Sponsored by
Show notes
Transcript
0:00:00 Intro
0:00:00 there's a group, and this is what I call
the Barefoot Developer Group, who really
0:00:04 do want more agency over their computers.
0:00:05 They're, like, currently using
Airtable and Notion and like
0:00:09 spreadsheets and like doing really
crazy, complex things with them.
0:00:12 But they also hit the
limits of these apps.
0:00:14 And they keep being like, well, how
do I make it like do this thing?
0:00:16 And I'm like, well, you can't, like this
other company controls that software.
0:00:19 You can't.
0:00:20 add that feature onto it.
0:00:22 but these people, if given access to
a language model, and kind of taught
0:00:25 how to prompt it in a way where they
can clearly ask for what they want,
0:00:29 it can write the code for them.
0:00:30 welcome to the Local First FM podcast.
0:00:33 I'm your host, Johannes Schickling,
and I'm a web developer, a
0:00:36 startup founder, and love the
craft of software engineering.
0:00:39 For the past few years, I've been on a
journey to build a modern, high quality
0:00:43 music app using web technologies.
0:00:45 And in doing so, I've been falling down
the rabbit hole of local first software.
0:00:50 This podcast is your invitation
to join me on that journey.
0:00:53 In this episode, I'm speaking to Maggie
Appleton, a designer, anthropologist,
0:00:58 and developer who's recently explored
the world of local first software
0:01:02 by giving the closing keynote of
the first local first conference.
0:01:06 In this conversation, we dive deep into
the topics of her talk, including home
0:01:11 cooked software, the ideas behind barefoot
developers, and how AI complements
0:01:16 local first software development.
0:01:17 Before getting started, also a
big thank you to Rosicorp and
0:01:21 Expo for supporting this podcast.
0:01:23 And now my interview with Maggie.
0:01:26 Cool.
0:01:26 Hey, welcome, Maggie.
0:01:28 Thank you so much for making
time today to come on the show.
0:01:30 How are you doing?
0:01:31 Hey, yeah, I'm really good.
0:01:32 I'm excited to be here.
0:01:33 Thank you so much for having me on.
0:01:35 So for the few in the audience who
don't know yet who you are, would you
0:01:38 mind briefly introducing yourself?
0:01:41 Sure.
0:01:41 Uh, I am a product designer
slash developer of some sort.
0:01:45 and I've been working
in developer tooling.
0:01:47 for I think my entire career at this
point, so pretty much 10 years, and I do
0:01:50 a lot of work that's like doing designs
for development tools, but most recently
0:01:55 I was a designer at a company called
Elicit where we were using language
0:01:58 models to make software for scientists
and researchers, so that dropped me into
0:02:02 the world of language models and AI and
machine learning, so I've been thinking
0:02:05 a lot about that for the past two years.
0:02:07 Amazing.
0:02:08 I want to give a little bit of
background since we just had
0:02:10 the local-first conference in
Berlin a couple of weeks ago.
0:02:14 And I've been getting in touch with
you, must have been at the beginning
0:02:18 of this year and through some common
friends we've been introduced.
0:02:22 And I was hoping to get you somehow
interested in speaking at the conference.
0:02:27 And when you and I were on a Zoom call and
I've been suggesting this topic to you.
0:02:32 You seem super polite, but you weren't
quite sure what you should say about
0:02:38 local-first since you haven't really
been building something with local-first.
0:02:41 And now a few weeks later, you
got back to me, said, okay,
0:02:46 you're, you're down to do it.
0:02:48 And you've delivered the closing
keynote and I think it was hands down
0:02:52 by most people, their favorite talk
of the day and really blew me away.
0:02:57 so we want to go into the talk,
momentarily, but maybe starting
0:03:02 first, like what convinced you
to do the talk and what has since
0:03:05 seemingly local-first pilled you?
0:03:08 Local-first and home cooked software
0:03:08 Yeah, um, So yeah, when you first
reached out, I mean, I had heard of
0:03:12 what, I had heard of local-first,
mostly from Ink And Switch's famous
0:03:15 essay, like, I follow a lot of Ink And
Switch's research, and I had read it,
0:03:19 but at the time was like, okay, this
has a lot to do with databases, and I,
0:03:22 I like, I do front end web development,
so I know nothing about back end.
0:03:26 I hadn't even thought to look into it,
or it just never appealed to me, because
0:03:30 Most of my motivation for development
is often, like, I want to make a
0:03:34 cool thing animate on, a screen, or,
like, I want to see, something move.
0:03:37 I like the reactivity, but, you
know, I'm, like, JavaScript.
0:03:40 so I'd never thought of backend
as an interesting area or really
0:03:43 understood, like, why I would
want to be interested in it.
0:03:46 so I'd read the essay, been
like, cool, that's not my thing.
0:03:48 I'm not one of these backend people.
0:03:50 Um, so when you reached out, I
had to go Googling again and be
0:03:53 like, what, what is local-first?
0:03:54 Like, what is this concept?
0:03:56 Like, why would I care?
0:03:57 because I was compelled by a lot of
the people who I saw, were going to
0:04:01 go to this event, or were involved
in the local-first community, were
0:04:05 people who I knew were working on
really interesting stuff and who had
0:04:08 really interesting original thoughts
about ways we should build software.
0:04:12 Not just from a technological
perspective, but from like a
0:04:14 cultural, social perspective.
0:04:16 Which is really the thing I love, is
thinking about, like, who's building these
0:04:20 tools, who are they building them for,
what are their implications in society,
0:04:24 like, how does this affect equality and
economics and, like, everything else
0:04:28 that is not just, like, what code are
we writing and, like, you know, how
0:04:30 do we architect the app kind of stuff.
0:04:33 so once he reached out and I started
googling and thinking about it, I saw this
0:04:36 connection in things that I care about.
0:04:39 with things that I saw that the
local-first community cared about,
0:04:41 which is things like data ownership,
people being able to kind of own
0:04:44 and control their own software.
0:04:46 and things working offline for me is, it's
quite important in like a very sort of
0:04:51 banal personal way in that, like I spend a
lot of time on the tube, like doing stuff
0:04:55 on my phone where there's no wifi, like
a lot of time on the tube and a lot of
0:04:59 time in airplanes where there's no wifi.
0:05:00 So just personally, I like need a lot of
stuff to work offline first and, and I
0:05:06 had also an understanding that in a lot
of the world there's very limited internet
0:05:10 connection or it's very slow, like the
vast majority of people like might not
0:05:13 have constant stable fast internet access
all the time, and I hate the idea that
0:05:18 we're always designing for a context
where we do have that, and so we assume
0:05:22 that everyone does, but this is like,
like, Gets into the inequality stuff
0:05:25 and the like, you know, designing for
everyone in the world, you need to have
0:05:27 different systems than if you're designing
on a 16 inch MacBook in San Francisco.
0:05:32 Yeah, totally.
0:05:33 I've lived briefly for a
year or so in London as well.
0:05:36 And I do remember that I think
you had some connectivity, at
0:05:40 stations, but maybe not in between.
0:05:43 And sometimes you're also like, as
you're leaving this, the station,
0:05:46 your device thinks, Oh, I have some
connectivity and it tries to do
0:05:50 stuff, but this is actually the worst.
0:05:53 So I feel like a lot of software works
more reliably on a plane where usually
0:05:58 you're just entirely offline, but on
a train, it's much more tricky because
0:06:04 you have sort of this intermittent,
and unreliable connectivity.
0:06:08 And I think this is where
software really struggles.
0:06:10 So I know exactly what
you're talking about there.
0:06:13 But as you were preparing for the talk,
how did you find sort of like the, the
0:06:18 topic that you wanted to talk about?
0:06:20 It was definitely quite hard.
0:06:21 I mean, I liked this because right
when you kind of reached out and then
0:06:24 I liked that you said, you know, you
can talk about whatever you want.
0:06:27 I mean, it needs to be about
local-first, but like take
0:06:29 whatever spin on it you want.
0:06:30 That's really nice because I like that
kind of freedom in conference talks or
0:06:34 doing any kind of research where I can
go, Okay, I'm going to think about the
0:06:37 things that I think are interesting
trends that are unfolding or that
0:06:40 are topics that I care a lot about.
0:06:42 And I'll just find a way to like
weave them in because most of the
0:06:44 time everything is interconnected.
0:06:46 Like this is all one, you know, one
like software development philosophy
0:06:50 that we're all trying to figure out.
0:06:51 and it was once I figured out how that
local-first, I felt like had a strong
0:06:55 connection to the world of, what, um,
Robin Sloan came up with this phrase
0:07:00 called home cooked apps or home cooked
software that I had been collecting
0:07:03 information on for like a year at that
point, I have this note taking system.
0:07:06 I will just like collect examples of
every single time I see someone building
0:07:10 something that is just for them or just
for their family or just for their local
0:07:14 neighborhood, this very like local,
you know, just for me and my people,
0:07:18 not meant to scale, never going to be
a VC startup, just like I'm solving
0:07:22 my own problems kind of software.
0:07:24 I'd written essays on this before.
0:07:26 Like I wrote one called folk interfaces
about people kind of like reusing existing
0:07:30 interfaces to make these kinds of apps.
0:07:32 So I had a ton of content and a
lot of thoughts on this topic.
0:07:36 And it was when I looked at it and
I was like, Oh, like local software.
