TryShadowing
Shadow YouTube. Speak English.
Home
Browse
Dictation
NEW
My Library
en
0
days
Sign in
Inside Claude Code With Its Crea… — Y Combinator shadowing | TryShadowing
TryShadowing
Shadow YouTube. Speak English.
Home
Browse
Dictation
NEW
My Library
en
0
days
Sign in
Home
Browse
Y Combinator
Inside Claude Code With Its Creator Boris Cherny
Inside Claude Code With Its Creator Boris Cherny
Y Combinator
·
50:10 · Feb 17, 2026
Start Shadowing
0:00
0:00
Record
×1
1x
VI
EN
JA
KO
ZH
FR
PT
TH
IT
DE
IPA
Pronunciation scoring isn't supported on this browser — you can still record and listen back.
At
Enthropic,
the
way
Translating…
Turn on Record to capture your voice and get scored
Smart
Karaoke
Line
1
/1397
0:00
At Enthropic, the way
0:00
that we thought about it is we don't build for the model of today.
0:03
We build for the model six months from now.
0:06
That's actually like still my advice to to founders that are building on LLM.
0:09
Just try to think about like what is
0:10
that frontier where the model is not very good at today cuz it's going
0:14
to get good at it.
0:14
All of Quad Code has just been written
0:16
and rewritten and rewritten
0:17
and rewritten over and over
0:18
and over.
0:18
There is no part of Quad Code that was around 6 months ago.
0:21
You try a thing, you give it to users, you talk to users,
0:23
you learn, and then eventually you might end up at a good idea.
0:26
Sometimes you don't.
0:27
Are you also in the back of your mind thinking
0:28
that maybe like in 6 months you won't need to prompt
0:30
that explicitly?
0:31
Like the model will just be good enough to figure out on its own?
0:34
>> Maybe in a month, >> no more need for plan mode in a month.
0:38
>> Oh my god.
0:46
Welcome to another episode of the light cone
0:49
and today we have an extremely special guest,
0:52
Boris Churnney, the creator engineer of Claude Code.
0:56
Boris, thanks for joining us.
0:58
>> Thanks for having me.
0:59
>> Thanks for creating a thing
1:00
that has taken away my sleep for about 3 weeks straight.
1:05
>> I am very addicted to Cloud Code and uh it feels like rocket boosters.
1:09
Has it felt like this for people like for you know months at this
1:13
point.
1:13
I think it was like end of November is where uh a lot of
1:16
my friends said like something changed.
1:19
>> I remember for me I felt this way
1:20
when I first created Quad Code
1:22
and I didn't yet know
1:23
if I was on to something.
1:24
I kind of felt like I was on to something
1:25
and then that's when I wasn't sleeping.
1:27
>> Yeah.
1:28
>> And that was just like three straight months.
1:30
>> This was uh September 2024.
1:33
Yeah.
1:33
It was like three straight months.
1:34
I I didn't take a single day vacation.
1:36
Worked through the weekends.
1:37
Worked every single night.
1:38
I was just like, "Oh my god,
1:39
this is I think this is going to be a thing.
1:41
I don't know if it's useful yet because it it couldn't actually code yet."
1:45
>> If you look back on uh those moments to now,
1:48
like what would be like the most surprising thing about this moment right now?
1:52
>> It's unbelievable that we're still using a terminal.
1:55
That was supposed to be the starting point.
1:56
I didn't think that would be the ending point.
1:57
And then the second one is
1:58
that it's even useful cuz uh you know at the beginning it didn't really
2:02
write code.
2:03
Even in February when we G it wrote maybe like 10% of my code
2:06
or something like that.
2:07
I didn't really use it to write code. it wasn't very good at it.
2:09
I still wrote most of my code by hand.
2:11
Uh so the fact
2:12
that it it actually like our bets paid off
2:15
and it got good at the thing
2:17
that we thought it was going to get good at
2:19
because it wasn't obvious.
2:20
At Enthropic, the way
2:21
that we thought about it is we don't build for the model of today.
2:24
We build for the model 6 months from now.
2:26
And that's actually like still my advice to to founders
2:29
that are building on LLM is,
2:31
you know, just try to think about like what is
2:32
that frontier where the model is not very good at today. um
2:35
because it's going to get good at it
2:37
and you just have to wait.
2:38
>> Going back though, but when do you remember when you first got the idea?
2:41
Can you just talk us through that?
2:42
Like was it some like a spark
2:43
or what was even the first version of it in your mind?
2:46
>> You know, it's funny.
2:47
It was like it was
2:48
so accidental that it just kind of evolved into this.
2:50
Um you know as
2:52
as anthropic I think for Ant the bet has been coding for a long
2:56
time and the bet has been the path to save to safe AGI is
3:01
through coding >> and this is this has kind of always been the idea
3:04
and the way you get there is you you teach the model how to
3:07
code then you teach it how to use tools
3:08
then you teach it how to use computers um
3:10
and you can kind of see
3:11
that because the the first team
3:12
that I joined at Enthropic it was called the anthropic labs team uh
3:16
and it produced three products it was quadcode MCP
3:19
and in the desktop app.
3:20
So you can kind of see how these like weave together.
3:23
The particular product that we built, you know,
3:26
like no one no one asked me to build a CLI.
3:29
Um we kind of knew maybe it was time to build some kind of
3:32
coding product cuz it seemed like the model was ready,
3:34
but no one had yet really built the product that harnessed this capability.
3:39
So like still there's this insane feeling of product overhang.
3:42
But at the time it was just like even crazier cuz like no one
3:44
had built this yet.
3:45
And so I I started like hacking around uh and I was like, "Okay,
3:49
we build a coding product.
3:50
What do I have to do first?
3:51
I have to understand how to use the API
3:53
because I hadn't used anthropic API at
3:55
that point."
3:56
Um and so I I just built like a little terminal app to use
3:59
the API.
4:00
That's all that I did.
4:00
And it was a little chat app
4:01
because you know like you think about the you know AI applications of the
4:05
time and you know for non-coders today most what what are most people using
4:08
is just a chat app.
4:09
So that's what I built.
4:11
Uh and you know it was in a terminal.
4:12
I can ask questions.
4:13
I give answers.
4:15
Then I think tool use came out.
4:16
I just wanted to try out tool use
4:18
because I I don't really understand what this is.
4:19
I was like to use this is cool.
4:21
Is this actually useful?
4:22
Probably not.
4:22
Let me just try it.
4:23
>> You built it in terminal just
4:25
because it was the easiest way to get something up
4:26
and running.
4:27
>> Yes.
4:28
Cuz I didn't have to build a UI.
4:30
>> Okay.
4:30
>> It was just me >> at that point.
4:31
It was like the IDEs, Cursor, Windsurf taking off.
4:35
Were you sort of under any pressure or getting lots of suggestions of, hey,
4:38
like we should build this out
4:40
as a plugin or
4:41
as a as a fully featured ID itself?
4:43
There was no pressure because we didn't even know what we wanted to build.
4:46
Like the the team was just in explore mode, you know,
4:48
like we we didn't we know vaguely we wanted to do something in coding,
4:51
but it wasn't obvious what no one was high confidence enough.
4:54
That was like my job to figure out.
4:55
And so I g I gave the model uh the batch tool.
4:58
That was the first tool
4:58
that that I gave it just cuz I think
5:00
that was literally the example in our docs.
5:03
I just like took the example.
5:04
It was in Python.
5:04
I just ported it to TypeScript because that that's how I wrote it.
5:07
You know, I didn't know like what the model could do with bash.
5:08
So I asked it to like read a file.
5:10
It could like cat the file.
5:11
So like that was cool.
5:12
And then I was like, "Okay, like what can you actually do?"
5:14
And and I asked her, "What music am I listening to?"
5:17
He wrote some like Apple script to script my my Mac
5:21
and look up the music in my music player.
5:23
>> Oh my god.
5:24
>> And this was Sauna 3.5.
5:25
>> And you know, like I I didn't think the model could do that.
5:30
And that was my first I think ever fuel the AGI moment >> where
5:33
I was just like,
5:33
"Oh my god, the model it just wants to use tools.
5:37
That's all it wants."
5:38
>> That's kind of fascinating.
