at it you throw
41:40it throw a ton of actions at it
41:43so what what are actions so for example
41:45for example actions could be
41:47all the mouse and keyboard events in a
41:50so I'll just take all that
41:52is raw stream of data and just put it
41:56train in Transformer to do that so now
41:58the Transformer encodes knowledge about
42:01how to use a browser
42:03that's wild all right what can you do
42:06with that you can now instruct the
42:09Transformer to go book an Airbnb for you
42:13to go do like more complex tasks like
42:16find the place with the best weather in
42:19this time of year and book an Airbnb for
42:21me and my family so that's quite
42:24I would say that wasn't possible before
42:27um so I think that's a fundamental
42:29area if if action Transformers
42:35mainstream and as powerful as GPT
42:39then I think it'll unlock a new
42:41programmable platform
42:44because now you can sort of you can sort
42:46of it's almost like everything has an
42:48API suddenly yeah that's fascinating
42:51actually because yeah in the past it's
42:53like a specific database has an API that
42:56you can plug into but now this idea
42:57where anything in the browser anything
43:00on the web could potentially be
43:01transformed into its own API without
43:03setting up the like specific API
43:06yourself like that the language model
43:08could actually figure that out that's
43:10fascinating and I think to your point we
43:12haven't seen these specific examples of
43:15wasn't possible now possible yet because
43:17we're really early but even just as
43:19we've discussed so far the amount of
43:23foundational development that is
43:26required from Engineers today that will
43:28soon be abstracted just opens that time
43:31up that brain capacity up to apply to
43:34something new so I think it's a little
43:36bit inevitable for something to emerge
43:39we're just not sure what that something
43:40is I want to ask you very quickly about
43:42mobile as well because up until now
43:46desktop has really captured most of
43:49development I can't think of many
43:50developers who code on their phone but
43:52it sounds like this might be changing in
43:54fact I heard you actually say that you
43:56think millions of people will code on
43:58their phone yeah I mean if you think
44:00about what we've been talking about with
44:01AI the idea that your primary
44:04development experience will include a
44:07big portion of chat and like that the
44:10best way to do chat is on your phone
44:12everyone texts on their phone everyone
44:14talks on their phone and I think that
44:18being able to generate software by
44:20talking to your phone
44:21is going to be a very clear thing that
44:24will happen at minimum being able to uh
44:27like instruct your AIS while you're on
44:30the go and review their work that's
44:33obviously going to happen but as typing
44:36becomes less of an issue on the phone
44:38then actually making complete pieces of
44:40software just will make a ton of sense
44:43because you're just prompting and you're
44:46re-prompting and you're reviewing and
44:49then prompting and reviewing or
44:50prompting that area of loop is very
44:54clearly going to work very well on the
44:59phone and actually it could be more
45:00delightful on the phone because you can
45:02you can do a lot of swiping set of
45:04activity like I you know with my team
45:06with joke we called like Tinder for code
45:08where so if you prompt Ai and it gives
45:12you a piece of code you can say yes or
45:13no say no then give you another another
45:16piece of code and maybe you give it
45:18another prompt or maybe it asks you
45:20another questions so that iterative
45:22piece of making software using AI I
45:27think really lends itself nicely to the
45:29the concept of a phone and then the
45:33other thing is that it's not just the
45:34phone it's a tablet like how crazy it is
45:37stuff that uh you know we don't like
45:40have good IDs on the tablet it's kind of
45:43surprising that you know Microsoft
45:44hasn't made vs code for the tablet so
45:47we're the first kind of major IDE for
45:50the tablet as well and I think you can
45:52do everything that you do you could do
45:54on desktop on tablet you can attach a
45:56keyboard to that and you can go to any
45:58coffee shop in the world everything
46:00happens at Cloud your storage in the
46:03cloud your ai's in the cloud you don't
46:06need that much local capability and you
46:09can just write software on that yeah
46:12yeah I haven't thought of that actually
46:13but I it's true I have never seen
46:15someone coding on a tablet just like
46:17I've never really seen someone coding on
46:19mobile but let's see how that evolves