0:07:39 If you take that like at face value,
that is like local to me, my family, my
0:07:43 community, my neighborhood, my country,
whatever it is like near and close to me.
0:07:47 it might also mean the data lives
on my device and the way that we
0:07:50 use it in local-first software.
0:07:52 But I realized there was like, Oh,
there's this interesting interplay
0:07:54 where we can say local as in data or
local as in culture and community.
0:07:59 And once I figured out that connection,
I went, Oh, this totally can work.
0:08:01 Like I can build an argument
around this, that actually connects
0:08:05 these two together and can make.
0:08:07 a bit of a, an ask to the local-first
community to also care about this other
0:08:11 version of local that I care about.
0:08:12 So you didn't just go for software
for yourself and like your local
0:08:17 context, but you found a way to pose
the question in a much more globally
0:08:23 impactful way by, I'm not sure, I think
you introduced and like you came up
0:08:29 with a new concept with a new term that
you call like the Barefoot Developer.
0:08:33 And put that into perspective and into
context with local-first software, home
0:08:38 cooked software, would you, for, for
those in the audience who haven't seen
0:08:43 your talk yet, highly, highly recommend
watching the talk also, like not just
0:08:47 hearing the talk, but watching it.
0:08:49 You've, as a designer, I think there's no
one out there who's able to as a visually
0:08:55 compelling and clear way, find visual
narratives that makes this even clearer.
0:09:00 So I highly recommend watching the
talk, but would you mind briefly
0:09:04 giving the idea of the talk?
0:09:07 Sure.
0:09:07 Um, so yeah, I came up with this
concept called Barefoot Developers.
0:09:11 and this, name comes from the
concept of barefoot doctors,
0:09:14 which was this initiative in
Maoist China in the 1950s and 60s.
0:09:19 It was not actually run by Mao,
you know, he's a bad person, etc.
0:09:22 But like, probably someone else
in his cabinet who was very smart
0:09:26 started this movement of barefoot
doctors where they would take
0:09:28 people from very rural villages.
0:09:29 And this was at a time where
there was very low health
0:09:33 care access in rural China.
0:09:34 There's just tons of people spread out.
0:09:36 There's not a lot of expertise
in these local places.
0:09:38 And they would pick, like, one person
from each village, bring them into
0:09:41 the cities, and train them up in,
kind of, basic medical healthcare
0:09:44 practices, like, immunization, right?
0:09:46 Like, treating basic ailments,
just, like, giving people
0:09:49 nutrients, and this kind of thing.
0:09:51 And then send them back to their
villages, where they are, like,
0:09:53 local, everyone trusts them, everyone
knows them, and then they provide
0:09:56 healthcare for the people around them.
0:09:59 And this worked really well.
0:10:00 It, like, raised life expectancy a
ton, and the whole philosophy of it
0:10:04 was, like, we're taking, the power,
not away from, but we're distributing
0:10:08 the power from these, like, very urban
centers where they have all the medical
0:10:11 expertise out to everyone, out to all
these very rural, poor villagers who
0:10:15 now can have a better quality of life
because we're giving them access to,
0:10:19 to proper health care in a sustainable
way, this is kind of a famous project.
0:10:22 And people ended up using
this phrase of barefoot.
0:10:25 Insert your profession here
in a number of different ways.
0:10:27 There's like barefoot designers,
there's barefoot architects, and
0:10:30 the whole philosophy is redistribute
the power away from the urban elite
0:10:35 and into like the rural masses.
0:10:38 So, People advocate for this in all
kinds of ways, but I kind of looked
0:10:41 around and was like, no one's advocated
for this in development so far.
0:10:44 And it's the same philosophy as end user
programming, as the other word we would
0:10:48 usually use for this philosophy, is like,
normal people should be able to program
0:10:54 and have agency over their computers
to the same degree that programmers do.
0:10:58 Now, In reality, we know that's really,
that's a hard, like, difficult ask, right?
0:11:04 As programmers, we deal with
all this, like, complexity.
0:11:06 There's a reason it's a profession.
0:11:07 There's a reason it's a
highly paid profession.
0:11:09 There's a reason that not many people,
like, have enough skills and knowledge
0:11:14 to do it at a really high level.
0:11:15 It's just, it's incredibly hard, right?
0:11:16 Computers are super complex.
0:11:18 It's all these edge cases.
0:11:19 We know it.
0:11:19 Just like layers of complexity, just 60,
000 layers of complexity in every app.
0:11:24 Like we are struggling to
make the apps work well, let
0:11:26 alone some end user doing it.
0:11:28 but the argument is, you know, maybe
they don't have to be able to build
0:11:31 Facebook by themselves, but surely
they could build some software, some
0:11:36 programs that serve their needs, that
solve their problems in a way that no
0:11:40 giant VC funded corporation in California
ever would or would ever think to,
0:11:45 because people who are far away in very
rich urban areas don't understand the
0:11:49 problems of people all around the world
in all kinds of different communities.
0:11:53 And probably shouldn't try because
it's usually quite, uh, I don't know,
0:11:57 like colonialist in a certain way.
0:11:59 Right.
0:11:59 Trying to say like, Hey, you know,
whatever it is, a street seller in
0:12:02 Turkey or like a homemaker in Tokyo.
0:12:04 Some like bro, California, VC guy
shouldn't be trying to solve their
0:12:07 problems because he can't, he can't,
there's not enough Zoom interviews that
0:12:10 are going to give him the context to
solve their problems well, and they
0:12:13 should have the power and the tools and
the agency to solve their own problems.
0:12:17 , and you know, sure in the areas of
design and architecture and healthcare,
0:12:20 but also software, computers are part
of everyone's lives and they should
0:12:23 have more agency and control over them.
0:12:26 So, sorry, that was the long version,
but like, that is, that is the concept
0:12:29 of Barefoot Developers are, people
who are part of local communities who
0:12:33 have enough programming knowledge and
skills to build software for their
0:12:37 communities and solve problems for them.
0:12:38 It's not everyone, right?
0:12:39 Like, not everyone, like, nurses
and teachers and whatever,
0:12:42 don't need to learn to program.
0:12:44 But if you can have one person, like,
who's in their neighborhood or in their
0:12:47 school who can, then you get that person
to build all the software for them.
0:12:52 Barefoot Developers
0:12:52 I love that idea and, uh, I agree
that this is thoroughly needed.
0:12:56 and I think most software that we use on
a daily basis is , by a super wide margin
0:13:03 built by folks in San Francisco or New
York or like in some like Western hub.
0:13:11 but this is where not everyone
in the world is living.
0:13:14 And so if you're just also
look, looking at the density.
0:13:17 and distribution of datacenters If
you're looking into the global south,
0:13:22 you don't have, nearly, nearly, uh,
as many data centers, et cetera.
0:13:27 And you can only think about
like how poorly the Western
0:13:29 made software is running there.
0:13:32 But we could.
0:13:33 from a technological perspective,
built as an entirely different way.
0:13:36 So this maybe leads to a
question I'm curious about.
0:13:40 why hasn't the idea of a
barefoot developer, why hasn't
0:13:44 that been a thing before?
0:13:46 Why hasn't that existed before?
0:13:48 And why can't this change now?
0:13:51 See, it's interesting.
0:13:52 I think, it probably did exist before
in the earlier days of computing.
0:13:56 When, in a way, like, before GUIs, and
before, like, we got all the complexity,
0:14:00 all the wonderful complexity of, like,
the web, and, like, more sophisticated,
0:14:03 more complex software engineering stacks,
right, before we all had, like, Node and
0:14:08 NPM, like, I don't know, just backends,
and just, we were able to build much
0:14:12 more impressive stuff in, like, the
maturity and rise of personal computing
0:14:17 and software, but it did mean that it
became such a highly specialized, complex,
0:14:21 professionalized, thing, versus I think
in like, let's try to figure out if these
0:14:25 are the right years, but the 80s to 90s
when it was still the command line, like a
0:14:28 lot of people could build their own little
things in like Linux, or just like little
0:14:31 command line scripts, it was probably a
lot more accessible that if you did on a
0:14:35 computer, you probably knew how to program
it versus now, that's not true at all.
0:14:39 so, and even I think in the early
days of the web, like, you know when
0:14:42 WordPress was still the hot thing?
0:14:43 I mean, it is still the vast majority
of the web is WordPress, right?
0:14:46 Still, vast majority of websites.
0:14:49 But I think early on, it was like, you
could just kind of get in there and like,
0:14:52 tinker on your own site in a way that
now feels a lot more inaccessible, like
0:14:56 now that we have, like, bigger build
chains, right, and all these frameworks.
0:14:59 It feels a little bit illegitimate
to just, like, spin up a WordPress
0:15:03 site to try to build an app.
0:15:04 You're probably going to reach for
React if you know what you're doing.
0:15:08 I mean, you should just reach for,
like, HTML and CSS, sure, like,
0:15:11 but most of us reach for React.
0:15:13 in a way that if you're just, like, a,
you know, someone in your neighborhood
0:15:17 trying to build something, like, set up
a community newsletter or, like, help
0:15:20 build some script for your local library.
0:15:23 you probably don't know what tools
to use or like how to use them well.
0:15:26 You probably are like just using
something like Airtable or Notion,
0:15:29 which is great, but you don't really
have the same agency and power as
0:15:32 someone writing command line scripts.