5:39
I mean it's very kind of contrarian
5:42
that clocker works so well in such an elegant simple form factor.
5:47
I mean terminals have been around for a really long time
5:50
and that seemed to be like a good design constraint
5:54
that allowed a lot of interesting developer experiences like it doesn't feel like working.
5:59
It just feels fun as a developer.
6:01
I don't think about files where everything is and that came by accident almost.
6:07
>> Yeah, it was an accident.
6:08
I remember so after the terminal started to take off internally.
6:12
Um and honestly like after building this thing I think like 2 days after
6:15
the first prototype I started giving it to my team just for dogfooting cuz
6:19
you know like you know
6:20
if you come up with an idea
6:21
and it seems useful the first thing you want to do is you want
6:23
to give it to people to see how they use it.
6:25
And then I came in the next day
6:26
and then Robert who sits across from me who's another engineer he he just
6:30
like had quad code on his computer
6:32
and he was like using it to code.
6:33
I was like I was like what what are you what are you doing?
6:35
Like this thing isn't ready.
6:36
It's just a prototype.
6:37
But yeah, it it was already useful in that form factor.
6:40
And I remember when we did our launch review to kind of launch quad
6:44
code externally,
6:44
this was in December, November, something like that in 2024.
6:49
Um Dario asked and he was like,
6:51
"The us chart internally like the the Dow chart is like vertical.
6:54
Are you like forcing engineers to use it?
6:56
Like why are you mandating them?"
6:58
>> And I was just like, "No, no, we didn't.
7:00
We I just like posted about it
7:01
and they they' just been like telling each other about it."
7:04
Honestly, it was it was just accidental.
7:05
We we started with the CLI
7:06
because it was the cheapest thing
7:07
and it just kind of stayed there for a bit.
7:09
>> So in that 2024 period, what how were the engineers using it?
7:14
Were they sort of shipping code with it yet
7:16
or were they using it in a different way?
7:18
>> The model is not very good at coding yet.
7:20
I I was using it personally for automating git.
7:23
Um I think at this point I I probably forgotten most of my git
7:26
because cloud code has just been doing it for
7:28
so long.
7:29
But yeah, like automating uh bash commands
7:31
that that was a very early use case
7:32
and like operating like Kubernetes
7:34
and kind of things like this.
7:35
People were using it for coding.
7:37
So there were some early signs of this.
7:38
I think the first use case was actually writing unit tests
7:40
because it's a little bit lower risk
7:42
and the model was still pretty bad at it >>
7:44
but people were were were kind of figuring it out
7:46
and and they were figuring out how to use this thing.
7:49
>> Um and one thing
7:49
that we saw is people started writing these markdown files for themselves
7:54
and then having the model read
7:55
that markdown file.
7:56
And this is where QuadMD came from.
7:58
Probably the single for me biggest principle in product is latent demand.
8:02
Um and the just every bit of this product is built through latent demand
8:05
after their initial CLI.
8:08
Uh and so quadmd is an example of that.
8:10
There's this other general principle
8:11
that I think is maybe interesting where you can build for the model
8:15
and then you can build scaffolding around the model in order to improve performance
8:19
a little bit and depending on the domain you can improve performance maybe 10
8:23
20% something like that
8:24
and then essentially the gain is wiped out with the next model.
8:27
So either you can build build the scaffolding
8:29
and then you know get some performance gain
8:31
and then rebuild it again
8:33
or you just wait for the next model
8:34
and then you kind of get it for free. the quantumd
8:36
and kind of the scaffolding is an example of
8:38
that and really I think that's why we stayed in the CLI is
8:41
because we felt there is no UI we could build
8:44
that would still be relevant in 6 months
8:46
because the model was improving
8:48
so quickly >> earlier we were saying like we should compare cloud MDs
8:51
but you said something very profound
8:53
which is you know yours is actually very short
8:56
which is almost like the opposite of what you know people might expect why
9:00
is that what's in your cloud MD >> okay
9:02
so I I checked this before we came
9:04
so my my cloud has two Um,
9:06
one is, uh, there it's just two lines.
9:10
So, the first line is whenever you put up a PR, enable automerge.
9:14
Um, so as soon as someone accepts it, it's merged.
9:16
That's just so I can like code
9:17
and I don't have to kind of go back
9:18
and forth with CR
9:19
or whatever.
9:20
And then the second one is whenever I put up a PR,
9:22
post it in our internal team stamps channel.
9:25
Uh, just so someone can stamp it and I can get unblocked.
9:28
Uh, and the idea is every other instruction is in our quadmd that's checked
9:33
into the codebase and it's something our entire team contributes to multiple times a
9:37
week.
9:38
And very often I'll see someone's PR
9:41
and they make some like mistake that's totally preventable
9:44
and I'll just literally tag Claude on the PR.
9:46
I'll just do like add quad, you know,
9:48
like add this to the quad MD and I'll do this, you know,
9:50
like many times a week.
9:51
>> Do you have to like compact the Claude MD?
9:53
Like I definitely reached a point where I got the message at the top
9:56
saying your cloud MD is like thousands of tokens now.
9:59
What do you do when you guys hit that?
10:01
>> So our quadm is actually pretty short.
10:03
I think it's like couple thousand tokens maybe something like that.
10:06
Um if if you hit this my recommendation would be delete your quadmd
10:09
and just start fresh.
10:11
>> Interesting.
10:11
>> I think a lot of people like they try to overengineer this right
10:14
and and really like the capability changes with every model.
10:17
And so the thing
10:18
that you want is do the minimal possible thing in order to get the
10:22
model on track.
10:23
And so if you delete your quadd
10:24
and then you know the model is getting off track,
10:26
it does the wrong thing.
10:27
That's when you kind of add back a little bit at a time.
10:29
And what you're probably going to find is with every model,
10:31
you have to add less and less.
10:33
For me, I consider myself a pretty average engineer to be honest.
10:36
Like I don't use a lot of fancy tools.
10:38
Like I I don't use like Vim.
10:39
I use, you know, VS Code because it's simpler.
10:41
Um I don't really >> Wait, really?
10:43
I would have assumed
10:44
that because you built this in the terminal
10:45
that you were sort of like a dieh hard ter terminal like Vim Vim
10:48
only person you know screw those VS code people you know >> well we
10:52
have people like that on the team there's you know like Adam Wolf for
10:54
example he's he's on the team he's like you will never take Vim for
10:57
my cold dead hands like yeah
10:59
so there's definitely a lot of people like
11:00
that on the team
11:01
and this is one of the things
11:02
that I learned early on is every engineer likes to hold their dev tools
11:05
differently they like to use different tools there's just no one tool
11:07
that works for everyone
11:08
but I think also this is one of the things
11:10
that makes it possible for quad code to be
11:13
so good because I kind of think about it
11:15
as what is the product
11:16
that I would use
11:17
that makes sense to me
11:18
and so to use quad code you don't have to understand Vim you don't
11:21
have to understand TMX you don't have to know how to like SSH you
11:24
don't have to know all the stuff you just have to open up the
11:26
tool and it'll guide you it'll it'll do all this stuff >> how do
11:29
you decide how verbose you want like sort of the terminal to be like
11:32
sometimes you have to go you know control O
11:34
and check it out
11:36
and is it like internal bike shed battles around like longer shorter I mean
11:41
every user probably has a for an opinion like how do you make those
11:44
sorts of decisions?
11:45
>> What What's your opinion?
11:46
Is it is it too verbose right now?
11:48
>> Oh, I love the verbosity cuz basically sometimes it just like goes off
11:51
the deep end and I'm watching
11:53
and then I can just read very quickly
11:55
and it's like,
11:55
"Oh, no, no, it's not that."
11:57
And then I escape
11:57
and then just stop it
11:58
and then it just like stops an entire bug farm like
12:02
as it's happening.
12:03
I mean, that's usually when I didn't do plan mode properly.
12:05
>> This is something that we probably change pretty often.
12:09
Um, I remember early on, this is maybe six months ago,
12:11
I tried to get rid of bash output just internally just to like summarize
12:14
it because I was like these giant long bash commands,
12:16
I don't actually care.
12:18
And then I gave it to anthropic employees for a day
12:19
and everyone just revolted.
