TryShadowing
Shadow YouTube. Nói tiếng Anh.
Trang chủ
Khám phá
Chép chính tả
NEW
Thư viện
vi
0
ngày
Đăng nhập
OpenClaw: The Viral AI Agen… — Lex Fridman luyện shadowing | TryShadowing
TryShadowing
Shadow YouTube. Nói tiếng Anh.
Trang chủ
Khám phá
Chép chính tả
NEW
Thư viện
vi
0
ngày
Đăng nhập
Trang chủ
Khám phá
Lex Fridman
OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491
OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491
Lex Fridman
·
3:15:52 · 12 thg 2, 2026
Bắt đầu học
0:00
0:00
Ghi âm
×1
1x
VI
EN
JA
KO
ZH
FR
PT
TH
IT
DE
IPA
Chấm điểm phát âm chưa hỗ trợ trên trình duyệt này — bạn vẫn ghi âm & nghe lại được.
I
watched
my
agent
happily
click
the
I'm
not
a
robot
button.
Đang dịch…
Bật Ghi âm để được thu giọng và chấm điểm
Thông minh
Karaoke
Câu gốc
Câu
1
/4161
0:00
I watched my agent happily click the I'm not a robot button.
0:03
I made the agent very aware.
0:05
Like it knows what its source code is.
0:07
It understands the how it
0:09
sits and runs in its own harness.
0:12
It knows where the documentation is.
0:14
It knows which model it runs.
0:16
It understands its own system.
0:17
That made it very easy for an agent to
0:20
Oh, you don't like anything?
0:21
You just prompt it into existence.
0:23
And then the agent would just modify its own software.
0:26
People talk about self-modifying software.
0:27
I just built it.
0:29
I actually think web coding is a slur.
0:31
You prefer agentic engineering.
0:33
>> Yeah, I always tell people I'd I do agentic engineering
0:35
and then maybe after 3:00 a.m.
0:37
I switch to web coding
0:38
and then I have regrets on the next day.
0:40
Wow, walk of shame.
0:42
Yeah, you just have to clean up and like
0:43
fix your >> We've all been there.
0:46
I used to write really long prompts.
0:49
And by writing I mean I don't write.
0:51
I I I talk.
0:53
You know, this these hands are like too
0:55
too precious for writing now.
0:56
I just I just
0:58
use bespoke prompts to build my software.
1:00
So, you for real with all those terminals are using voice. Yeah.
1:05
I used to do it very extensively.
1:08
To the point where there was a period where I lost my voice.
1:12
>> I mean, I have to ask you, just curious.
1:15
I I know you've probably gotten
1:17
huge offers from major companies.
1:21
Can you speak to who you're considering working with?
1:30
The following is a conversation
1:31
with Peter Steinberger, creator
1:34
of Open Claw, formerly known
1:37
as Mold Bot, Clawed Bot, Clawdis,
1:39
Clawd spelled with a W
1:41
as in lobster claw,
1:42
not to be confused with Claude,
1:44
the AI model from Anthropic
1:46
spelled with a U.
1:48
In fact, this confusion
1:49
is the reason Anthropic
1:51
kindly asked Peter to change the name to Open Claw.
1:55
So, what is Open Claw?
1:58
It's an open-source AI agent
1:59
that taken over the tech world in a matter of days,
2:02
exploding in popularity, reaching over
2:07
stars on GitHub and
2:09
spawning the social network Mold Bot
2:11
where AI agents post
2:12
manifestos and debate consciousness,
2:15
creating a mix of excitement
2:17
and fear in the general public
2:19
in a kind of AI
2:21
a mix of click bait fear mongering
2:24
and genuine fully justifiable
2:26
concern about the role of AI
2:27
in our digital interconnected human world.
2:31
Open Claw, as its tagline states,
2:33
is the AI that actually does things.