0:15:34 So, I have some sense that we had more
of this earlier on in computing, and
0:15:38 as it's gotten more complex, it's like
we, maybe we're on like a little bit
0:15:41 of a curve where it needs to come back
down the other side of complexity, or
0:15:44 we need to build tools that simplify
it for people to be able to build their
0:15:48 own software, but we're in some, what I
call like the high modernism industrial
0:15:52 phase, where everything's done in these
like big centralized hubs and it's not
0:15:56 distributed across everyone's tool set.
0:15:59 Yeah.
0:15:59 So just from the perspective of a
software developer, I can think of so
0:16:04 many use cases where I would love to
build a little tool for myself, but with
0:16:09 the stacks that we have available for
the last 10 years or so, there's like,
0:16:14 Just my mind is typically approaching
it the same way as like how I would
0:16:18 build a software product for a startup.
0:16:22 And that is like way overkill.
0:16:24 There's like so many layers of
complexity that by now we're just
0:16:27 used to, like there's like the hammer
that we're used to using, that it
0:16:32 sort of like kills the entire idea.
0:16:35 In the initial phase where I
don't even get to build it because
0:16:38 I already know like, Oh, I've,
but I know all the implications.
0:16:42 I need to build this.
0:16:43 I need to deploy that.
0:16:44 And like, if that thing changes, then,
this will ruin all my weekends to come.
0:16:49 So we, we have like
done so much innovation.
0:16:53 That, was always adding more
functionality, cooler stuff, but it always
0:16:58 came at the price of complexity, just
like layering on and like compounding.
0:17:04 And I think where I see local-first
enabling this age of the barefoot
0:17:09 developer is by bringing down the, the
complexity in a massive way that you can,
0:17:16 with a much simpler technology stack build
very interesting things that don't need to
0:17:22 scale to Mark Zuckerberg Facebook scale.
0:17:25 but if it works for your family and
you build it when you want it, and
0:17:30 then you don't really, it doesn't
break constantly because you depend
0:17:34 on all so many other external things.
0:17:37 I think not just for, use cases that
I can't think about, but just for, for
0:17:42 my own personal context in a, my family
stuff that I might as a Barefoot developer
0:17:48 build for my friends and extended
family, but also what other people in
0:17:52 extended parts of the world can build.
0:17:54 I think this is what's changing now
that the complexity is coming massively.
0:18:00 But there's another thing happening
now that you've also, built in as
0:18:04 sort of the second layer of your
talk, which is AI is happening now.
0:18:09 so it seems like that is also a
fundamental aspect that's also ushering
0:18:14 in a new era of end user programming.
0:18:22 AI
0:18:22 Yeah, so this was kind of, like, the
big claim I tried to make in the talk,
0:18:26 which I do believe, but it's definitely
a claim that, like, needs a lot of
0:18:29 evidence and a lot of, like, questioning.
0:18:31 But at the moment, I think I'm still
pretty firmly in this camp, where
0:18:34 I genuinely believe language models
make a significant difference to end
0:18:39 user programming being more possible.
0:18:41 when language models first started getting
popular, and, like, we realized they were
0:18:44 really good at code assistance, like,
They're questionably good at other things.
0:18:47 I'm not going to say that they generate,
like, useful articles for the web.
0:18:50 I don't believe they do.
0:18:52 But for code assistants, it's like,
anyone who uses Copilot regularly, or
0:18:56 Curso, or any of these other assistants,
I think you would have to be crazy to
0:18:59 claim that they are not, like, speeding
you up, like, helping you debug things.
0:19:04 I know a lot of senior developers
are like, well, I could just do that.
0:19:06 myself.
0:19:07 And it's like, well, yes,
you are a senior developer.
0:19:08 you understand the bugs
in the way that I do not.
0:19:11 like, I've started to joke, I'm
kind of a mediocre developer without
0:19:14 language models, but with them,
I'm, like, almost a full developer.
0:19:17 I'm, like, actually, I can pretty
much do anything you assign me to.
0:19:21 Like, in the world of front end, like, a
little bit of making, you know, back end
0:19:24 calls, but, like, it just, it just knows,
you know, I just ask what I want it to do.
0:19:29 And it pretty much gives me working code.
0:19:30 Sometimes I do have to go in and debug it
or like, ask it to do revisions, but like,
0:19:35 the amount it can do that I could not
do on my own is really just astonishing.
0:19:38 it's very much a sort of
subjective, personal qualia.
0:19:43 Everyone that I talk to, like, feels
this way, especially people who aren't
0:19:47 great programmers, like more experienced
programmers are more skeptical, like,
0:19:50 oh, this is causing more bugs, but
I don't know, if you're more on the
0:19:53 junior end, you just feel like you've
just been given a superpower and you
0:19:56 can like, make way more than you could.
0:19:59 so this big claim is that like, not,
I'm not saying all normal people, right?
0:20:03 As I said in my talk, there are a bunch
of normal people out in the world who just
0:20:07 don't, they don't care about programming
and we shouldn't force them to.
0:20:10 Like, I think there was this
moment where we were like, everyone
0:20:12 needs to learn to code, right?
0:20:14 Every nurse needs Every, like,
accountant, every teacher,
0:20:17 you all have to learn to code.
0:20:18 And they were like, well, we
don't want to learn to code, like,
0:20:20 why are you making us do this?
0:20:21 So I feel like we should just leave
people alone who don't want to code.
0:20:24 Like, let them do their other very
important jobs in the world that
0:20:26 have nothing to do with computers.
0:20:28 but there's a group, and this is
what I call the Barefoot Developer
0:20:31 Group, who really do want more
agency over their computers.
0:20:34 They're, like, currently using
Airtable and Notion and like
0:20:38 spreadsheets and like doing really
crazy, complex things with them.
0:20:41 But they also hit the
limits of these apps.
0:20:43 Like, I feel like I have a lot of friends
in this category who have a lot of
0:20:46 software needs and the software doesn't
meet their needs and they keep being like,
0:20:50 well, how do I make it like do this thing?
0:20:51 And I'm like, well, you can't, like this
other company controls that software.
0:20:54 You can't.
0:20:55 add that feature onto it.
0:20:57 but these people, if given access to
a language model, and kind of taught
0:21:00 how to prompt it in a way where they
can clearly ask for what they want,
0:21:04 it can write the code for them.
0:21:05 It can build the app for them.
0:21:07 If we, like, get the infrastructure
right, it could set up some kind of
0:21:10 local-first database, ideally, but
some kind of database, put it on a
0:21:14 URL, let, let people log in, store
data somewhere, let, you know, figure
0:21:20 out a way where we can have language
models set up, the schema for them.
0:21:23 It's like, It doesn't have to be
super complex stuff, it might just be
0:21:26 something like, Hey, help me organize
a trip for 12 people for a week.
0:21:30 Like, that involves a lot of logistics
and ops and planning, or like, you know,
0:21:33 help me plan meal prep and groceries.
0:21:35 There's a thousand apps for this, but
what if you could just prompt your
0:21:38 own, tweak every single feature that
you wanted, like make it work just the
0:21:41 way that you want the thing to work.
0:21:43 These are not complex situations,
but tons of people have these basic
0:21:46 needs and language models could
easily get them to working local
0:21:50 apps that they own and control.
0:21:52 Right.
0:21:52 That makes a lot of sense.
0:21:53 And I think this kind of puts it
nicely into two compounding steps.
0:21:58 Step number one is even for more
experienced engineers like myself, Like
0:22:03 even for me, it would be too much to
build all of like these ideas that I
0:22:08 have in my mind, because I know like,
okay, this is easy enough to think about,
0:22:13 but to build, you need to wrestle with
all of those layers of complexity that
0:22:17 will just like drown you at some point.
0:22:19 And this is where step one is local-first,
simplifies everything to the core again,
0:22:24 that it becomes feasible, but it's
feasible, but you still need to have
0:22:29 quite a bit of experience to build it.
0:22:31 And now step number two is LLMs that
just bring down the barrier to entry
0:22:36 by so much, by having someone who
knows what they want to build and
0:22:40 having a little bit of an intuition,
how computers work, et cetera, and how
0:22:44 to put those things together, who are
not afraid to get some autocomplete.
0:22:49 Sure, I have no idea what it is, but let
me try and then, like, iterate from there.
0:22:53 And this is the second part where LLMs
bring down the barrier to entry and now,
0:22:58 like, make this possible on a wider scale.
0:23:00 Did I get this correctly?
0:23:02 Yeah, yeah, exactly.
0:23:03 I think I should add on too that, I just,
I showed in my talk some demos from both
0:23:07 Vercel's V0 and TLdraw drawing interface.
0:23:11 Because as much as I was saying, you know,
of course you'll describe and text what
0:23:14 you want, You can do that, but often it's
sometimes hard, or there's an ambiguity.
0:23:18 but with TLdraw, and with, uh, B0 from
Vercel, you're drawing a visual interface.
0:23:24 Like, in TLdraw, you can,
like, draw boxes, right?
0:23:26 And you can, like, point to a button
and say, when I click this button,
0:23:28 you know, increment this counter.
0:23:30 And, like, when I click this button,
start a timer for three minutes.
0:23:32 And you can just sketch out what
you want with these instructions.
0:23:36 And then they have this make real button,
so you select what you're drawing.
0:23:40 You click make real and
it makes the app for you.
0:23:42 And like, it doesn't work a
hundred percent of the time, but
0:23:45 like 90 percent and these models
are only going to get better.