12:22
I want to see my dash because it it actually is quite useful for,
12:25
you know, like for something like git output, maybe it's not useful,
12:28
but if you're running, you know, like Kubernetes jobs or something like this,
12:31
you actually do want to see it.
12:32
We recently hit the hid the file reads and uh file searches.
12:36
So you'll notice instead of saying, you know, like read food.md said,
12:39
you know, like read one file, search searched one pattern.
12:43
And this is something I think we could not have shipped six months ago
12:45
because the model just was not ready.
12:47
It would have, you know, it still read the wrong thing pretty often.
12:49
As a user, you still had to be there
12:50
and kind of catch it
12:51
and debug it.
12:52
But nowadays, I just noticed it's on the right track almost every time.
12:55
And because it's using tools so much,
12:58
it's actually a lot better just to summarize it.
13:00
Um, but then we shipped it.
13:01
Uh, we dog fooded it for like a month
13:03
and then people on GitHub didn't like it.
13:05
Uh so there was a big issue where people like no like I want
13:07
to see the details
13:08
and that was really great feedback.
13:09
Um and so we added a new verbose mode
13:12
and so that's just like in slash config you can enable verbose mode
13:15
and if you want to see all the file outputs you can continue to
13:17
do that and then I posted on the issue
13:20
and people still still didn't like it
13:21
which is again awesome
13:22
because like my favorite thing in the world is just hearing people's feedback
13:25
and hearing how they actually want to use it.
13:27
Um and so we just like iterated more
13:29
and more and more to get
13:29
that really good and to make it the thing
13:31
that people want.
13:32
I'm amazed like how much I enjoy uh fixing bugs now.
13:35
And then all you have to do is uh have really good logging
13:39
and then even just say like hey check out
13:41
that you know this particular object it messed up in this way
13:45
and it like searches the log.
13:47
It figures everything out.
13:48
It can like go into your you can make a production tunnel
13:50
and it'll look at your production DB for you.
13:52
It's like this is insane.
13:54
Bug fixing is just going to sentry copy markdown.
13:57
You know pretty soon it's just going to be straight MCP.
13:59
It's like an autobug fixing like
14:01
and test making sort of uh what's the new uh term they call it
14:06
like a making a startup factory.
14:08
Oh yeah.
14:08
>> Right.
14:09
There's like all these concepts now of rather than having to review the code,
14:13
you know, I'm I'm old school, so I like the verbosity.
14:17
I like to say, "Oh, well, you're doing this,
14:19
but I want you to do that."
14:20
Right?
14:21
But there's a totally different school of thought now
14:23
that says like anytime an a real human being has to look at code
14:28
uh that's bad.
14:29
>> Yeah.
14:29
Yeah.
14:30
Yeah.
14:30
>> Which is fascinating.
14:31
>> I think like Dan Chipper talks about this a lot
14:33
as kind of when whenever you see the model make a mistake try to
14:36
put in the quadmd try to put it in like skills
14:38
or something like that
14:39
so it's reusable.
14:40
But I I think there's this meta point
14:42
that I actually struggle with a lot.
14:44
And I people talk about like agents can do this, agents can do that,
14:47
but actually what agents can do, it changes with every single model.
14:51
And so sometimes there's a new person
14:52
that joins the team
14:53
and they actually use quad code more than I would have used it.
14:56
>> And I'm just constantly surprised by this.
14:59
Like for example, there was a we had like a memory leak
15:02
and we were trying to debug it.
15:03
Um and by the way,
15:04
like Jared Sumar has just been on this crusade killing all the memory leaks
15:07
and it's just been amazing.
15:08
But before Jared was on the team,
15:10
I had to do this and there was this memory leak.
15:12
I I was trying to debug it.
15:13
And so I I took a heap dump.
15:15
I opened it in DevTools.
15:16
I was looking through the profile.
15:17
Then I was looking through the code
15:18
and I I was just trying to figure this out.
15:20
And then another engineer on the team, Chris, he just like asked Quad Code.
15:25
He was like, "Hey, I think there's a memory leak.
15:26
Can you like run this?"
15:27
And then like try to figure it out.
15:28
And Quad Code like took the heap dump.
15:30
It wrote a little tool for itself to like analyze the heap dump.
15:33
And then it found the leak faster than I did.
15:36
And this is just something I have to constantly relearn
15:39
because my brain is still stuck somewhere six months ago at times.
15:44
>> So what would be some advice for technical founders to really become maximalists
15:50
at the latest model release?
15:52
It sounds like people off of fresh off of school
15:54
or that don't have any assumptions might be better suited than maybe sometimes engineers
15:59
who have been working at it for a long time.
16:01
And how do the experts get better?
16:05
I think for yourself it's kind of beginner mindset
16:07
and uh I don't know maybe just like humility like I feel like engineers
16:12
as a discipline we've learned to have very strong opinions
16:15
and senior engineers are kind of rewarded for this in my old job at
16:18
a big company when I hired like architects
16:20
and this kind of a type of engineer you look for people
16:22
that have a lot of experience
16:23
and really strong opinions
16:25
but it actually turns out a lot of this stuff just isn't relevant anymore
16:28
and a lot of these opinions should change
16:30
because the model is getting better um
16:33
so I think actually the biggest skill is people
16:34
that can think scientifically
16:36
and can just think from first principles.
16:38
>> How do you screen for
16:39
that when you try to hire someone now for for your team?
16:42
>> I sometimes ask about what's an example of when you're wrong.
16:45
It's a really good one.
16:46
You know, some of these like classic behavioral questions like not even coding questions
16:49
I think are quite useful
16:50
because you can see
16:51
if people can recognize their mistake in hindsight,
16:53
if they can claim credit for the mistake
16:55
and if they learn something from it.
16:58
And I think a lot of these like very senior people especially there there
17:01
are some founder types like this
17:02
but I think founders in particular are actually quite good at it.
17:05
Um but other people sometimes will never really take uh they'll never take the
17:09
blame for a mistake.
17:11
But I don't know like for me personally I'm wrong probably half the time.
17:14
Like half my ideas are bad
17:16
and you just have to try stuff
17:17
and you know you try a thing you give it to users you talk
17:20
to users you learn
17:21
and then eventually you might end up at a good idea.
17:23
Sometimes you don't.
17:25
And this is the skill
17:26
that I think in in the past was very important for founders,
17:30
but now I think it's very important for every engineer.
17:32
>> Do you think um you would ever hire someone based on the uh
17:36
claude code transcript of uh them working with the agent cuz we're actively doing
17:41
that right now.
17:43
We just added uh just
17:44
as a test like you can upload a transcript of you coding a feature
17:48
with cloud code or codeex
17:50
or whatever it is.
17:51
Personally, I think that like it's going to work.
17:53
I mean, you can figure out uh how someone thinks,
17:56
like whether they're looking at the logs or not,
17:59
like can they correct the agent if it goes off off the rails?
18:02
Like, does do they use plan mode?
18:04
You know, when they use plan mode,
18:06
do they make sure that there are tests or you know,
18:08
all of these different things that, >> you know, do they think about systems?
18:11
Do they even understand systems?
18:12
Like, there's just so much that's sort of embedded in that that I imagine.
18:16
I just want like a spider uh a spiderweb graph, you know,
18:19
like in those video games like NBA 2K.
18:22
It's like, oh, this person's really good at shooting or defense.
18:24
It's like you could imagine a spiderweb graph of like, you know,
18:28
someone's claude code skill level.
18:30
>> Yeah.
18:30
What would what would the skills be?
18:31
What would be those?
18:32
>> I mean, I think it's like systems testing must be like user behavior.
18:37
I mean, there's got to be a design part like product sense maybe also
18:41
just like automating stuff.
18:42
Mhm.
18:42
My favorite thing in CloudMD uh for me is I have a thing
18:46
that says for every plan decide whether it's overengineered,
18:50
underengineered, or perfectly engineered and why.
18:53
>> I think this is something that we're trying to figure out, too,
18:55
cuz I I think uh
18:56
when I look at engineers on the team
18:57
that I think are the most effective,
18:59
there's essentially two, it's very biodal.
19:01
Um there's one side where it's extreme specialists.