2:36
It's an autonomous AI assistant
2:38
that lives in your computer,
2:39
has access to all of your stuff if you let it,
2:42
talks to you through Telegram,
2:44
WhatsApp, Signal, iMessage, and whatever else messaging client,
2:47
uses whatever AI model you like,
2:50
including Claude Opus 4.6
2:53
and GPT 5.3 all to do stuff for you.
2:59
Many people are calling this one of the biggest moments in the recent history
3:02
of AI since the launch of ChatGPT in November 2022.
3:07
The ingredients for this kind of AI agent
3:09
were all there, but putting it all together
3:11
in a system that definitively
3:13
takes a step forward
3:15
over the line from language to agency,
3:17
from ideas to actions
3:19
in a way that created a useful assistant
3:21
that feels like one who gets you
3:23
and learns from you
3:25
in an open-source community-driven
3:26
way is the reason Open Claw
3:29
took the internet by
3:31
Its power, in large part,
3:33
comes from the fact
3:34
that you can give it access to all of your stuff
3:36
and give it permission
3:38
to do anything with that stuff
3:39
in order to be useful to you.
3:42
This is very powerful,
3:44
but it is also
3:46
Open Claw represents freedom,
3:48
but with freedom comes responsibility.
3:51
With it, you can own and have control
3:53
over your data, but precisely because
3:56
you have this control,
3:57
you also have the responsibility
3:59
to protect from cybersecurity
4:01
threats of various kinds.
4:03
There are great ways to protect yourself,
4:05
but the threats and vulnerabilities are out there.
4:08
Again, a powerful AI agent
4:10
with system-level access is a security minefield,
4:14
but it also represents the future
4:16
because when done well and securely,
4:18
it can be extremely
4:19
useful to each of us humans
4:22
as a personal We discuss all of this with Peter
4:26
and also discuss his big picture programming
4:29
and entrepreneurship life story,
4:31
which I think is truly inspiring.
4:34
He spent 13 years building PSPDFKit,
4:37
which is a software
4:38
used on a billion
4:40
He sold it and
4:42
for a brief time fell out of love with programming,
4:45
vanished for 3 years,
4:47
and then came back,
4:48
rediscovered his love for programming,
4:51
and built in a very short time an open-source
4:53
AI agent that took the internet by storm.
4:57
He is in many ways a symbol
4:59
of the AI revolution
5:00
happening in the programming
5:02
There was the ChatGPT
5:03
moment in 2022, the Deep Seek moment
5:06
in 2025, and now
5:09
in '26 we're living through the Open Claw moment,
5:13
the age of the lobster,
5:15
the start of the agentic AI revolution.
5:18
What a time to be alive.
5:21
This is the Lex Fridman podcast.
5:22
To support it, please check out our sponsors
5:24
in the description where you can also find links to contact me, ask questions,
5:29
give feedback, and so on.
5:31
And now, dear friends,
5:33
here's Peter the one and only, the Clawdfather.
5:39
Actually, Benjamin predicted in this tweet,
5:42
"The following is a conversation
5:43
with Claude, a respected crustacean."
5:46
There's a hilarious-looking picture
5:48
of a lobster in a suit.
5:50
So, I think the prophecy has been fulfilled.
5:54
>> Let's go to this moment
5:56
when you built a prototype in 1 hour.
5:59
That was the early version of Open Claw.
6:01
I think this story is really inspiring to a lot of people because
6:05
this prototype led to something
6:07
that just took the internet by storm
6:09
and became the fastest-growing
6:11
repository in GitHub history with
6:14
now over 175,000 stars.
6:16
So, what was the story of the 1-hour prototype?
6:19
You know, I wanted that since April.
6:23
A personal assistant, AI personal assistant.
6:25
>> Yeah, and I I played around with some other things
6:28
like even stuff to
6:32
get some of my WhatsApp
6:34
and I could just run queries on it.
6:35
That was back when I
6:37
we had GPT 4.1
6:39
and it was the 1 million context window.
6:42
And I I pulled in all the data and then I asked him questions
6:47
like, "What makes this friendship meaningful?" Mhm.