0:23:48 So we would hope if you get closer to a
hundred percent, it just builds exactly
0:23:51 what you told it to build and it works.
0:23:52 And it's like, it's on a URL.
0:23:54 It's like, it's a real app.
0:23:55 You can click it, all the things work.
0:23:57 it's hard to look at that and still
be skeptical that we're not at some
0:24:01 step change in end user programming.
0:24:04 Definitely.
0:24:04 and I think this is like this going
beyond the medium of auto completing
0:24:09 code, but that you have like the modality
of like a visual modality where you
0:24:14 can draw a napkin sketch and from that
napkin sketch, you have like some working
0:24:20 React code that you now like I feel
like the next step in the evolution of
0:24:25 putting together all of those systems.
0:24:27 It would be something like TLdraw,
make it real, but for your local-first
0:24:33 data model that you basically describe
like your, your use cases, or you
0:24:37 could even use your, your visual
sketches that you've made through
0:24:41 TLdraw, feed it into an LLM and say
like, Hey, can you suggest to me Data
0:24:47 stuff that I'll probably need for this.
0:24:49 And it might spit out some code
or it might spit out like some,
0:24:53 some other artifacts that you can
now use to iterate from there.
0:24:56 So I think that multimodality
just brings down the barrier to
0:24:59 entry even more, as well as making
experienced people more productive.
0:25:04 So now it really clicks how the, that
answers the question of like, why now?
0:25:09 Since I think the idea was
compelling before as well.
0:25:13 But with all of that complexity and, with
all of the learning curve that needs to
0:25:18 be done before, that wasn't feasible.
0:25:21 So, uh, I love that.
0:25:23 Like I had also cited in my essay, uh,
an essay written by, well, I guess more
0:25:27 of a short post written by Clay Shirky
in I don't want to mess up the year.
0:25:32 Was it 20 2004 or 2014?
0:25:34 I want to bet 2014.
0:25:35 called Situated Software, which
was about which was the same
0:25:38 claim as Barefoot Developers.
0:25:39 Like, software should be situated
in the local environment that where,
0:25:42 like, it's the people that it serves,
it should have been made there.
0:25:45 He made this idea at a time where
it was, like, completely infeasible.
0:25:49 Just, like, No one could build
sophisticated software, which
0:25:52 didn't have the tooling, like
we hadn't gotten this far.
0:25:56 And, like, what would be the interface
for normal people to build this program?
0:25:59 It was one of those, like, wouldn't
this be nice kind of concepts.
0:26:03 And now it does feel like, yeah, we've
reached this point where it's like, wow.
0:26:07 The coming together of like better
tooling, like reducing complexity,
0:26:10 like better abstractions, plus
language models trained on all those
0:26:14 abstractions and all that tooling.
0:26:15 It is like, Oh, this is possible.
0:26:18 Yeah, that, that makes so much sense.
0:26:20 I can't wait to see like how
quickly that's going to unfold.
0:26:25 This was also a big part of what we
try to do with the conference is, not
0:26:30 just get the, people who are already
excited about local-first from, and who
0:26:35 are trying to figure out how to build
it in the best way, but also people
0:26:39 who've maybe heard about local-first
before, or haven't heard about it before,
0:26:44 and that they get to understand what
could be the possible impact of that.
0:26:49 And I think this is where you've
just knocked it out of the park
0:26:53 with, at the end of the day, with
your talk, where I think before.
0:26:57 Your talk, everyone was already
excited about like, Oh, this is going
0:27:01 to make software development so much
easier, brings down the level of
0:27:05 entry for people to build local-first.
0:27:08 And it also enables next generation
user experience, like Linear, et cetera.
0:27:14 And that's also why I initially got
excited about local-first to build sort
0:27:19 of like a really cutting edge, next
generational products with Overtone.
0:27:23 But I think you really expanded
the horizon even further, what
0:27:28 on a kind of societal level,
local-first could have an impact for.
0:27:33 So, do you have some theory of
like, what are some, like milestones
0:27:38 that you would like to see happen
there as sort of success stories?
0:27:42 Yeah, um, I talked about a bit
at the end of my talk that we're
0:27:45 missing a few bits of like tooling
or infrastructure, I think, to make
0:27:48 this work, because at the moment,
what you can do easily with language
0:27:51 models is prompt front end code, right?
0:27:54 It's mostly what people
are doing, like even a 4.
0:27:57 3.
0:27:57 5 Sonic came out, I think,
a week ago, and it's really,
0:28:01 really good at code generation.
0:28:02 People are making crazy stuff, and I
have this whole fat collection on Twitter
0:28:06 every time someone makes something.
0:28:07 I'm just, like, saving all this stuff as
evidence of, like, here's where we're at.
0:28:11 Here's the current state of things.
0:28:12 But it's just making
front end code, right?
0:28:14 It's giving you React
components, interactive, it
0:28:17 looks cool, it looks capable.
0:28:18 But it's not connected to a back end.
0:28:20 Connecting it to a back end would
be like a whole different thing.
0:28:23 Um, user authentication and logging
in would be a whole different thing.
0:28:27 Like security, like
multiplayer collaboration.
0:28:30 Like, there's just, there's all this other
stuff to have a real working product.
0:28:34 Like, deployed to a real domain.
0:28:36 Imagine if you're not a software
developer, like, you would never figure
0:28:40 out how to glue that stuff together.
0:28:41 No matter how much you have, like,
chat GPT helping you and, like, you
0:28:44 asking it questions and trying to put
together a working piece of software.
0:28:48 You wouldn't even know
those things exist, right?
0:28:49 Like, no one knows what a database
is if you don't work in software.
0:28:53 so we're missing, all these
other pieces of tooling.
0:28:56 I think what I'm waiting to see is,
like, I don't know, you know, I've kind
0:29:00 of tried to put this idea out into the
world, tried to make more people aware
0:29:03 of like, hey, we should be trying to
build things to serve these developers.
0:29:06 and I know of two projects, Jazz and
DXOS, who are both at local-first, Jess
0:29:11 and, um, Anselm are hopefully building
this kind of thing, which is like some
0:29:16 sort of full stack local-first, like,
framework for people to build within.
0:29:21 I think if you could train a model on
those kinds of things, where it's like,
0:29:25 You're talking to a language model,
but its whole goal is to get you to
0:29:28 build a full stack local-first app.
0:29:30 That would be the goal for me.
0:29:32 That's like, oh, I want to come in,
I want to be able to sketch, I want
0:29:35 to be able to prompt things in it.
0:29:36 and I want it to sort of be architected
and trained in a way where it's
0:29:40 acting like a product manager, like
helping me move towards my final goal.
0:29:44 Like, you know, it's It gets me
to tell it the requirements and
0:29:47 it gets me to tell it what I want.
0:29:49 And then it's doing, it's making
technical decisions for me, right?
0:29:52 It's like, hey, I'm just going to
figure out how to model this data.
0:29:55 I'm going to figure out how
to make this front end code.
0:29:57 All you have to worry about is looking
at the end result and telling me,
0:30:00 yes, no, you know, what should I edit?
0:30:02 What needs to change?
0:30:03 And just looping.
0:30:04 Like that, that's, that's what I'm
waiting for someone to build, honestly.
0:30:08 I agree.
0:30:08 I think there's sort of the, this gap
in going from A to B where like, A,
0:30:14 where we're right now, where I think
we have the conceptual right ideas,
0:30:20 how things can be simpler, and a few
pioneers such as the folk from linear.
0:30:27 And a few other products who have
already had those right ideas, but
0:30:31 they're technically very savvy so
that they could like build that new
0:30:36 hammer, that new framework themselves
so they can live in the future.
0:30:41 And, then you have the, the other
group, what you've described with like
0:30:45 DxOS, Jazz, and there's many other
technologies who are not primarily
0:30:50 building it for their own apps.
0:30:51 But to kind of build this for
developers, so that the next Linear
0:30:56 has a much easier time, can focus
on the product and doesn't need to
0:30:59 build off like the underlying tech.
0:31:01 So that's coming along, and once
that is there, then we're kind of
0:31:06 halfway there, since at that point,
now the AI also needs to pick this up.
0:31:11 Since, I think this is why the
AIs are so great at bringing down
0:31:16 the barrier to entry for things
like React development, et cetera.
0:31:20 There's already a lot of training data
out there, whereas for local-first,
0:31:25 this is all like very little training
data, all like very, very new.
0:31:30 So I think once we have those ingredients
and the AIs can also integrate that into
0:31:36 their more encompassing, solution corpus,
then I think we've, I guess like to, to
0:31:44 a much more comprehensive degree enable
a Barefoot developer, since I guess
0:31:48 right now a Barefoot developer already
lives in the future by, things like
0:31:54 front end React programming, et cetera.
0:31:56 That already works super well for them.
0:31:59 And then they get sort of this
glimmer of hope, Oh, a local-first
0:32:03 is going to make everything simpler.
0:32:05 But the, the fact of reality today
is you either need to use a very
0:32:10 Early technologies such as jazz, et
cetera, and where you still need to
0:32:14 be a bit more on the more expert y
end, or you need to wait a little
0:32:18 bit until all of that is AI ready.
0:32:22 But I, I like the way how you
think about that sort of like as
0:32:25 a big next step function change.
0:32:28 Yeah, I expect they
might meet in the middle.