46:21another feature you've developed that
46:23you alluded to earlier was bounties so
46:25can you tell listeners a little bit more
46:27about what this is this idea of bounties
46:30and also just a little bit more about
46:32how it's going so far yeah um so
46:35bounties is part of our sort of
46:39portfolio of products that make it easy
46:42to make software so we realize that not
46:45everyone in the world wants to be a
46:47coder and I think a combination of the
46:50network effects on droplet and being
46:51able to discover a lot of software
46:53Engineers like the kind of guy who made
46:56the GPT times for biology and the AI
47:05um sort of what's happening in
47:07currencies with a cryptocurrency or
47:09centralized currency there's a lot of
47:11interesting things I think it's a sort
47:12of a bit of a trifecta that allows us to
47:16kind of build what I think is a fluid
47:18software labor market
47:21um uh like a big fan of uh this um
47:24Theory uh called the theory of the uh of
47:28the firm the fundamental observation is
47:30that full-time employment is a bug it's
47:33not a feature of the market the reason
47:34why full-time employment exists is
47:37because the transaction cost of doing
47:40something is really high Uber is an
47:42example of something bringing the
47:44transaction costs almost down to zero
47:46and by doing that it creates a a huge
47:49amount of flexibility in the in the
47:50market so anyone can enter the market
47:52anyone can do the work and they can do
47:54it at their own terms and there's no
47:56binding contract uh between the
47:58different parties software has been
48:01something that's been like very
48:03hard to actually contract out and when
48:06you do contract out you get a lot of
48:08problems a lot of issues running the
48:10software you get a lot of quality issues
48:13so the fact that replied is a fully
48:15integrated place with high quality
48:17software Engineers using it allows us to
48:20be a place where someone can go put a
48:22description for a piece of software then
48:25they need to get done and then a
48:28developer high quality developer can go
48:31and using AI can make that software very
48:35quickly and we're seeing people making
48:37software in 30 minute increments so
48:40someone the other day put a landing page
48:42a request for a landing page up they
48:45attach the figma file they got it back
48:47in 30 minutes that's really never been
48:49done before and why shouldn't have it
48:52existed uh like that before and so so I
48:55think the combination of all the
48:56Technologies we're building allows us to
48:58create this this Marketplace how are you
49:01thinking about this bounce T's
49:03Marketplace among the existing
49:05marketplaces because
49:07it sounds like and let me know if I'm
49:08wrong this is more peer-to-peer I post a
49:11project someone within the community
49:12puts their hand up or maybe they're
49:14matched but when I think about the
49:17existing marketplaces for developers out
49:19there their job in many of these cases
49:22is to vet a bunch of developers and say
49:24okay you have this skill set you're this
49:27good of a developer and then also vet
49:29their clients and say okay this is you
49:31know a good project this is worthwhile
49:32for us to introduce into this
49:34Marketplace and then they match people
49:36and then of course those marketplaces
49:38take some commission for doing so and so
49:42since many of these already exist like
49:44how do you see bounties as an ecosystem
49:47differentiating or maybe providing
49:49something new to people within the
49:53community I think it's already
49:54differentiated and the reason it's
49:56already differentiated is because the
49:58development environment is built into
49:59the system if Uber was a Marketplace
50:03to connect here to people and then they
50:05have to go get their own car right like
50:07so you have to go meet someone at a
50:10coffee shop and then you and them go get
50:12a car and then go all right that's an
50:14absurd example but that's what happens
50:16at upwork or some of these marketplaces
50:19like I asked for a piece of software
50:21and then you go make it in something
50:23that I don't know and then you send me a
50:25zip file and then what do I do with that
50:28right and replied I just sent you a link
50:30a link to a computer that's running
50:32that's running your application that's
50:34like fundamental Innovation on top of
50:37that and then like all the services just
50:39being integrated right there like you're
50:41opening out AI API key the cloud runtime
50:44like all that stuff the database we just
50:47added uh postgres support just