6:50
And I got some
6:52
some really profound results.
6:56
Like I sent it to my friends and they all got like teary eyes.
6:59
So, there's something there. Yeah.
7:02
But then I I thought
7:04
all the labs will will will work on
7:07
So, I I moved on to other things and I was still very much
7:10
in my early days of experimenting and playing.
7:14
You know, you have to
7:16
That's how you learn.
7:17
You just like you do stuff and you play.
7:22
time flew by and it was November.
7:24
I wanted to make sure that
7:26
the thing I started is actually happening.
7:28
I was annoyed that it didn't exist.
7:29
So, I just prompted it into existence.
7:36
>> I mean, that's the beginning of the hero's journey of the entrepreneur,
7:40
And you've even with your
7:42
original story with PSPDFKit,
7:44
it's like, "Why does this not exist?
7:46
Let me build it."
7:47
And again, here's diff- whole different realm,
7:50
but similar maybe spirit.
7:52
Yes, I had this problem.
7:53
I tried to show PDF on an iPad,
7:55
which should not be hard.
7:56
>> This is like 15
7:57
years ago, something like that. Yeah.
7:59
Like the most the most random thing ever.
8:02
Suddenly, I had this problem.
8:04
And I I wanted to help a friend.
8:06
And that was that was was not like nothing existed, but it was just not good.
8:10
I'm like, like I tried it and it was like very mad.
8:15
I'm like, "I can do this better."
8:17
By the way, for people who don't know,
8:18
this led to the development
8:20
of PSPDFKit that's used on a billion devices.
8:23
So, the it turns out that it's pretty useful to be able to open a PDF.
8:28
You could also make the joke that I'm really bad at naming.
8:31
>> Like name them the five on the current project.
8:34
And even PSPDF doesn't really
8:37
roll from the tongue.
8:39
Anyway, she said, "Screw it.
8:42
Why don't I uh do it?"
8:44
So, what was the what was the prototype?
8:46
What was the thing that you
8:47
What was the magical
8:48
thing that you built in a short amount of time that you're like, "This
8:51
might actually work as an agent
8:53
where I talk to it and it does things."
8:55
That was like one of my projects before
8:58
already did something where I could bring my terminals onto the web.
9:03
And then I could like interact with them, but they also would be terminals on my Mac. Mhm.
9:07
Vibe Tunnel, which was like
9:09
a a weekend hack project.
9:12
That was still very early and
9:15
I was cloud code times.
9:17
You know, you got a dopamine hit when you got something right.
9:20
And now I get like mad when you get something wrong.
9:22
And you had a really great,
9:24
not to dig a tangent, but a great blog post describing
9:26
that you converted Vibe
9:29
you vibe coded Vibe Tunnel from TypeScript
9:31
into Zig of all programming languages
9:34
with a single prompt.
9:36
One prompt, one shot.
9:38
Convert the entire code base into Zig. Yeah.
9:41
That was this one thing where
9:43
part of the architecture
9:45
>> was took too much memory.
9:48
Every terminal used like a node.
9:52
Um and I wanted to change it to Rust and
9:56
I mean, I can do it.
9:57
I can manually figure it all out,
10:01
but all my automated attempts
10:05
failed And then I revisited
10:10
about four or five months later
10:12
and I'm like, "Okay, now let's use something even more
10:16
And I and I just typed,
10:18
"Convert this and this part to sig."
10:20
And then let Codex run
10:23
And it basically got it right.
10:25
There was one little detail that I had to like
10:27
modify afterwards, but it just ran for
10:32
overnight, like 6 hours, and just did the thing.
10:35
And it's like this is just
10:39
So, that's on the LLM programming side
10:42
refactoring, but uh back to the actual story of the uh of the prototype.
10:48
So, how did Vitale connect to the first prototype
10:50
where your like agents can actually work?
10:52
Well, that was still very limited.