0:32:30 Like what I think will happen is, like,
people will build these kind of like
0:32:33 full stack language model helpers, which
are not local-first, of course, right?
0:32:37 They will just build one that like
uses cloud databases entirely.
0:32:40 And it's like, all like
not, doesn't work offline.
0:32:43 and then the local-first tools will
get more mature and maybe in a few
0:32:46 years, like these come together
in the middle so that you have.
0:32:49 You know, something for barefoot
developers that just has natural
0:32:52 language inputs and visual inputs
for them to define what they want.
0:32:56 and the stack is, like,
mature enough, right?
0:32:59 We have, like, stable enough technologies
that they get all the stuff around
0:33:02 local-first, like, CRDTs, all the
fancy stuff for free, just because we,
0:33:06 like, have enough training data and
we figured out the complexities of it.
0:33:10 I don't want to put too many timelines
on it, but I'm like, five years?
0:33:14 Maybe less.
0:33:14 I don't know.
0:33:15 Some people think that we're all
going to die from AGI in five years.
0:33:17 So maybe just speed up timelines, like,
but I think reasonably five years.
0:33:24 , I think a very frequent philosophical
question that one can entertain
0:33:28 is like, my life as a programmer
has certainly become a lot better
0:33:34 over the past couple of months and
years with things like Copilot and
0:33:38 Cursor and, ChatGPT, et cetera.
0:33:41 I can like take a screenshot of an error
message and it gives me some useful clues.
0:33:47 so that is.
0:33:48 Definitely getting better, but is this
just like a very nice time before, I
0:33:54 don't have to do my, anyway, this is
not a question to debate here, but
0:33:58 it's certainly also, I'm not immune
to, to those sort of thoughts as
0:34:01 well, but I'm, I'm choosing to see
the world as like glass half full.
0:34:06 and, I'm a tool builder, so I'm excited
for new tools that make me more productive
0:34:11 and, make me more capable in doing stuff.
0:34:15 Yeah.
0:34:16 I mean, I was gonna say, if we,
if we automate all the code,
0:34:19 we'll all just become designers
and product managers, you know?
0:34:21 Like, we'll just, we'll just help.
0:34:24 Help the AIs, like, write the
code and we'll just be like,
0:34:26 Hey, you know, focus on this.
0:34:28 This is your priority.
0:34:29 You know, this usability
piece isn't working very well.
0:34:31 I still think it would scratch the
same itch that a lot of developers have
0:34:34 around like building beautiful systems
and like solving like logical puzzles.
0:34:38 Cause it's still design.
0:34:39 You're solving a lot of logical puzzles.
0:34:41 I don't know.
0:34:42 So maybe that's why I'm not
as afraid of like the whole.
0:34:44 Automation fears, but I also
think they're ridiculous.
0:34:46 Like we'll never get to a hundred percent
code automation and it seems crazy.
0:34:51 Elicit
0:34:51 Yeah, definitely.
0:34:52 So I want to shift gears slightly,
and not just look at AI as an enabling
0:34:59 technology to make us more productive.
0:35:01 But if we put ourselves into
the shoes of a product builder,
0:35:05 possibly of a tool builder, how.
0:35:08 Can AI be a ingredient that we can put
into possibly local-first powered apps?
0:35:16 and so given that you've worked at
Elicit, which is maybe you can briefly
0:35:20 give a summary what Elicit is about and
what you've done there, but after that,
0:35:25 maybe we could explore a little bit of.
0:35:28 maybe you have thoughts on how to best
bring AI into products and how is that
0:35:34 a challenge or an even better benefit
to do it in a local-first context?
0:35:39 yeah, I actually, this is funny, like,
after I had written the talk and was
0:35:42 talking to more people, I had that
realization of, like, oh, language models
0:35:46 and local-first go really well together.
0:35:48 I think this is actually Jeffrey Litt,
who's at Ink and Switch, that kind of
0:35:51 got me onto realizing this, how they're,
like, actually perfect for each other.
0:35:55 but I'll back up a bit
and talk about Elicit.
0:35:57 So, yeah, I was the designer there
for two years, It is a tool that helps
0:36:01 scientists and researchers do this thing
called systematic literature review,
0:36:05 which is this frankly very long manual
boring process that no human wants
0:36:08 to be doing, but they're sort of like
forced to do because they have to.
0:36:12 where let's say they want to
do a study on like, vitamin D
0:36:14 deficiency or something like this.
0:36:16 They have to read every single
paper that's ever been done on
0:36:19 like vitamin D deficiency studies.
0:36:21 And like, what do we know?
0:36:22 What's already been tested?
0:36:23 What needs to be tested in the future?
0:36:26 Just like read tens of thousands of
papers and extract all this data from it.
0:36:29 Like how many people did
they study in each paper?
0:36:31 And like, what were the results?
0:36:33 , what was the effect
size in the population?
0:36:35 Like just all this stuff.
0:36:36 And usually humans, Open up PDFs one by
one and type this data into a spreadsheet.
0:36:41 They're just like sitting there
for, it takes six months usually,
0:36:44 it takes a couple people six
months to do one of these.
0:36:46 And you have to do it for every
like new drug release to market
0:36:49 or any new medical device.
0:36:51 So there's this whole consultancy whose
whole job is to do manual data entry.
0:36:55 and it turns out language models
are quite good at extracting
0:36:58 structured data from these PDFs.
0:36:59 Or if you prompt them in the right way
and you have Frankly, at times, quite
0:37:03 sophisticated architectures to make them
accurate, but you can build them, which
0:37:07 is what we focused on at Elicit, is like,
raising accuracy rates and making sure
0:37:10 models got the right data out of papers.
0:37:13 But you can get them to do it to over 95
percent accuracy, which is what humans
0:37:16 roughly get out of doing this process.
0:37:19 So you just speed up this thing, used
to take six months, you can do it in
0:37:22 a couple days, maybe a week, of you
just going through and reviewing all
0:37:25 the answers that the models extracted.
0:37:27 So, yeah, working there for two
years, I learned a lot about, both
0:37:31 just like prompt techniques and
architectures for making models accurate.
0:37:34 It's actually quite hard, but,
but good to know it can be done.
0:37:37 That was like really useful to learn.
0:37:39 it definitely made me very skeptical
of most use cases of language
0:37:43 models, I think, in that we found
one that was really appropriate.
0:37:46 It's like, okay, you There's
data and a bunch of text.
0:37:49 Get the data out of the text.
0:37:50 That's like quite straightforward.
0:37:51 But there's a lot of stuff people
are trying to use them for where
0:37:54 I'm like, is a language model
the right hammer for this nail?
0:37:57 Like, I don't know.
0:37:58 You're trying to write
marketing copy for your company.
0:38:01 I don't know that you want the
most generic, predictable text to
0:38:04 really be the thing that you're
like putting out into the world.
0:38:06 Or like, okay, some people are like,
yeah, I want to write 10, 000 SEO keyword
0:38:11 stuffed articles to like get to the
top of Google, but and everyone does
0:38:14 that and then there's no point in like
you, you doing that because you're not
0:38:16 actually going to get any traffic from it.
0:38:18 So, I'm strangely quite skeptical
or like negative on a lot of uses
0:38:22 of language models, but the data
extraction, like summarization,
0:38:26 like getting it to do very small
structured things over text, I actually
0:38:30 think they're really good at that.
0:38:31 And code generation and debugging
is actually one of those, right?
0:38:35 Like, that's one of the things
where like, okay, we've proven
0:38:37 they're actually useful for this, so
like, we should use them for this.
0:38:40 This is like, they are shining here and
like, ignore the areas where they're
0:38:43 not shining, like all the other, like,
everything turns into a chatbot kind
0:38:47 of stuff, which we don't really need.
0:38:49 yeah, so that's, that's the long
version of like, that was, that
0:38:51 was what I realized at Elicit.
0:38:53 but the thing with local-first
and language models that makes
0:38:56 them go really well together.
0:38:58 Language models perform better when
you have lots of really specific data
0:39:01 for them to run things over, right,
to summarize or to extract data from.
0:39:05 And if all the data is local on
your machine already, and you're
0:39:09 not having to, like, go up to the
cloud, or it's not trapped in someone
0:39:11 else's app to run a model over it,
You can just run a model locally.
0:39:15 You don't need to be making a
call to OpenAI or to Anthropic,
0:39:18 which gets really expensive.
0:39:20 You could just be doing this all
on your machine with no delay,
0:39:22 which is like, why are local-first
makes perfect sense if you want to
0:39:25 actually be doing language model work.
0:39:27 Yeah.
0:39:28 So thank you so much for, for
the background, on, Elicit.
0:39:30 I was not as aware of like the
entire, process that researchers
0:39:35 need to go through and just
hearing the word manual data entry.
0:39:39 it gives me shivers.
0:39:40 So that sounds like a well
needed thing and whatever , makes
0:39:45 research and truth easier.
0:39:47 that is very welcome.
0:39:49 also your insights in regards to where
LMs are a good fit versus not to.
0:39:54 I've coming to on a much smaller
scale, but I'm coming to similar
0:39:58 conclusions, for example, when I'm
writing documentations for, for some
0:40:03 technologies that I'm working on.
0:40:04 And this is where I typically disable LLMs
because like what they typically suggest
0:40:10 to me is exactly not what I'm doing.