50:49replicating this complete platform just
50:52like makes this process a lot more
50:53efficient that's actually a great point
50:55because when I think about other
50:59scenarios where people hire developers
51:01to create projects I think one of the
51:04massive gaps is as you said the
51:07standardization but it's also if we
51:09think about the AI tools that you're
51:10integrating if someone gets a project
51:12and they can actually read the code and
51:15maybe they're not a developer but have
51:16some level of understanding of like this
51:18was what I got back I think that's
51:21actually massive differentiation because
51:22in the past they just get back this code
51:24that they can't understand at all if
51:26they need to refactor anything or change
51:28anything or get a new developer to work
51:30on at they really struggle and they
51:32really also struggle to have some sort
51:34of sentiment in terms of how good the
51:38exactly and we want to do like
51:42we want to create more market dynamics
51:44in the future that are more interesting
51:47like for example like
51:49uh you know this concept from crypto
51:54people like being able to stake their
51:58in order to say I'm going to build this
52:00I'm going to build this better than
52:01anyone so I'll put up a bounty for a
52:05and you stuff will say I'm so sure of my
52:08I'm willing to put up a hundred bucks
52:10that'll be able to deliver on time and
52:12if I don't you can take my hundred bucks
52:13I love that and so I I think there's a
52:16lot of innovation to do in in market and
52:19you can also integrate some kind of AI
52:22things on top of that so for example you
52:24can have a sort of AI project manager
52:27where now it takes a one Bounty and like
52:29splits it into 10 bounties so we have
52:33this experimental product that Twitter
52:35about recently but basically
52:38the more you put in a bounty we actually
52:43the scaffolding for the code what's you
52:45know scaffolding and programming is
52:46basically the structure of the code
52:49and then every function is like that is
52:52not implemented that function could be
52:54its individual bounty
52:56so I think once you add all these sort
52:58of Innovations together I think you're
53:00going to get this like super fluid
53:03Market of AIS and humans that it you can
53:08go from an idea to a product just like
53:10that and crucially like you said the
53:12product that you're going to get is
53:14something you're going to be able to
53:15iterate on the future and get more
53:17Bounty creators to contribute to it's a
53:20living artifact that's working on the
53:22platform as opposed to like again like a
53:24zip file for over the wall to you I love
53:27the idea of people being able to like
53:31legitimately bet on themselves if they
53:33want to participate if they want to take
53:37actually when I talked to biology in our
53:39interview we talked about this idea of
53:40evidence versus confidence so someone
53:42can have a lot of confidence and say hey
53:44I have these requisites on my CV or I
53:47have this degree from somewhere and that
53:49distills a lot of confidence in other
53:50people that they can get something done
53:51but what's better than confidence is
53:54evidence that you can actually get it
53:55done and if you if you can show hey I
53:59have I've done this project before like
54:01literally here's my work and also I'll
54:03put as you said put 500 down to say hey
54:06if you don't like it like you can have
54:07my 500 like that is more than confidence
54:10that is like evidence that you can do
54:12the work and you're willing to put your
54:14money where your your skills are with
54:15that said let's move on to the very
54:17final section and we'll just rapid fire
54:18go through these questions they are
54:21questions that I've actually seen Sam
54:24from openai talk about as four different
54:27important questions that he thinks are
54:31relevant to this Advent of AI especially
54:34as it progresses again taking the lens
54:36of not just Ghost Writer V1 but
54:38Ghostwriter V10 quite a ways away from
54:41now and how how good how proficient it
54:46I'll just read you each question I just
54:48want to get your like raw take on each
54:51the first question is you know with that
54:55how how do you think that fundamentally
54:57changes Society something I care deeply
55:01a quality of opportunity so I think just
55:05the ability that anyone in the world
55:07being able to contribute and build
55:10something is just beautiful it's just
55:12amazing it allows to include everyone
55:15and anyone who's willing to work hard
55:18and that's like the real
55:20americanism in my opinion that's like