10:54
You know, like I had this one experiment with WhatsApp.
10:58
Then I had this experiment,
10:59
and both felt like
11:01
not the right answer.
11:03
And then my third try was literally just hooking up
11:08
WhatsApp to Cloud Code.
11:11
One shot, the CLI. Message comes in.
11:14
I call the CLI with minus P.
11:17
It does its magic.
11:18
I get the string back and I send it back to WhatsApp and I
11:21
I built this in 1 hour.
11:24
I have >> already felt really cool.
11:26
It's like, "Wow, I could I can like
11:27
talk to my computer, right?"
11:29
This That that was that was cool, but I I wanted
11:33
images cuz I I often use images when I prompt.
11:36
I think it's such a
11:37
such an efficient way to give the agent more context and they're really good
11:40
at figuring out what I mean even if it's like a a weird cropped out screenshot.
11:45
Um so, I used it a lot and I wanted to do it in WhatsApp as well.
11:49
Also, like you know, just you run around, you see like
11:52
a post of an event,
11:54
you just make a
11:56
figure out if I have time there, if this is good, if my friends
11:59
are maybe up for that.
12:00
It's like images seem important.
12:03
I worked a few
12:05
It took me a few more hours to actually get that right.
12:09
and then it was just
12:12
I I used it a lot.
12:14
And funny enough, that was
12:17
just before I went on a trip to Marrakech
12:20
with my friends for birthday
12:23
And there it was even better because
12:26
internet shaky, but WhatsApp
12:28
just works, you know?
12:29
It's like doesn't matter.
12:30
You have like edge, it still works.
12:32
WhatsApp is just it's just made really well.
12:37
So, I ended up using it a lot.
12:41
Translate this for me,
12:42
explain me to find me places.
12:44
Like you're just having
12:46
a clanker doing having Google for you.
12:49
That was I basically was still nothing built,
12:51
but it still could do so much.
12:53
So, if we talk about the full journey
12:55
that's happening there with the agent,
12:58
you're just sending on this very thin line
13:01
WhatsApp message via CLI,
13:04
it's going to Cloud Code, and Cloud Code is doing all kinds of
13:09
work and coming back to you with a thin message. Yeah.
13:13
It was slow because
13:15
every time I boot up the CLI,
13:17
but it it was really cool already.
13:20
And it could just use all the things that I already had built.
13:23
And I built like a whole bunch of
13:25
CLI stuff over the months.
13:27
So, it it really powerful.
13:31
There's something magical about that experience
13:33
that's hard to put into words.
13:36
Being able to use a chat client
13:40
to talk to an agent
13:43
versus like sitting behind a computer
13:45
and like I don't know, using
13:47
cursor or even using Cloud Code CLI
13:49
in the terminal, it's a different experience than being able to sit back and talk to it.
13:53
I mean, it seems like a trivial step,
13:56
but in some in some sense it's
13:58
uh it's like a phase shift
14:01
in the integration of AI into your life
14:03
and how it feels, right? >> Yeah. Yeah.
14:05
I I read this tweet this morning where someone said,
14:08
"Oh, there's no magic in it.
14:10
It's just it does this and this and this and this and this and this and this."
14:15
And it it almost feels like a hobby just as Cursor or Perplexity.
14:20
And I'm like, "What if that's a hobby?
14:22
That's kind of a compliment, you know?"
14:25
>> They're like, "You're not doing too bad.
14:29
thank you, I guess."
14:32
Yes, I mean, isn't isn't isn't magic often just like you take a lot
14:36
of things that are already there,
14:39
but bring them together in new ways?
14:41
Like I I don't
14:42
There's no Yeah, maybe there's no magic in there, but
14:46
sometimes just rearranging things and like adding a few new ideas is all the
14:49
magic that you need.
14:51
Yeah, it's really hard to convert into words what is
14:54
what is magic about a thing.
14:56
If you look at the
14:57
the scrolling on an iPhone,
14:59
why is that so pleasant?