0:40:13 I'm building a different thing
that challenges the status quo
0:40:16 for very particular reasons.
0:40:18 So I found it actually, it increases my
cognitive overhead of like trying to.
0:40:24 like come up with a thought
and like a very fuzzy process.
0:40:27 And then like, it's basically
constantly like slapping me in the face.
0:40:31 Like, do you want to say this?
0:40:32 No, exactly not.
0:40:34 So, there's like use cases for it,
where it's good and where it's not good.
0:40:38 but, I think also that giving it more
contextual information is exactly
0:40:43 the, the key to make it feasible
and maybe make it even so that I do
0:40:48 want to have it enabled for docs.
0:40:49 So if I imagine that I give it
access to all of my source code.
0:40:53 then it could actually draw,
draw the right conclusions.
0:40:57 So , this use case where you have
your source code and you want to
0:41:01 generate documentation, I think is
a much simpler one from a technical
0:41:06 perspective because everything is static.
0:41:08 I, don't need , my source code,
sure, maybe some of it, I don't
0:41:12 want to have leaked, but a lot
of it's going to be public.
0:41:15 So privacy concerns are
not as much of a concern.
0:41:18 But where that is the context, if
we're now talking about the health
0:41:22 app, this is where my context
is possibly very, very sensitive
0:41:27 data or financial data, et cetera.
0:41:30 So this is where the contextual
data is even more important to
0:41:34 drive good results for the AI,
but it's so much more sensitive.
0:41:39 And I think this is going to be a huge
challenge and a huge question, how
0:41:44 different AI systems will be built.
0:41:46 And that's a key.
0:41:47 topic that I'm interested in,
this is where I think where
0:41:50 we've seen with Apple's recent AI
announcements, where they hijack the
0:41:55 term AI to mean Apple intelligence.
0:41:57 Well done.
0:42:00 And now what Apple has announced
at the latest WWDC is where they
0:42:06 basically shared their AI strategy.
0:42:09 and what I like about it is that
most of the AI stuff is also
0:42:13 happening locally on device.
0:42:15 And I think the, the way how
to do that is, an approach that
0:42:20 hopefully more technologies follow.
0:42:22 And I think this will be sort of like
a bit of a dividing line, whether
0:42:26 you're gonna do it locally or whether
you're going to do it in a cloud.
0:42:29 And I hope that more products,
more technologies will enable
0:42:35 all of that to happen locally.
0:42:37 Since if I can keep my most sensitive
data locally on the device, and if I
0:42:42 can run the AI model locally, it's not
just gonna give me the fastest results,
0:42:48 but it's also, in my perspective,
kind of like the only sane way
0:42:52 how to respect the user's privacy.
0:42:55 It was more of a coin flip to go
with Google versus Apple before
0:43:00 in terms of products, et cetera.
0:43:02 But if Google's only way in the future
is to drive everything into the cloud
0:43:07 and do AI there, and Apple's is to do it
locally, that's a very clear call for me.
0:43:13 But I think that's a
major question overall.
0:43:20 Apple Intelligence
0:43:20 Yeah, this is something that I haven't
seen yet come up as like a SaaS or
0:43:24 a service of any kind, or I guess
it'd be more of developer tooling.
0:43:27 Like, there's plenty of, wherever it
is, like, toolchains on top of the, all
0:43:30 the foundation model APIs, and like,
you know, analytics and, visibility
0:43:36 and, like, tracking and all this
stuff, like, coming up around, around
0:43:39 doing language models and product.
0:43:40 But I'm waiting for someone to just be
like, hey, We help you run LLAMA locally
0:43:44 on the user's device and then only when
needed, like, when, like, as Apple said
0:43:48 in their presentation, only when you need
something that's more compute intensive
0:43:51 do they send it up to a cloud server.
0:43:53 Something that allows you to do hybrid,
local, and uh, remote, so that at least
0:43:59 ideally the user, I mean, this is probably
more likely to happen in a B2B case, could
0:44:02 say only do these kind of like, functions
on this kind of data, uh, locally, maybe
0:44:07 this is more sensitive data, maybe this
is more sensitive stuff, but stuff that's
0:44:10 maybe more complex or stuff over here,
you can send to OpenAI if needed, or to
0:44:15 a cloud, I mean, to Anthropic if needed.
0:44:17 So I haven't seen that happen yet.
0:44:19 I don't know what that is.
0:44:19 That's like hybrid language
model prompting, but that seems
0:44:22 like a very obvious piece of
infrastructure we could build in.
0:44:25 I'm wondering, like, what will
be the thing that wins the
0:44:28 argument one way or another?
0:44:30 one possible path could be, Oh, it gives
you the best performance or that it needs
0:44:35 to work offline in certain scenarios,
whereas like, I think with local-first,
0:44:41 we already see that the, the works offline
case is kind of like a, Argument that
0:44:47 is not as compelling for many or not as
intuitive for many as you might think.
0:44:53 but I think what I would hope will
be the winning argument is privacy.
0:44:59 But I'm saying that as a European, and
I know that other people might, uh, not
0:45:04 value privacy as much, but do you have
thoughts on what will be like a big
0:45:09 lever in terms of one way or another?
0:45:12 Yeah, I mean, I also agree on the
privacy bit, but I also know a lot of
0:45:15 people who just, I don't know, doesn't,
don't, they don't care, or it's just
0:45:18 like, oh, this is just whatever.
0:45:19 This is like my, you know, notes
on my product development app.
0:45:22 Why do I care, about privacy?
0:45:24 But I think it's speed.
0:45:25 Like, one of the biggest UX
challenges of the list that we faced
0:45:28 is just that loading times are Or
just on a whole different level.
0:45:31 It's just, it's, we're so used
to quick, quick loads, right?
0:45:35 At the most one, two seconds.
0:45:36 And even then users are getting, you
know, antsy and like, oh, this feels slow.
0:45:40 This app isn't working.
0:45:41 Sometimes you have a load
time with a language model.
0:45:43 If it's doing a complex call of
like three minutes, like, What do
0:45:47 you do with a user for 3 minutes?
0:45:49 Like, there's only so many animations
you can play on the screen.
0:45:52 Like, there's only so many
minigames you can have them
0:45:54 play while this thing loads in.
0:45:56 so having models locally on your machine
that could run language model functions
0:46:00 faster would be a huge UX benefit.
0:46:03 A lot of the time we're just
waiting for the OpenAI API to
0:46:05 reply, or like, Maybe we're doing
multiple, you know, a hundred calls.
0:46:09 So every single time there's a couple
seconds of delay on each one, this is just
0:46:13 adding up into a huge, huge time sink.
0:46:15 so that would be a big benefit, but I
of course agree on the privacy front,
0:46:19 especially for companies will care
about this, right, for proprietary data,
0:46:23 but yeah, individual uses for finance
to health and anything where you just
0:46:25 don't want it going off your machine.
0:46:28 Yeah, so this is a topic I'm
highly interested in, want to
0:46:31 explore also in future episodes.
0:46:34 I think luckily there's so much
momentum and so much good stuff
0:46:37 happening on the AI capability side
where models are getting more powerful,
0:46:43 but also more efficient that makes
it possible to run them locally.
0:46:46 I think Google announced or released,
uh, Recently, like a Google, uh,
0:46:51 Google Chrome version, I think a
Canary version where you have, I think,
0:46:56 Gemini Nano working actually already
locally where it can say window.
0:47:01 ai and run some little prompts.
0:47:04 So this shows some first,
building blocks in that direction.
0:47:09 But, I'm more interested in
exploring and figuring out a path.
0:47:14 How can we.
0:47:16 Figure out the contextual data side,
since I think this is where I'm less
0:47:21 concerned about a AI model that runs
locally, temporarily having, some
0:47:28 contextual information about me that it
has like flushed again, like in a few
0:47:32 seconds, what I'm more interested in is
exploring all of that contextual data
0:47:38 that is highly personalized, that might
be text messages I'm sending to my close
0:47:43 friends or to my family, that might be
health information, that might be my
0:47:47 emails, that might be my personal notes
I'm taking after a nightmare, or whatever.
0:47:53 Like, all of that stuff needs to be
accessible to the AI in a local way.
0:47:59 And that is much easier to do if,
some entity, for example, Apple
0:48:04 controls the overall ecosystem
and can do it in a way that is
0:48:08 secure, privacy respecting and fast.
0:48:12 But where we are all building
things for, for the web, this is
0:48:16 where our playing field is much
more constrained to the browser.
0:48:19 So one possible path That I could
imagine where this is going is
0:48:24 that the browser takes even more
responsibilities of an operating system
0:48:29 and where a lot of my contextual data
lives in the browser, which raises
0:48:35 the stakes and the risks even further.
0:48:37 That now like some cross site scripting
could access even more personal data.
0:48:42 Since right now, with the stuff
that we have really saved in the
0:48:46 browser is like close to nothing.
0:48:49 This is actually one of the problems
we tried to solve with local-first
0:48:53 that Right now we treat the web
browser like it's not capable to do
0:48:57 anything, but we need to flip that.
0:48:58 We need to store a lot in it to
make that happen, which brings
0:49:02 up a lot of privacy questions.
0:49:03 So lots of interesting
questions to explore there.
0:49:08 Well, okay, wait, I have a good,
a good dumb question for you as
0:49:11 like a non backend person, like
a non, like, databasey person.