55:22what what America is built upon is the
55:24idea is that if you're willing to do do
55:26the work you're going to get the chance
55:27at doing the work and if you do the work
55:29right you're going to get rewarded for
55:31it and I think that's beautiful
55:33however I I think that comes with also
55:37increased inequality I I think there's
55:39going to be the pi is going to be bigger
55:41but also there's going to be differences
55:43between again like that 10x different
55:46becomes a thousand acts difference
55:48and that's going to create some
55:50political issues because I think it's
55:53going to create envy and it's going to
55:56accentuate some some differences and I
56:00think that part's not going to be fun
56:01and um and I think reasonable people
56:04should be able to like say it like it is
56:08and that's why I think
56:10you know the 10x denialism
56:13um is a small sort of symptom of a
56:16larger issue in the issue of just being
56:20of people being rewarded more because
56:23they have a a better ability of doing
56:26something so that that's just going to
56:28be a big issue that Society is going to
56:30have to deal with I agree exponential
56:33technology means potentially exponential
56:36gains from certain people and not others
56:39and that's not necessarily inherently
56:40bad especially if the pie is growing but
56:43yes I agree with you that's going to be
56:44something that is divisive that actually
56:47relates to question number two which is
56:49how do we ensure that it benefits us all
56:54now that does not mean it benefits us
56:56all equally but how do we ensure that
56:58this technology is something that widely
57:01benefits the people participating in
57:04society I think as much as possible
57:06having competition is going to be
57:08important I think a world in which just
57:12one company controls the biggest AI
57:15it's not going to be great but I neither
57:17do like one two or three companies I
57:19think the lesson we learned from the
57:21last generation of big Tech is that
57:24especially if there's like a cultural
57:26monoculture like the Silicon Valley
57:28you're going to have similar decisions
57:31made and so how do we prevent that in
57:34the in the world of AI That's something
57:36that's very important I think open
57:38source uh plays a part in that I think
57:40as much as possible you know pushing
57:43this technology to be easy to run and
57:46easy to develop but not necessarily
57:48something that's sort of closed off and
57:49like a few people can control I think a
57:52lot of what's called AI alignment today
57:54is not really aligning with what the
57:58average human being wants it's aligning
58:01with like what the sort of
58:04Silicon Valley average sensibility is
58:07which I don't think it's good like I
58:10think that we should try to build as
58:12much as possible neutral technology
58:14so my bias is going to be more freedom
58:17more decentralization yeah ultimately
58:20these models are a tool the tool itself
58:22is neutral but the application of the
58:26tool or the technology often isn't
58:28neutral or it doesn't always happen to
58:30be neutral and they hate they hate the
58:33fact that it can be neutral and they
58:35think that neutral is bad and I think
58:37that's where we really need to push back
58:39is that no actually it should be neutral
58:42and then the user have to decide what to
58:44do with it right yeah I think that's a
58:46good modifier I'm going to combine the
58:49final two into one question you can
58:51maybe touch on either or both of them
58:53but the question is given again if we
58:56imagine the exponential nature of this
58:59technology the evolved form of it how do
59:02we deploy it safely and you can kind of
59:05interpret the word safe in in your own
59:07terms how do we deploy it safely and
59:10also maybe how does governance play a
59:12role in that you know I I just like the
59:15word safety and I think a lot of the
59:18problems in the world today and and so
59:23the way our world is shaping up is this
59:25concept of safetyism is that yeah this
59:29this concept was just like that like
59:30companies need to ensure our safety and
59:34then they encroach on our freedom to to
59:37ensure our safety like it's such a
59:39stupid topic but there was a recent gas
59:42stove debate right you know whatever
59:45government official that kind of said
59:48like they might consider Banning gas
59:49stoves that that kind of impulse of like
59:52we know better than you sort of the
59:54nanny State impulse I think it's bad and
59:56I think it actually every time
59:58governments have created sort of
01:00:01atrocities have come from that idea like