15:01
There's a lot of elements about that interface that makes it incredibly
15:04
pleasant that is fundamental
15:05
to the experience of using a smartphone.
15:07
And it's like, "Okay,
15:09
all the components were there. Scrolling was there. Everything was there."
15:13
And nobody did it. Yep.
15:14
And then afterwards it felt so obvious. It's so >> Right? Yeah.
15:17
But still um you know, the moment where
15:22
it it blew my mind was when
15:25
when I used it a lot and then at some point I just sent it a message.
15:28
And and then a typing indicator appeared
15:33
and I'm like, "Wait, I didn't build that.
15:36
Uh it's only It only has image support.
15:38
So, what is it even doing?"
15:40
And then it would just
15:41
What was the thing you sent it?
15:43
Oh, just a random question like, "Hey,
15:45
what about this and this restaurant, you know?"
15:47
Um cuz we were just running around and and
15:50
checking out the city.
15:52
So, that's why I
15:53
I didn't didn't even think when I used it because sometimes when you're in
15:56
a hurry, typing is
15:58
So, you are you did an audio message? >> Yeah.
16:01
And it was just it just worked and I'm like,
16:03
And it's not supposed to work because it
16:05
you didn't give it that No, >> capability.
16:08
I literally wrote, "How the did you do that?"
16:11
And it was like, "Yeah,
16:12
the metal did the following.
16:13
You sent me a message,
16:15
it only only was a file
16:17
and no file ending.
16:18
So, it I checked out the header of the file
16:22
and it found that it was like Opus.
16:24
So, I used FFmpeg to convert it.
16:26
And then I wanted to use Whisper, but it didn't have it installed.
16:29
But then I found the OpenAI key
16:32
and just used curl
16:34
to send the file to to OpenAI to translate.
16:37
And here I am."
16:39
And I just looked at the message and I'm "Oh, wow."
16:42
You didn't teach it any of those things and the agent just figured it
16:45
out that it has to do all those conversions, the translation,
16:47
to figure out the API,
16:49
it figured out which program to use, all those kinds of things.
16:52
And you were just absentmindedly
16:54
just sending an audio message.
16:55
And like he was so clever even because he would have gone the Whisper
16:58
local path, would have had to download a model, it would have been too slow.
17:02
So, like there's so much world knowledge in there, so much creative problem-solving.
17:07
A lot of it I think mapped from
17:09
if you get really good at coding, that means you have to be really
17:12
good at general-purpose program solving.
17:15
So, that's a skill, right?
17:16
And that just maps into other domains.
17:18
So, it had the problem of like,
17:21
"What is this file with no file ending?
17:22
Let's figure it out."
17:24
Um and that's where it
17:27
kind of clicked for me.
17:28
It was like I was like very impressed.
17:32
somebody sent a pull request for Discord support.
17:35
And I'm like, "This is a WhatsApp relay.
17:37
That doesn't doesn't fit at all."
17:40
At that time it was called War Relay.
17:43
And so, I debated
17:45
with me like, >> "Do I want that
17:48
or not want that?"
17:51
>> And then I thought, "Well, maybe
17:53
maybe I do that cuz
17:56
that could be a cool way to show people.
17:58
Cuz I so far did it in WhatsApp with like groups, you know, but
18:03
>> don't really want to
18:04
give my phone number to every internet stranger. Yeah.
18:07
Um journalists managed to do that anyhow now, so that's a different story.
18:14
Uh so, I I merged it
18:16
from Shadow who helped me a lot with the whole project. So, thank you.
18:22
>> And and I put my
18:24
my bot in there on Discord.
18:28
Yeah, no security cuz I didn't
18:30
I hadn't built sandboxing in yet.
18:32
>> I I just prompted it to like
18:35
only listen to me.
18:38
And then some people came and tried to hack it.
18:40
And I just I I just watched and I just kept working in the open, you know?
18:45
Like you used my agent
18:49
to build my agent harness
18:53
and to test like various stuff.