0:49:14 Could you just make either an Electron
app or web app that just accesses,
0:49:18 like, a bunch of local files?
0:49:19 Like, let's say, like, you've
got 10, 000 local markdown files
0:49:22 that, like, your personal notes,
and a lot of them very sensitive.
0:49:24 Maybe you're, like, journaling or
nightmares or whatever's in there.
0:49:27 Is there a way for some web app to
just access those without it being uh,
0:49:31 at risk of like cross site scripting
or some other attack like this.
0:49:35 So, I mean, there's always risk and
there's always like the figuring out the
0:49:40 right balance between making something
possible, making something convenient and
0:49:44 keeping it secure and privacy respecting.
0:49:47 So I think one capability that browsers
already offer that could enable
0:49:53 that would be Chrome extensions.
0:49:55 But this is also, I think in a
parallel universe, Chrome extensions
0:49:59 would be even more common.
0:50:01 and we'd use them maybe even more,
maybe in a parallel universe,
0:50:06 Electron wouldn't have taken off as
much as it did and Chrome extensions
0:50:10 instead would have been that medium.
0:50:13 So, Chrome extensions could, or
browser extensions, generally,
0:50:16 could be a bridge towards local
data, and I think that will be a, a
0:50:21 mechanism that might be used for that.
0:50:24 I'm actually building a Chrome
extension over the last couple
0:50:28 of weeks for LiveStore, the
staff tool that I'm building.
0:50:31 And it's, really complex to build.
0:50:34 And one of the big reasons why it's
complex to build is because the
0:50:37 browsers are already very intentional
about trying to minimize the,
0:50:43 the sort of like security risks.
0:50:45 And I think this will
be even more important.
0:50:48 And I think we need to, strike like
an interest, find a nice balance
0:50:52 where it's secure, but still
gives you the capabilities and in
0:50:58 a way that a end user is able to
make the right decision for them.
0:51:03 And so I think, yeah, that's going
to be interesting balance to strike.
0:51:08 Okay.
0:51:08 Okay.
0:51:09 Okay.
0:51:09 That gives me more hope.
0:51:10 Cause yeah, I'm in the phase
of experimenting with, um, you
0:51:13 know, Obsidian, local-first,
like note taking tools.
0:51:16 Yeah.
0:51:16 Right.
0:51:16 So like all my All my notes are in there.
0:51:19 I was in something else before
that was all cloud based and
0:51:21 actually going to local-first.
0:51:22 I was already on the edge of being like,
Oh, personal notes should really be, Oh,
0:51:25 you should own that on your own machine.
0:51:27 That should be marked down.
0:51:28 That shouldn't be stored in
someone else's cloud server.
0:51:30 so went through the process
of moving them all into this
0:51:32 like local-first Obsidian app.
0:51:34 and then now exploring building
plugins to like build on top of that,
0:51:37 like integrate language models, like
do more interesting things with this
0:51:40 like huge database of notes I have.
0:51:42 but it made me realize all I needed
was some sort of like, Interface or
0:51:45 runtime or primitives in place, which
is what the browser gives you, or what
0:51:48 something like Obsidian gives you.
0:51:50 And then people being able to
build things on top of that.
0:51:52 That's when you get into
the more interesting stuff.
0:51:54 But making sure that base layer of like
somewhere to store data, some security
0:51:58 and making sure it's local-first.
0:52:00 Like once that's in place, people
can build much more capable things.
0:52:03 Local-first in the tools for thought space
0:52:03 So maybe this is a good time to
shift gears again a little bit.
0:52:07 you've been mentioning Obsidian
and in the past, and I think still
0:52:11 today, you're a very active person
in the tools for thought space and
0:52:16 Obsidian, obviously being one of
the prominent tools in that space.
0:52:20 Do you have thoughts on how
local-first could be an interesting
0:52:25 lever to make those experiences even
better, both in terms of like end
0:52:29 user apps that you're using, but
also in terms of you as a builder?
0:52:34 Yeah, yeah, I mean, I kind of had this
like, I've been so dumb moment, I guess,
0:52:38 with the, with the tools for thought
note taking thing, or realization, I
0:52:41 should say, like around, around the time
of the local-first conference, where
0:52:44 for years I had been using, I was like
early to roam research, and I moved to
0:52:48 this app called Tana, which is like,
Gorgeously designed, like, totally acts
0:52:51 like a really flexible database, and they
have a really beautiful outliner, things
0:52:55 where everything has types and schemas,
and it's just the sort of thing I love.
0:52:58 Like, types and schemas
everywhere, super structured data.
0:53:01 and their product is great,
but it's not local-first.
0:53:03 I don't think it's ever going to be.
0:53:04 It's nowhere near their roadmap.
0:53:05 And I just had this moment of realizing,
like, After, after going to the conference
0:53:09 and like hearing all the arguments
against it, I mean, for it and against it.
0:53:14 And I was like, Oh, I, even if there's a
UX cost to switch into something that's
0:53:18 local-first, I like absolutely have to.
0:53:19 This is like crazy that I've been using
something that doesn't work on airplanes.
0:53:23 That is my note taking database.
0:53:24 It's just like nuts.
0:53:25 Doesn't work on that too, but it's wild.
0:53:27 Well, you just don't think about, you
just don't think during airplanes.
0:53:32 No, no, I know, I know, because you
kind of go, oh that's just, you know,
0:53:37 all software doesn't work offline, of
course, this is just the way of the
0:53:40 world, like I just have to accept this
reality, it was like my mindset before,
0:53:45 and then had this switch of like, oh
no, like that's totally unacceptable,
0:53:48 I absolutely cannot use, like, Software
that doesn't work online anymore.
0:53:52 That's just like nuts.
0:53:53 even Figma is sometimes on the edge of it.
0:53:55 It doesn't always work
like perfectly offline.
0:53:57 If you're trying to finish like a huge
design presentation on the flight on the
0:54:01 way over to your job, it's like, this
needs to, this needs to work, you guys.
0:54:04 You can't have this not
work when I go offline.
0:54:07 so anyway, I made the
big switch to Obsidian.
0:54:09 It was like a fair amount of work.
0:54:11 Something only someone who
like really cares a lot about.
0:54:14 notes and structured notes and like
knowledge bases, I guess would enjoy
0:54:17 doing it, but it was worth it for me.
0:54:19 and it opened up my mind to what
was possible now that I had my
0:54:23 whole giant notes database local.
0:54:26 And now I could, and as I said,
they have this plugin ecosystem
0:54:29 where anyone can build plugins.
0:54:30 It's all in JavaScript on top of it.
0:54:32 And I was like, oh, wow,
I can extend this now.
0:54:35 This is my environment.
0:54:36 Like, This is not just my data.
0:54:38 I now have control over the
functionality of this thing.
0:54:40 This is totally malleable software.
0:54:42 This is end user programming.
0:54:43 Like, I'm going to build my own
spaced repetition language model
0:54:47 like crazy, Zettelkasten reviewer
system on top of this thing.
0:54:50 Um, but no one else, I don't,
well, other people might want it.
0:54:54 We'll see if I get them to work properly.
0:54:55 But, um, It just, it just was like,
oh, so much more is possible in
0:55:00 this really exciting, empowering
way where I went, oh, I own this.
0:55:05 Like I fully have control of this in a
way that I didn't with the previous tools.
0:55:09 I was always checking the change log.
0:55:11 What have they just released?
0:55:12 Like, are they going to
prioritize my feature?
0:55:15 Are they going to make the changes I want?
0:55:16 I would often write crazy CSS
stuff to hack the interface to
0:55:20 make it the way I wanted, right?
0:55:21 Just removing elements, moving
elements, changing the, the size and
0:55:25 structure of things for my preferences.
0:55:27 And again, not, most people do
not do this, but I am someone
0:55:31 that wants it to work and feel the
way I want it to work and feel.
0:55:34 so it was very liberating, and that
made me realize like, oh, local-first
0:55:36 software is actually much more
about the user having more agency
0:55:40 and control over the software.
0:55:41 They can like write their own hacks
around it in a way that's just not
0:55:44 possible with cloud based apps.
0:55:46 Yeah.
0:55:47 And I think this is another way to think
about the, the affordances that enable
0:55:52 the Barefoot developer movement, is
that also for the experienced folks who
0:55:59 build, much more sophisticated products,
they have now much more capacity that
0:56:06 they don't need to put into running
massive Kubernetes backend systems.
0:56:12 if they want to go a little bit
more ambitious about the products
0:56:15 they're building, they could now
put those resources, that capacity
0:56:19 into facilitating more end user
programming inside of that software.
0:56:24 I think unfortunately, it's like more the
exception, than the norm that software
0:56:29 facilitates that, because it is quite
difficult to find the, what is the, the
0:56:34 right interface, the right modalities,
If there's a plugin system or those
0:56:39 sorts of things, you need to be very
intentionable of like, how far do we go?
0:56:43 How flexible do we want to make this?
0:56:46 It was like, if you want to go even
more flexibility, you might as well
0:56:50 just like throw up your hands and
say, okay, you built this as a web
0:56:53 app and we only give you an SDK.
0:56:55 and so I think there's an
interesting spectrum there.
0:56:58 And, I think.
0:56:59 Building things local-first
can free up capacity that can
0:57:03 be reinvested into making more
software end user programmable.
0:57:08 So I love those, um, those use cases
and like your recent success stories
0:57:13 that you found with, sort of the
local-first mindset, bringing that
0:57:16 into the tools for thought space.