19:04
Um and so like I named Jared before,
19:06
like he's a really good example of this
19:07
and kind of the bun team is a really good example.
19:09
Just hyper specialist.
19:10
They understand dev tools better than anyone else.
19:13
They understand JavaScript runtime systems better than anyone else.
19:16
And then there's the flip side of kind of hyper generalists
19:18
and that's kind of the rest of the team.
19:20
And a lot of people they span like product
19:22
and info or product
19:23
and design um or you know like product
19:25
and user research,
19:26
product and business.
19:27
I really like to see people that just do weird stuff.
19:31
I think that's one of these things
19:32
that was kind of a warning sign in the past
19:34
because it's like can these people actually build something useful?
19:37
>> Um that's the limits test.
19:39
Yeah, that's what must
19:40
but but nowadays like for example an engineer on the team Daisy,
19:44
she was on a different team
19:45
and then she transferred onto our team
19:47
and the reason that I wanted her to transfer is she put up a
19:51
PR for Claude Code like a couple weeks after she joined
19:54
or something and the PR was to add a new feature to Claude Code
19:58
and then instead of just adding the feature what she did is first she
20:01
put up a PR to give Claude code a tool
20:03
so that it can test an arbitrary tool
20:06
and verify that that works.
20:07
And then she put up
20:08
that PR and then she had Quad write its own tool instead of herself
20:11
implementing it.
20:12
And I think it's this kind of out of the box thinking
20:14
that is is just
20:15
so interesting because not a lot of people get it yet.
20:17
You know, like we use the Quad agents SDK to automate pretty much every
20:21
part of development.
20:22
It automates code review, security review.
20:24
Uh it labels all of our issues.
20:26
It shephards things to production.
20:28
It does pretty much everything for us.
20:30
But I think externally I'm seeing a lot of people start to figure this
20:33
out,
20:34
but it's actually taken a
20:35
while to figure out how do you use LMS in this way?
20:37
How do you use this new kind of automation?
20:39
So it's kind of a new skill.
20:40
>> I guess one of the uh funnier things
20:42
that I've been having office hours with various founders about is um you have
20:46
like sort of the visionary founder who has like the idea they've like built
20:50
this like crystal palace of the product
20:53
that they want to build. they've totally loaded in their brain,
20:56
you know, who the user is
20:57
and what they feel
20:58
and what they're motivated by
21:00
and then they're sitting in claude code
21:02
and they can do like,
21:03
you know, 50x work
21:05
and then but they have engineers who work for them who like don't have
21:08
the,
21:09
you know, crystal memory palace of like the platonic ideal of the product
21:12
that the pro founder has
21:14
and they can only do like 5x work.
21:16
Are you hearing stories like that?
21:18
there's usually a person who's like the core like designer of a thing
21:23
and they're just like,
21:25
you know, trying to blast it out of their brain.
21:27
What's the nature of like teams like that?
21:30
You know, it seems like that's almost a stable configuration.
21:33
Like you're going to have the visionary who like now is unleashed,
21:37
but you know, maybe going back to the top of it,
21:40
like I'm experiencing this right now.
21:41
I was like, "Oh, well, I'm only a solo person and you know,
21:45
I need to eat and sleep and I have, you know, a whole job.
21:49
It's like, how am I going to do this?"
21:50
You know, >> you know, like we just launched quad teams and, you know,
21:53
this is a way to do it,
21:54
but you can also just build your own way to do it.
21:56
It's pretty easy.
21:57
>> What's the vision for cloud teams?
21:59
>> Just collaboration.
22:00
It's like there's this whole new field of like agent top apologies
22:03
that people are exploring.
22:03
Like what are the ways that you can configure agents?
22:06
There's this one sub idea which is uncorrelated context windows.
22:09
And the idea is just multiple agents,
22:11
they have fresh context windows
22:12
that aren't essentially polluted with each other's context
22:14
or their own previous context.
22:16
And if you throw more context at a problem,
22:18
that's like a form of test time compute.
22:19
Um, and so you just get more capability that way.
22:22
And then if you have the right topology on top of it,
22:24
so the agents can communicate in the right way,
22:26
they're laid out in the right way, then they can just build bigger stuff.
22:30
And so Teams is kind of like one idea.
22:32
There's a few more that are coming pretty soon.
22:34
Um, and the idea is just maybe it can build a little bit more.
22:38
I think the first kind of big example where it worked is our plugins
22:41
feature was entirely built by a swarm over over a weekend.
22:45
It just ran for like a few days.
22:46
There wasn't really human intervention.
22:49
And plugins is pretty much in the form
22:50
that it was when
22:51
when it came out.
22:52
>> How did you set that up?
22:53
Like did you spec out sort of the outcome
22:56
that you were hoping for
22:57
and then let it sort of figure out the details
23:00
and then like let it run?
23:02
>> Yeah. an engineer on the team just gave uh gave Quad a spec
23:06
and um told Quad to use a Asauna board
23:09
and then Quad just put up a bunch of tickets on a sauna
23:13
and then spawned a bunch of agents
23:14
and the agent started picking up tasks.
23:16
The main quad just gave it instructions
23:17
and they all just figured it out >> like independent um agents
23:21
that didn't have the context of the bigger spec.
23:23
Right.
23:24
>> Right.
23:24
If you if you think about the way
23:25
that uh you know like how our agents actually started nowadays
23:28
and you know I haven't pulled the data on this
23:30
but I would bet the majority of agents are actually prompted by quad today
23:34
in the form of uh sub agents cuz like a sub agent is just
23:37
like a recursive quad code that's all it is in the code
23:39
and it's just prompted by we call her mama quad >>
23:43
and that that's all it is
23:44
and I think probably
23:45
if you look at most agents they're launched in this way >> my claude
23:48
insights just told me to do this more for debugging
23:50
so that I get like I spend a lot of time on debugging
23:53
And it would just be better to have like multiple sub agents spin up
23:57
and like debug something in parallel.
23:59
And so then I just like added
24:00
that to my claude MD to just be like,
24:02
hey, like next time you try
24:02
and fix a bug like have one agent
24:04
that like looks in the log,
24:05
like one that looks in the code path.
24:07
That just seems sort of inevitable.
24:09
>> For weird scary bugs,
24:10
I try to uh fix bugs in plan mode
24:12
and then it seems to use the agents to sort of search everything.
24:16
Whereas like when you're just trying to do it in line, it's like, okay,
24:19
I'm going to do like this one task instead of search wide.
24:22
This is something I do all the time too.
24:24
I I just say if the if the test seems kind of hard,
24:27
this kind of research test,
24:28
I'll calibrate the number of sub aents I ask it to use based on
24:30
the difficulty of the task.
24:32
>> So if it's like really hard,
24:33
I'll say like use three or maybe five or even 10 sub aents,
24:36
research in parallel and then see what they come up with.
24:39
>> I'm curious.
24:39
So then why don't you put that in your clawed MD file?
24:42
>> It's kind of case by case, you know, like quadm like what is it?
24:46
It's just a it's a shortcut.
24:48
Like if you find yourself repeating the same thing over and over,
24:50
you put in the quad MD.
24:52
But otherwise, you don't have to put everything there.
24:53
You can just prompt quad.
24:54
>> Are you also in the back of your mind thinking
24:56
that maybe like in six months,
24:57
you won't need to prompt that explicitly?
25:00
Like the model will just be good enough to figure out on its own.
25:03
>> Maybe in a month.
25:05
>> No more need for plan mode in a month.
25:07
>> Oh my god.
25:08
>> I think plan mode probably has a limited lifespan.
25:11
>> Interesting.
25:11
>> That's some alpha for everyone here.
25:12
What would the world look like without plan mode?
25:14
Do you just describe it at the prompt level
25:17
and it would just do it?
25:18
One shot it?
25:19
Yeah, we've uh we've started experimenting with this
25:21
because quad code can now enter plan mode by itself.
25:24
I don't know if you've you guys have seen that.
25:26
>> Yeah.
25:26
>> So, we're trying to kind of get this experience really good.
25:29
So, it would enter plan mode the same point where a human would have
25:32
wanted to enter it.
25:33
So, I think it's like I think it's something like this,
25:36
but actually plan mode there's no there's no big secret to it.