01:00:03the government wanted uh the American
01:00:06people to be safe from the Japanese
01:00:08that's why they create Japanese
01:00:09internment camps right you can use
01:00:12safety to do the most abhorrent thing in
01:00:15in the world and so I think we need to
01:00:17be skeptical anyone's talking about
01:00:18safety they probably want something bad
01:00:21for you and so uh that that would be my
01:00:24bias is that safety is bad or at least
01:00:26when when they want to impose something
01:00:29in the under the guise of safety I would
01:00:32be super skeptical and then on the on
01:00:35the question of of governance I would
01:00:37also be skeptical of that so maybe I'm
01:00:41uh impulses here but um like like I
01:00:46don't think anyone is inherently
01:00:48responsible for governance of Technology
01:00:51look I think government has a plate has
01:00:54a role to play we got from a sort of
01:00:56move slow and that's good that it moves
01:00:59slow because it needs to learn what's
01:01:00happening what's the system is like you
01:01:03know if the government moved really fast
01:01:04and regulating cars they would have gone
01:01:06in super wrong and it's it's now like
01:01:09making self-driving move a lot slower
01:01:11than it should be it's just because the
01:01:13regulation just mounted up but I I think
01:01:16the the sort of the process in the US is
01:01:18actually fine for regulation just like
01:01:21move it slow and let's learn what's
01:01:24happening in the real world and then
01:01:26let's let's react to that and let's have
01:01:28a reasonable debate and discussion I
01:01:31think a democracy at the end of the day
01:01:33will arrive at the right decisions if
01:01:35there's a sufficient uh freedom and I'm
01:01:37a big fan of debate and vicious debate
01:01:40in order to arrive uh the truth but but
01:01:44that just going to take time and and the
01:01:47problem now is that there's some people
01:01:49talking about AI takeoff being so fast
01:01:52that you know you need to react to it
01:01:54really quickly and I think that'll get
01:01:56us into trouble actually if you're going
01:01:58to force politicians to understand this
01:02:00technology before it's even deployed
01:02:02you're really asking for trouble yeah
01:02:04but I do think on the flip side of
01:02:06side something beautiful that's happened
01:02:08in the last year or so is that the
01:02:11technology has taken off really quickly
01:02:12but it also has done so in a way where
01:02:14it's gotten in the hands of consumers
01:02:16and you see this if you look at the
01:02:18charts of like chat gbt for example
01:02:20their speed to 10 million users was so
01:02:22much quicker than you know Facebook or
01:02:24SnapChat or other other apps in the past
01:02:27and so I think while it is inviting some
01:02:30skepticism to say oh my gosh this is
01:02:32happening so quickly this is scary we
01:02:34don't know the implications I also think
01:02:36you're going to see a level of pushback
01:02:39right because at that point you're
01:02:41taking something away from people the
01:02:43superpower as we talked about it it's
01:02:46important to notice stuff that nothing
01:02:48as catastrophic happened right like this
01:02:51technology is now massive deployment and
01:02:53nothing as bad as the sort of less wrong
01:02:56set of AI safetyism sort of part of the
01:03:00debate have said it's going to happen
01:03:03um and uh and so it's it's important to
01:03:07recognize this because the people that
01:03:10are going to be argue doing for extra
01:03:12controls and everything are gonna you
01:03:16know always paint a future picture of
01:03:19catastrophe but like the question to ask
01:03:21them is like like okay now there are
01:03:23millions of people already using this
01:03:25like nothing really bad happened so you
01:03:28know what are you actually worried about
01:03:30in concrete terms yeah and I think your
01:03:32time Dimension is so important here
01:03:34because to your point you look at
01:03:35Technologies like cars and they were
01:03:37worried in the early days about cars
01:03:38scaring horses and I think just having a
01:03:41layer of humility to say we don't know
01:03:45how this is going to shake out I'm super
01:03:46awesome to say about the future uh of
01:03:49this technology and everything that's
01:03:50happening I think I think it's really
01:03:52lovely and I think it like uh it's like
01:03:55the best time to be a builder
01:03:58thanks for listening to the a16z podcast
01:04:01if you like this episode don't forget to
01:04:03subscribe here on YouTube to get our
01:04:04exclusive video content we'll see you