0:57:20 Is there, any other thing in regards to
local-first where you wish or you hope
0:57:26 that local-first enables a certain thing.
0:57:29 what is like on your local-first wishlist?
0:57:32 Maggie's local-first wishlist
0:57:32 Ooh, I mean, I definitely
want the quick spin up an app.
0:57:36 I mean, I know this is like, we mentioned
like things like Jazz and DxOS are trying
0:57:39 to do this, just develop a tooling that
makes building a local-first app easy.
0:57:43 Does not exist.
0:57:44 I tried.
0:57:44 I tried to build a local-first
app a couple of times.
0:57:47 And it was just that the, the developer
experience is not quite there yet.
0:57:52 So I think there's a lot to do
around DX and UX of just people being
0:57:55 able to build these kind of apps.
0:57:56 in terms of things they enable, I
don't know if they have that many more.
0:57:59 I mean, I certainly have specific examples
where I'm like, there's lots of personal
0:58:04 things that I would want to build my own
little apps for and I just want easy ways
0:58:08 to do that and that's like I mentioned
things like, I always want to do trips
0:58:11 with people like, you know, like groups of
like a dozen or even more but it's so hard
0:58:15 to just do all these small things in the
middle that like software so far doesn't
0:58:19 enable in these weird ways like Find out
when everyone's free during the same week
0:58:22 in an entire year and then figure out
all the flights and all the like finances
0:58:26 and the cooking and find like an airbnb
somewhere that would fit everyone You It's
0:58:30 just like a big ops and logistics problem.
0:58:32 And there's tons of these examples
in everyone's personal lives.
0:58:35 it's the kind of thing that like,
if it was local-first, who knows
0:58:37 if you're going to have like phone
signal out in the middle of wherever
0:58:40 you're going, and you're still going
to want to have access to all your
0:58:42 like timetables and finance sheets
and who's cooking dinner that night.
0:58:46 that's just the kind of smaller examples
that I think I'd, I'd like pull on is
0:58:49 there's always those for personal stuff.
0:58:52 So let me say what, once we are reaching
that point where it is super easy for
0:58:58 front end developers to build local-first,
home cooked software, I think your close
0:59:04 friends can be very happy to and very
much looking forward to be traded to
0:59:09 some locally cooked software from you.
0:59:12 Yeah, I will host a local-first
software retreat if we ever
0:59:15 manage to get those apps working.
0:59:17 If you could just make that easier, I'd
just be like, great, we're just getting
0:59:20 whole groups of people together to do
these weird camping weeks where we all
0:59:23 just build home cooking software together.
0:59:25 I think that's a, that's
a nice goal to aspire to.
0:59:28 And luckily there's many, many smart
folks like some of the ones that you've
0:59:32 mentioned who are working towards that.
0:59:34 It's just, yeah, to, to absorb as much
complexity as possible and to build it in
0:59:40 a way that developer experience is nice
and where it's not a leaky abstraction.
0:59:44 That is quite the undertaking
and takes a little bit of time,
0:59:47 but I think it's on, on its way.
0:59:49 Maybe as a last question.
0:59:51 Given that you've come to Berlin to
the local-first conference and there
0:59:54 were so many interesting people there,
I'm wondering whether besides kind of
0:59:59 opening the eyes for you, in regards
to local-first software and not
1:00:04 local-first software and ruining some
non local-first software for you, where
1:00:08 there's some other takeaways, some other
insights from conversations with people?
1:00:13 Maggie's takeaways from Local- First Conference
1:00:13 I was impressed at how
much work had been done.
1:00:15 I mean, we had a few people speak
who had been part of like the offline
1:00:18 first movement before local-first
got coined as a term, who said
1:00:21 like, this has been around since,
I want to forget the earliest year,
1:00:23 what, 2008, 2009 or something,
people started trying to do this.
1:00:27 I was impressed by how much like
previous art there was, but also
1:00:31 how it hadn't progressed that much.
1:00:33 Like maybe other things became popular
in the tech space at the time, or it
1:00:36 didn't have this same like cultural
energy behind it that it kind of does now.
1:00:40 But it was both like, wow, this has
been going on for ages, and we haven't
1:00:43 actually gotten to a place where it's,
like, standardized or easy to do yet,
1:00:46 But then all the people who presented
at the conference, it made me get this
1:00:49 sense of like, Oh, things are picking up.
1:00:51 People are working things out.
1:00:53 Like a bunch of these like
wonderfully smart people have
1:00:56 solved a bunch of complex problems
over the last couple of years.
1:00:58 Like, great.
1:00:59 We are like hitting inflection points,
you know, CRDTs and like Automerge
1:01:02 and things like this are like
actually making a huge difference.
1:01:05 so I think I was just kind of blown away
by how much sophisticated, hard technical
1:01:10 work had already been invested into this.
1:01:13 philosophy or dream, that made me
go, Oh, this has to happen now.
1:01:16 Like this has to, this has to pan out.
1:01:18 Like we've hit the apex,
we're over the hump.
1:01:21 Yeah, this was certainly the
goal with the conference to just
1:01:24 bring together all the people.
1:01:26 There's so much.
1:01:27 Momentum already happening, but a lot
of it is like mostly felt on a given
1:01:34 day on someone's solitary desk and
like they're working on something and
1:01:39 think, Oh my gosh, this is so cool,
but like, does anyone else care?
1:01:43 And then bring all of
those people together.
1:01:45 The momentum compounds, and then you
go even further and show like how this
1:01:50 can have like a massive global impact.
1:01:53 this is sweet.
1:01:54 We certainly met all of those goals
and exceeded them for ourselves.
1:01:57 And thank you so much for contributing in
such a meaningful way to, to that event.
1:02:02 Yeah, no, I loved it.
1:02:03 And I'll also add on that I think I
was also impressed by, everyone in this
1:02:06 community is like, maybe, maybe by nature
of being at a conference like this, which
1:02:10 is trying to say, let's build software
differently, like, let's do a radical
1:02:14 change to the way software works, was,
you know, Then the kind of person who
1:02:18 was also thinking very creatively out
of the box, like out of the norm about
1:02:22 how we should build software, what
software is, like who builds software.
1:02:26 it was a, it's a very like holistic,
I think is the word I want to use,
1:02:29 approach to understanding what.
1:02:32 Software and programming is, which is what
I always love at the right conferences.
1:02:35 It's like, okay, we are not
just showing each other like how
1:02:37 to do API calls with some new
framework, like React on the screen.
1:02:41 Like, please God, no, not anymore.
1:02:42 But if someone could please talk about
like how this fits into like developer
1:02:46 culture or wider culture, I felt
like there was a lot of that at the
1:02:49 conference that I really appreciated.
1:02:51 Yeah, that's what I think makes it such
an inclusive and wide space, since at the
1:02:58 end of the day, the, also the local-first
essay was not very prescriptive in terms
1:03:04 of like, you ought to use CRDTs, but
laid out those ideals of like, of things
1:03:11 that I think most people can get behind.
1:03:14 and some might see it as like a nice
to have bonus and some others say
1:03:18 like, Absolutely, I won't touch any
software that doesn't respect all
1:03:22 of those ideals, but I think this,
kind of spans a very wide open tent.
1:03:27 Yeah, I feel like it's sort of.
1:03:29 Like the older people who
have been in software a while
1:03:32 sort of already knew this.
1:03:32 They're like, yeah, you don't trust
random companies with your data.
1:03:35 Like we've seen enough of them die.
1:03:37 Like they're very, like,
I keep everything locally.
1:03:39 I make it all myself.
1:03:40 And then I feel like, you know, younger
people, let's say people like born after
1:03:45 like 1985 or something, you know, we
just, we Didn't live through that, that
1:03:49 period of everything dying off as much.
1:03:51 And so a little bit more naive, a little
bit more like, yeah, of course, all my
1:03:54 data lives in the cloud at some random
person's server, of course I don't own
1:03:58 it, of course I don't have the ability
to edit my apps, the sort of, blinkered
1:04:02 by what we think is possible and what we
think we're allowed to do in software.
1:04:05 And this is kind of like unblinkering
people being like, okay, you've been
1:04:08 living in this paradigm, but this
is not the way things have to be.
1:04:13 We could change this.
1:04:14 This is just like how things unfolded
because of market like forces.
1:04:18 But.
1:04:19 What if, what if you actually
had control over your software?
1:04:21 Like crazy idea.
1:04:23 Outro
1:04:23 I love it.
1:04:24 Maggie, thank you so much for
this wonderful conversation.
1:04:27 Thank you so much again for
contributing in such a great way
1:04:31 to the conference, coining a whole
new term and inspiring many, many
1:04:35 developers and barefoot developers,
soon to be barefoot developers.
1:04:40 So thank you so much for
taking the time today.
1:04:42 Yeah, I really loved it.
1:04:43 Thanks so much for having me chat.
1:04:45 Thank you for listening to
the local-first FM podcast.
1:04:48 If you've enjoyed this episode and haven't
done so already, please subscribe and
1:04:51 leave a review wherever you're listening.
1:04:53 Please also share this
episode with others.
1:04:55 Spreading the word about the
podcast is a great way to
1:04:58 support it and to keep it going.
1:05:00 A special thanks again to Rocicorp
and Expo for supporting this podcast.
1:05:04 See you next time.