25:38
All it does is it adds one sentence to the prompt that's like please
25:42
don't code.
25:43
>> That's all it is.
25:44
You can you can actually just say that.
25:45
>> Yeah.
25:45
So it sounds like a lot of the feature development for clock code is
25:48
very much a what we talk about a YC talk to your users >>
25:52
and then you come
25:53
and implemented it.
25:54
It wasn't the other way
25:55
that you had this master plan
25:56
and then implemented all the features.
25:58
>> Yeah.
25:58
Yeah.
25:58
I mean that that's all it was like plan mode was we saw users
26:01
that that were like hey quad come up with an idea plan this out
26:04
but don't write any code yet.
26:05
And there was kind of various versions of this.
26:07
Sometimes it was just talking through an idea.
26:09
Sometimes it was these very sophisticated specs
26:11
that that they were asking Claude to write,
26:13
but the common dimension was do a thing without coding yet.
26:16
And so literally like this was like Sunday night at 10 p.m.
26:19
I was I was just like looking at GitHub issues
26:21
and kind of seeing what people were talking about
26:23
and looking at our internal Slack feedback channel
26:25
and I just wrote this thing in like 30 minutes
26:28
and then uh shipped it
26:29
that night.
26:29
It went out Monday morning.
26:30
That was plan mode.
26:31
So do you mean
26:32
that there will be no need for plan mode to in the sense of
26:35
I'm worried that the model's going to do like it's going to do like
26:38
the wrong thing or head off in the wrong direction
26:40
but there will still be a need for
26:42
that.
26:42
You need to think through the idea
26:44
and figure out exactly what it is
26:45
that you want and you have to do
26:46
that somewhere.
26:47
>> I kind of think about it in terms of like kind of increasing
26:50
model capabilities.
26:51
So maybe 6 months ago a plan was insufficient.
26:54
So you get Claude to make a plan.
26:56
Let's say even with plan mode you still have to kind of sit there
26:58
and babysit cuz it can go off track.
27:00
Nowadays what I do is probably 80% of my sessions I say I say
27:03
plan mode has a limited lifespan
27:04
but I I'm a heavy plan mode user.
27:06
Um I probably 80% of my sessions I start in plan mode
27:10
and claude will you know it'll start it'll start making a plan.
27:13
I'll move on to my second terminal tab
27:15
and then I'll have it make another plan
27:17
and then when I run out of tabs I open the desktop app
27:19
and then I go to the code tab
27:20
and then I just start a bunch of tabs there
27:22
and they all start in plan mode probably know like 80% of the time.
27:25
Once the plan is good, and sometimes it takes a little back and forth,
27:28
they just get clawed to execute.
27:30
And nowadays, what I find with Opus 4.5,
27:33
I think it started with 4.6 it got really good.
27:36
Once the plan is good,
27:37
it just stays on track
27:38
and it'll just do the thing exactly right almost every time.
27:41
And so, you know,
27:42
before you had to babysit after the plan and before the plan,
27:45
now it's just before the plan.
27:46
So, maybe the next thing is you just won't have to babysit.
27:49
You can just kind of give a prompt and Quad will figure it out.
27:51
>> The next step is Claude just speaks to your users directly.
27:56
>> >> Yeah, it just bypasses you entirely.
27:58
>> It's funny.
27:58
This is actually the current stuff for us.
28:00
Our quads actually like they talk to each other.
28:01
They talk to our users on Slack, at least internally pretty often.
28:06
Um, my quad will like tweet once in a while.
28:08
>> No way.
28:08
>> Um, but I actually like delete it.
28:10
It's just like it's a little like cheesy.
28:13
Like I don't love the tone.
28:14
>> What does it want to tweet about?
28:16
>> Sometimes it'll just like respond to someone cuz I always have like co-work
28:19
in the background and it's like it's the co-work
28:21
that really loves to do
28:22
that because it likes using a browser.
28:23
>> That's funny.
28:24
A a really common pattern is I ask Quad to build something.
28:26
It'll look in the codebase.
28:28
Uh it'll see some engineer touch something in the git flame
28:31
and then it'll message
28:32
that engineer on Slack.
28:33
Um just like asking a clarifying question and then once it gets answer back,
28:36
it'll keep going.
28:37
>> What are some tips for founders now on how to build for the future?
28:42
Sounds like everything is really changing.
28:44
What are like some principles that will stay on and what will change?
28:48
>> So I think some of these are pretty are pretty basic,
28:50
but I think they're even more important now than they were before.
28:53
Um, so one example is latent demand.
28:55
Like I mentioned it a thousand times for me.
28:57
It's just like the single biggest idea in product.
29:00
It's a it's a thing that no one understands.
29:02
It's a thing I certainly did not understand my first few startups.
29:06
And and the idea is like people will only do a thing
29:08
that they already do.
29:09
You can't get people to do a new thing.
29:10
If people are trying to do a thing and you make it easier,
29:13
that's a good idea.
29:14
But if if people are doing a thing
29:15
and you try to make them do a different thing,
29:17
they're not going to do that.
29:18
And so you just have to make the thing
29:19
that they're trying to do easier.
29:20
And I think quad is going to get increasingly good at kind of figuring
29:23
out these kind of product ideas for you just
29:25
because it can look at feedback,
29:26
it can look at debug logs, it can kind of figure this out.
29:28
>> That's what you mean by plan mode was latent demand
29:31
that people were already like I don't know had their clawed chat window open
29:34
in a browser and were like talking to it to figure out like the
29:38
spec and and what it should do.
29:40
And now that like pi mode just became
29:43
that you just do it in claw code.
29:44
>> Yeah.
29:44
Yeah, that's it.
29:45
Some sometimes what I'll do is I'll just walk around the office on on
29:48
our floor and I'll just kind of stand behind people like I I'll say
29:52
like hi so it's not
29:53
and then um I'll I'll just see kind of like how they're using quad
29:56
code.
29:56
Um and this is also just something I saw a lot um
29:58
but it also came up in GitHub issues like people were talking about it.
30:01
It seems like so you're surprised how far the terminal has gone
30:05
and how far it's been pushed like how far do you think it has
30:08
left to go just given with this world of swore multiple agents like do
30:13
you think there's going to be a new a need for a different UI
30:17
on top of it?
30:17
>> It's funny if you asked me this a year ago I would have
30:19
said the terminal has like a threemonth lifespan
30:21
and then we're going to move on to the next thing.
30:23
Uh and you can see us experimenting with this right
30:25
because quad code started in a terminal
30:27
but now it's in you know it's on web you can like quadcode it's
30:31
in the desktop app you know we've had
30:32
that for you know like three months
30:33
or six months or something just in the code tab um it's in the
30:37
iOS and Android apps just like in the code tab it's in slack it's
30:40
in GitHub there's VS Code extensions there's Jet Brains extensions
30:44
so we're just like we're always experimenting with different form factors for this thing
30:48
to figure out what's the next thing I've been wrong
30:51
so far about the of the CLI.
30:53
So, I'm probably not the person to forecast that.
30:56
>> What about like your advice to DevTool founders?
30:59
Like, someone's building a DevTool company today.
31:02
Should they just like be building for engineers
31:04
and humans or should they be thinking more about like what Claude going to
31:08
think and want and build for sort of like the agent?
31:11
>> The way I would frame it is think about the thing
31:13
that the model wants to do
31:15
and figure out how do you make
31:17
that easier.
31:19
And that's something that we saw, you know,
31:20
like when I first started hacking on quad code,
31:22
I I realized like this thing just wants to use tools.
31:24
It just wants to interact with the world.
31:27
And how how do you how do you enable that?
31:29
Well, the way you don't do it is you put it in a box
31:32
and you're like,
31:32
here's the API, here's how you interact with me,
31:34
and here's how you interact with the world.
31:36
The way you do it is you see what tools it wants to use.
31:38
You see what it's trying to do,
31:39
and you enable that the same way that you do for your users.
31:43
And so, like for if you're building a dev tool startup,
31:45
I would think about like what is the problem you want to solve for
31:47
the user?
31:48
And then when you use when you apply the model to solving this problem,
31:50
what is the thing the model wants to do?
31:52
>> And then what is the technical
31:53
and product solution that serves the weight
31:55
and demand of both?
31:56
YC's next batch is now taking applications.
31:59
Got a startup in you?
32:01
Apply at y combinator.com/apply.
32:04
It's never too early and filling out the app will level up your idea.
32:08
Okay, back to the video.
32:10
Back in the day, more than 10 years ago,
32:13
you were a very heav heavy user and you wrote a book about TypeScript,
32:17
right?
32:19
Before Typescript was cool.
32:20
This is when everyone was a deep in JavaScript.
32:23
This is back in early 2010s, right?
32:25
>> Yeah, something like that.
32:27
>> Before Typescript was a thing because back then is a very weird language.
32:32
It's not supposed to do a lot of things with being typed in JavaScript
32:36
and now it's the right thing
32:38
and it feels like clot code in the terminal has a lot of parallels
32:41
with TypeScript at the beginning.
32:44
>> TypeScript makes a lot of really weird language decisions.
32:48
So if you look at the type system pretty much anything can be a
32:51
literal type for example
32:53
and this is like this is super weird cuz like even like like Haskell
32:57
doesn't even do this.
32:58
It's just like it's too extreme
33:00
or it has like conditional types
33:01
which I don't think any language thought of at all.
33:05
>> It was like very strongly typed.
33:06
>> Yeah, it was very strongly
33:08
and and the idea was like
33:09
when you know like
33:09
when Joe Pamer and Anders
33:11
and the early team was like building this thing,
33:14
the way they built it is we okay,
33:15
we have these teams with these big untyped JavaScript code bases.
33:18
We have to get types in there,
33:19
but we're not going to get engineers to change
33:21
that the way that they code.
33:22
You're not going to get JavaScript people to have like, you know,
33:25
15 layers of class inheritance like you would a Java programmer, right?
33:28
They're going to write code the way they're going to write it.
33:30
They're they're going to use reflection
33:31
and they're going to use mutation
33:32
and they're going to use all these features
33:34
that traditionally are very very difficult to type.
33:36
>> They're a very unsafe type to any strong functional programmer.
33:40
>> That's right.
33:40
That's right.
33:40
That's right.
33:41
And so the thing
33:41
that they did instead of getting people to kind of change the way
33:44
that they code,
33:45
they they built a type system around this.
33:47
And it was just it's brilliant
33:49
because there's all these ideas
33:50
that no one was thinking about even in academia like no one thought of
33:54
a bunch of these ideas.
33:55
It purely came out of the practice of observing people
33:57
and seeing how JavaScript programmers want to write code.
34:00
And so you know for for quad code it there there are some ideas
34:03
that are kind of similar in
34:04
that you know like you can use it like a Unix utility.
34:07
You can pipe into it.
34:08
You can pipe out of it.
34:09
Um in some ways it is kind of rigorous in this way
34:11
but in in almost every other way it's just the tool
34:14
that we wanted. like I I build a tool for myself
34:17
and then the team builds the tool for themselves
34:19
and then for anthropic employees
34:21
and then for users
34:22
and it just ends up being really useful.
34:24
It's not it's not this like principled
34:26
and academic thing which I think the the proof is actually in the results.
34:30
Now fast forward more than 15 years later not many codebases are in Haskell
34:36
which is more academic
34:38
and there's tons of them now on TypeScript
34:40
because it's way more practical >> right >>
34:42
which is interesting.
34:43
Yeah, it is interesting, right?
34:44
It's like TypeScript solves a problem.
34:45
>> I guess one thing that's cool, I don't know how many people know,
34:48
but the terminal is actually one of the most beautiful terminal apps out there
34:53
and is actually written with React terminal.
34:57
>> When I first started building it, you know,
34:58
like I I did front-end engineering for for a while.
35:01
So, and I was also like a, you know,
35:02
I'm I'm sort of like a hybrid,
35:04
like I do like design and user research and, you know,
35:07
write code and all this stuff.
35:08
And we love hiring engineers that are like this.
35:10
Um, so we just we love generalists.
35:12
So for me it's like okay, I'm building a thing for the terminal.
35:14
I'm actually kind of a shitty Vim user.
35:16
So like how do I build a thing for people like me
35:18
that um you know are are going to be working in a terminal.
35:22
And I think just the delight is so important.
35:24
And I feel like at YC this is something you talk about a lot,
35:27
right?
35:27
It's like build a thing that people love.
35:29
If the product is useful but you don't fall in love with it,
35:31
that's not great.
35:32
Um so it kind of has to do both.
35:34
Designing for the terminal honestly has been hard, right?
35:36
It's like uh it's like 80 by 100 characters
35:38
or whatever. you have like 256 colors,
35:40
you have one font size, you don't have like mouse interactions,
35:44
there's all this stuff you can't do,
35:45
and there's all these very hard trade-offs.
35:47
So, like a little known thing, for example,
35:48
is you can actually enable mouse interactions in a terminal.
35:51
So, you can enable like clicking and stuff.
35:53
>> Oh, how do you do that in cloud code?
35:54
I've been trying to figure out how to do this.
35:56
>> We don't we don't have it in cloud code
35:58
because we actually prototyped it a few times
35:59
and it felt really bad
36:00
because the trade-off is you have to virtualize scrolling
36:02
and so there's all these weird trade-offs
36:05
because like the way terminals work is like there's no DOM,
36:07
right?
36:07
It's like there's like anti- escape codes
36:09
and these kind of weird organically evolved specs since like the 1960s
36:13
or whatever.
36:14
>> Yeah.
36:14
It feels like BBS's.
36:15
It's like a BBS door game.
36:17
>> Yeah.
36:17
>> Oh my god.
36:18
>> That's like that's like a great compliment.
36:20
Yeah.
36:20
Yeah.
36:20
Like it should feel like you're discovering >> Lord of the Red Dragon.
36:23
It's fantastic.
36:24
Oh my god.
36:24
>> Yeah.
36:25
But we have we've had to just like discover all these kind of UX
36:28
principles for building the terminal cuz no one really writes about this stuff.
36:31
And if you look at the big terminal apps of, you know,
36:34
like the 80s or 90s or 2000s or whatever,
36:36
they use like ed curses
36:37
and they have all these like windows
36:38
and things like this.
36:39
And it just looks kind of like janky by modern standards.
36:42
It just looks too heavy and complicated.
36:44
And so we had to like reinvent a lot.
36:46
And you know, for example, something like the terminal spinner,
36:49
like just like the spinner words,
36:50
it's gone through probably I want to say like 50 maybe 100 iterations at
36:55
this point.
36:55
And probably 80% of those didn't ship.
36:58
So we tried it, it didn't feel good,
36:59
move on to the next one. try it, didn't feel good,
37:01
move on to the next one.
37:02
Uh, and this was like sort of one of the amazing things about quad
37:05
code,
37:05
right?
37:05
Is like you can write these prototypes
37:07
and you can just do like 20 prototypes back to back,
37:10
see which one you like,
37:10
and then ship that and the whole thing takes maybe a couple hours.
37:13
>> Whereas in the past,
37:14
what you would have had to do is like wen to use origami
37:16
or framer or something like this.
37:17
You built like maybe three prototypes, it took like two weeks.
37:20
It just took much much longer.
37:22
>> And so we have this luxury of we have to discover this new thing.
37:25
We have to build a thing.
37:26
We don't know what the right endpoint is,
37:28
but we can iterate there
37:29
so quickly and that's what makes it really easy
37:32
and that's what lets us build a product that's like joyous
37:34
and that people like to use.
37:36
>> Boris, you had other advice for for builders
37:39
and we kept interrupting you
37:40
because we have so many questions,
37:42
but >> I would say um so okay,
37:45
so maybe two pieces of advice
37:47
that are kind of weird
37:48
because it's like about building for the model.
37:50
So one is uh don't build for the model of today,
37:53
build for the model of 6 months from now.
37:54
This is like sort of weird, right?
37:55
Because like you can't find PMF if the product doesn't work.
37:58
But actually this is the thing
37:58
that you should do
37:59
because otherwise what will happen is you spend a bunch of work you find
38:02
PMF for the product right now
38:03
and then you're just going to get leaprogged by someone else um
38:06
because they're building for the next model
38:08
and a new model comes out every few months.
38:10
Use the model, feel out the boundary of what it can do
38:12
and then build for the model
38:13
that you think will be the model maybe 6 months from now.
38:16
I think the second thing is um you know actually in the in the
38:19
quad code where in the quad code area where we sit we have a
38:21
framed copy of the bitter lesson on the wall.
38:24
Um and this is this like rich sutton uh I like everyone should read
38:27
it if if you haven't uh
38:29
and the idea is the more general model will always be the more specific
38:34
model and there's a lot of corlaries to this
38:36
but essentially what it boils down to is never bet against the model.
38:40
Uh, and so this is just like a thing to
38:42
that that we always think about where we could build a feature into cloud
38:46
code.
38:46
We could make it better as a product and we call this scaffolding.
38:49
That's all this code that's not the model itself.
38:51
But we could also just wait like a couple months
38:52
and the model can probably just do the thing instead.
38:56
Um, and there's always this trade-off, right?
38:57
It's like engineering work now
38:59
and you can kind of extend the capability a little bit,
39:01
maybe 10 20% or whatever in whatever domain on this like, you know,
39:04
like the spider chart of what you're trying to extend.
39:07
Um, or you can just wait and the next model will do it.
39:09
So just always always think in terms of this trade-off where where do you
39:12
actually want to invest
39:13
and assume that whatever the scaffolding is it's just tech.
39:16
>> How often do you rewrite the code ways of uh clock code is
39:20
every six months with this with this >> is there scaffolding
39:23
that you've deleted because you don't need it anymore
39:25
because the model just improved.
39:26
>> Oh so much.
39:27
Yeah.
39:28
Like all of quad code has just been written
39:29
and rewritten and rewritten
39:30
and rewritten over and over
39:31
and over.
39:32
We unhip tools every couple weeks.
39:34
We add new tools every couple weeks.
39:37
There's no part of quad code that was around six months ago.
39:39
It's just constantly rewritten.
39:41
>> Would you say most of the code base for current cloud code is
39:44
only say 80% of it is only less than a couple months old.
39:48
>> Yeah, definitely.
39:49
It might it might even be like less than Yeah,
39:51
maybe like a couple months.
39:52
That that feels about right.
39:53
>> So it's like the life cycle of code now.
39:55
That's another alpha is expecting it to be the shelf life to be just
39:58
couple months.
39:59
>> Yeah.
39:59
>> For the best founders.
40:00
>> Do you see uh Steve Yaggi's uh post about how awesome working at
40:04
Anthropic is?
40:05
And I think there's a line in there
40:06
that says that an anthropic engineer uh currently averages 1,000x more productivity than a
40:14
Google engineer at Google's peak
40:16
which is really an insane number honestly like 1,000x like you know we're 3
40:22
years ago we were still talking about 10x engineers now we're talking about 1000x
40:26
on top of a Google engineer in the prime like this is unbelievable honestly.
40:30
Yeah, I mean internally if you if you look at like technical employees,
40:33
they all use quad code every day.
40:35
Um, and even non-technical employees,
40:37
I think like half the sales team uses quad code.
40:39
Um, they they've started switching to co-work
40:41
because it's a little easier to use.
40:42
It has like a VM, so it's a little bit safer.
40:44
But yeah, we actually we just pulled a stat
40:46
and the I think the team doubled in size last year,
40:50
but productivity per engineer grew something like 70%.
40:53
>> It's measured by >> just like the simplest stupidest measure, pull requests.
40:57
Um, but we also kind of cross check
40:58
that against like commits
40:59
and like uh the lifetime of commits
41:01
and things like this.
41:02
And since quad code came out, productivity per engineer at anthropic has grown 150%.
41:07
>> Oh my god.
41:07
>> Um, and this is crazy
41:08
because I in my old life I was responsible for code quality at Meta.
41:13
>> Um, and I was responsible for the quality of all of our code
41:15
bases across every product across like you know Facebook,
41:18
Instagram, WhatsApp, whatever.
41:20
>> And one of the things that the team worked on was improving productivity.
41:24
And back then seeing a gain of something like 2% in productivity
41:27
that was like a year of work by hundreds of people.
41:30
And so this like 100% this is just like unheard of just completely unheard
41:34
of.
41:35
>> What drove you to come over to Anthropic?
41:37
I mean basically as a builder you could go anywhere.
41:40
What was the moment
41:40
that made you say like actually this is the set of people
41:43
or this is the approach.
41:45
I was living in rural Japan
41:47
and I was opening up Hacker News every morning
41:50
and I was reading the news
41:51
and uh it was all it just started to be like AI stuff at
41:55
some point and uh I started to use some of these early products
41:58
and uh I remember like the first couple times
42:01
that I used it I was just like it just took my breath away.
42:04
That was like very cheesy to say,
42:05
but that was actually that was actually the feeling.
42:06
Like it was just like it was amazing like
42:08
as a as as a builder,
42:10
I've just never kind of felt felt this feeling like using these very very
42:13
early products.
42:14
That was like in the quad 2 days or you know something like that.
42:17
And so I I just talking started talking to friends at Labs um just
42:20
to kind of see what was going on.
42:22
Um and uh I met Ben man who's one of the founders at uh
42:27
at Anthropic and uh he just immediately won me over.
42:30
Um and as soon
42:31
as I met kind of the rest of the team at an it just
42:33
won me over and I think I think probably in two ways.
42:35
So one is it operates as a research lab.
42:38
Um so the product was teeny teeny tiny.
42:41
It's really all about building a safe model.
42:44
That's all that matters.
42:45
Um and so this idea of just being very close to the model
42:47
and being very close to development
42:49
and being not the most important thing
42:52
because the product isn't anymore.
42:53
It's just the model is the thing that's the most important.
42:56
Um that really resonated with me after building product for many years.
42:59
And then the second thing was just how missiondriven it is.
43:03
Um like I'm I'm a huge sci-fi reader.
43:05
My bookshelf is just like filled with sci-fi.
43:07
And so like I just know how bad this can go.
43:10
>> And when I kind of think about what's going to happen this year,
43:12
it you know it's going to be totally insane.
43:15
And in the worst case it can go very very bad.
43:17
>> Um and so I just wanted to be at a place
43:19
that really understood that
43:20
and kind of really internalized
43:21
that.
43:22
And at Ant, you know,
43:23
like if you overhear conversations in the lunchroom or in the hallway,
43:26
people are talking about AI safety. this is really the thing
43:29
that everyone cares about more than anything.
43:32
Um, and so I just wanted to be in a place like that.
43:34
I I know I know for me personally the mission is just so important.
43:37
>> What is gonna happen this year?
43:40
>> Okay.
43:40
So if you think back like six months ago
43:43
and uh kind of what are the predictions
43:44
that people are making?
43:45
So Daario predicted that 90% of the code at Anthropic would be would be
43:49
written by Quad.
43:50
This is true.
43:51
Um for me personally it's been 100% for like since Opus 4.5.
43:56
Um I just I uninstalled my IDE.
43:58
I don't edit a single line of code by hand.
44:00
It's just 100% quad code and Opus.
44:03
Um and you know I land you know like 20 PR a day every
44:06
day.
44:06
If you look at Enthropic overall it ranges between like 70 to 90% uh
44:11
you know depending on the team.
44:12
For a lot of teams it's also like 100% for a lot of people
44:14
it's 100%.
44:15
And I remember making this prediction back in May
44:18
when we ged cloud code
44:20
that you wouldn't need an ID to code anymore.
44:22
Uh and it was totally crazy to say.
44:25
I feel like people in the audience gasped >>
44:27
because it was such like a silly prediction at the time.
44:29
But really all it is is like you just like trace the you know
44:31
the exponential >> and this is just like
44:33
so deep in you know the DNA at cuz like you know three of
44:37
our founders were co-authors of the scaling laws paper they kind of they saw
44:40
this very early and
44:41
so this is just like tracing the exponential this is what's going to happen
44:44
and yes that happened.
44:45
So continuing to trace the exponential I think what will happen is coding will
44:49
be generally solved for everyone.
44:51
Um, and I think today coding is practically solved, you know,
44:53
for me and I think it'll be the case for everyone.
44:56
Um, you know, regardless of domain,
44:58
I think we're going to start to see the title software engineer go away.
45:01
And I think it's just going to be maybe builder, maybe product manager,
45:04
maybe we'll keep the title as kind of a vestigial thing,
45:07
but the work that people do, it's not just going to be coding.
45:10
It's software engineers are also going to be writing specs.
45:12
They're going to be talking to users. like this thing
45:14
that we're starting to see right now in our team where engineers are very
45:18
much generalists and every single function on our team codes like our PM's code,
45:22
our designers code, our EM codes,
45:25
our um like everyone our our finance guy codes like everyone on our team
45:29
codes.
45:30
We're going to start to see this everywhere.
45:31
So this is sort of uh this is kind of like the lower bound
45:34
if we just continue the trend.
45:36
The upper bound I think is a lot scarier.
45:38
Um, and this is something like, you know, we hit ASL4.
45:42
Um, and this, you know, at anthropic, we talked about these safety levels.
45:45
ASL3 is where the models are right now.
45:47
ASL4 is the model is recursively self-improving.
45:50
Um, and so if this happens, essentially,
45:51
we have to meet a bunch of criteria before we can release a model.
45:54
And so the the extreme is that, you know,
45:56
this happens um or there's some kind of catastrophic misuse like people are using
46:00
the model to design bioiruses,
46:02
design zero days, stuff like this.
46:04
Um, and this is something
46:05
that we're really really actively working on
46:07
so that doesn't happen.
46:08
I think uh it's just been honestly it's just been like
46:11
so exciting and humbling like seeing how people are using quad code like uh
46:15
you know I just wanted to build a cool thing
46:16
and it ended up being really useful uh
46:19
and that was so surprising
46:20
and so exciting.
46:21
>> My impression from Twitter
46:22
or just the outside is basically everyone went away over the holidays
46:25
and then like found out about Claude code
46:27
and it's just been crazy ever since.
46:29
Is that how it was for you at like internally?
46:32
Did you were you having like a nice Christmas break
46:35
and then came back
46:36
and like what happened?
46:37
Well, actually for all of December, I was traveling around.
46:39
Uh, and I I took a coding vacation.
46:41
So, we were kind of traveling around
46:42
and I was just like coding every day.
46:44
So, that was really nice.
46:45
Uh, and then I also started to use Twitter at the time cuz like
46:47
I I worked on Threads back
46:49
then way back when.
46:50
So, I've been a Threads user for a while.
46:52
So, I just like tried to see kind of like other platforms where people
46:55
are.
46:55
Yeah.
46:55
I think for a lot of people they kind of discovered
46:57
that was the moment where they discovered Opus 4.5.
46:59
I kind of already knew.
47:01
>> Mhm.
47:01
>> Uh, and internally quad code's just been on this like exponential tear for
47:04
many many months now.
47:05
So that just like it it became even more steep.
47:09
That's what we saw.
47:10
And if you look at cloud code now, you know,
47:12
there was some stat from Mercury
47:13
that like 70% of startups are you know choosing cloud
47:16
as their model of choice.
47:18
There was some other stat from like semi analysis
47:20
that 4% of all public commits are made by cloud code. um like of
47:24
all code written everywhere.
47:26
All the companies, you know,
47:27
use squad code from like the biggest companies to kind of, you know,
47:30
smallest startups, you know,
47:31
like it it wrote it it plotted the course for Perseverance like for like
47:35
the Mars rover.
47:36
This is just like this is the coolest thing for me.
47:38
And we like we even printed posters cuz the team was like, "Wow,
47:41
this is just like so cool that NASA chooses to use this thing."
47:43
So, yeah, it's just like it's humbling.
47:45
Um but it also feels like the very beginning.
47:47
What's the sort of interaction between uh claude code
47:50
and then co-work like you know was it a fork of cla code?
47:54
Was it like you had cla code look at the cloud code
47:58
and say let's make a new spec for nontechnical people
48:02
that you know keeps all the lessons
48:04
and then you know it sort of went off for a couple days
48:06
and did that.
48:07
What's the genesis of that and you know where do you think that goes?
48:11
>> This is going to be like my fifth time using the word wait
48:13
and demand.
48:15
It was just that I mean like we we were looking at Twitter
48:17
and there was like
48:17
that one guy that was using quad code to like monitor his tomato plants.
48:21
>> Mhm.
48:21
>> Uh there was like this other person
48:22
that was using it to like recover wedding photos off of a corrupted hard
48:25
drive.
48:26
There were people that using it for like uh for finance.
48:29
When we looked internally at anthropic,
48:32
every designer is using it all the entire finance team at this point is
48:35
using it.
48:35
The entire data science team is using it not for coding.
48:38
People are jumping over hoops to install a thing in the terminal
48:40
so that they could use this.
48:41
So we knew for a
48:42
while that we wanted to build something
48:43
and so we're experimenting with a bunch of different ideas
48:46
and the thing that kind of took off was just you know a little
48:49
cloud code wrapper in a guey in the desktop app
48:51
and that's all it is.
48:52
It's just quad code under the hood.
48:53
It's the same agent.
48:54
>> Oh wow.
48:55
>> Um and uh Felix and the team and Felix was early Electron contributor.
48:59
He kind of knows
49:00
that stack really well
49:01
and he was hacking on various ideas
49:03
and uh they they built it in I think something like 10 days.
49:06
It was it was just like 100% written by quad code.
49:09
Uh and it just felt ready to release.
49:11
There was a lot of stuff that we had to build for nontechnical users.
49:15
So it's a little bit different than a technical audience.
49:18
Uh it runs in a all the code runs in a virtual machine.
49:21
Uh there's a lot of delete uh protections for deletion and things like this.
49:25
There's a lot of permission prompting and kind of other guardrails for users.
49:29
Um yeah, it was honestly pretty obvious.
49:32
Boris, thank you so much for making something
49:35
that uh is taking away all my sleep,
49:37
but in return, it's making me feel creator mode again,
49:41
sort of founder mode again.
49:42
It's been an exhilarating 3 weeks.
49:44
I like can't believe I waited
49:46
that long since November to actually get into it.
49:48
Thank you so much for being with us.
49:50
Thank you for building what you're building.
49:52
>> Yeah, thanks for having me.
49:53
And uh send bugs.
49:56
>> Sounds good.
50:01
Come on now.
Like
Share
Y Combinator
View all →
C1
Business
Karaoke
22:36
OpenClaw Creator: Why 80% Of Apps Will Disappear
Y Combinator
5
C1
Business
Karaoke
7:51
The New Way To Build A Startup
Y Combinator
C1
Business
Karaoke
13:07
Inside The Startup Reinventing The $6 Trillion Chemical Manufacturing Industry
Y Combinator
C1
Business
Karaoke
40:57
Demis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough
Y Combinator
C1
Business
Karaoke
10:27
How To Build A Company With AI From The Ground Up
Y Combinator
C1
Business
Karaoke
53:20
The Future Of Brain-Computer Interfaces
Y Combinator
C1
Business
Karaoke
21:49
How to Make Claude Code Your AI Engineering Team
Y Combinator
1
C1
Business
Karaoke
16:35
Conductor CEO Charlie Holtz Walks Us Through His AI Coding Setup
Y Combinator
1
Suggested videos
C1
Business
Karaoke
5:49
You Need to Be Bored. Here's Why.
Harvard Business Review
7
C1
Business
Karaoke
4:28
Lean Into Imposter Syndrome, Don't Give In to It
Harvard Business Review
C1
Business
6:30
Identity Crisis: Why Defining Yourself by Your Career Is a Problem
Harvard Business Review
C1
Business
Karaoke
22:36
OpenClaw Creator: Why 80% Of Apps Will Disappear
Y Combinator
5
C1
Business
Karaoke
6:36
Fighting Workaholism: You Are Not a Success Machine
Harvard Business Review
C1
Business
Karaoke
2:47
Can Work Make You Happy? Should It?
Harvard Business Review
C1
Business
Karaoke
7:51
The New Way To Build A Startup
Y Combinator
C1
Business
Karaoke
13:07
Inside The Startup Reinventing The $6 Trillion Chemical Manufacturing Industry
Y Combinator