18:56
>> And that's very quickly when it clicked for people.
18:59
So, it's almost like
19:00
it needs to be experienced.
19:03
And from that time on, that was
19:05
January the 1st, I
19:08
I got my first
19:10
really influencer being a fan.
19:12
He did videos, The Kizu. I thank you.
19:16
And and from there on I saw
19:19
I saw it gaining up speed.
19:21
And at the same time my
19:23
my sleep cycle went shorter and shorter because I
19:27
I felt the storm coming.
19:30
And I just worked my ass off to get it to into a state
19:33
where it's kind of good.
19:38
There's a few components
19:38
we'll talk about how it all works, but basically
19:41
you're able to talk
19:42
to it using WhatsApp, Telegram, Discord.
19:46
So, that's the component that you have to get right. Yeah.
19:48
And then you have to
19:50
figure out the agentic loop, you have you have the gateway,
19:52
you have the harness,
19:53
you have all those components
19:54
that make it all just work nicely. Yeah.
19:57
It felt like Factorio times infinite. Right.
20:00
I I feel like I built my little >> my little playground.
20:04
Like I never had so much fun than building this project.
20:07
You know, like you have like oh
20:09
I go like level one agentic loop.
20:11
What can I do there?
20:11
How can I be smart at queuing messages?
20:14
How can I make it more human like?
20:16
Oh, then I just idea of
20:18
because the loop always
20:20
the agent always replies something, but you don't always want an agent to reply
20:24
something in a group chat.
20:25
So, I gave him this no reply token.
20:28
So, I gave him an option to shut up.
20:30
So, it it feels more natural. That's level two. Yeah, yeah, yeah.
20:35
On the on the Factorio.
20:36
On the agentic loop and then I go to memory, right?
20:39
You want them to like remember stuff.
20:40
So, maybe maybe the end
20:43
the ultimate boss is
20:44
continuous reinforcement learning, but I'm I'm like at I feel like I'm level two
20:48
or three with markdown files and a vector database.
20:52
And then you you can go to the level
20:54
community management, you can go to the website and marketing.
20:58
There's just so many hats that you have to have on.
21:01
Not even talking about native apps.
21:03
That's just like infinite
21:05
different levels and infinite
21:07
level ups you can do.
21:08
So, the whole time you're having fun.
21:10
We should say that for the most
21:12
throughout this whole process,
21:14
you're a one-man team.
21:16
There's people helping, but you're doing so much of the key core development. Yeah. And having fun?
21:23
You did in January 6,600 commits. Probably more.
21:28
I sometimes posted the meme I'm limited by the technology of my time.
21:31
I could do more if agents would be faster.
21:34
But we should say you're running multiple agents at the same time. Yeah.
21:38
Depending on how much I slept
21:39
and how difficult of the tasks I work on 4 and 10.
21:45
4 and 10 agents.
21:47
Uh there's so many possible directions speaking of Factorio
21:50
that we can go here, but one
21:52
uh big picture one
21:53
is why do you think
21:56
your work OpenClaw in this world
22:02
if you look at 2025,
22:04
so many startups, so many companies
22:06
were doing kind of agentic
22:07
type stuff or claiming
22:09
to and here OpenClaw
22:11
comes in and destroys everybody.
22:14
Like why did you win?
22:15
Because they all take themselves too serious. Yeah.
22:19
Like it's hard to compete
22:21
against someone who's just there to have fun. Yeah.
22:24
I wanted it to be fun.
22:25
I wanted it to be weird.
22:27
And if you see like all the
22:29
all the lobster stuff online,
22:31
um I think I I managed weird.
22:35
I And then for the longest time,
22:37
the the only way to install it was
22:41
get clone pmpm build pmpm gateway.
22:46
Like you clone it, you build it, you run it.
22:48
Um and then the
22:50
the agent I made the agent very aware.
22:53
Like it knows that it is
22:55
what its source code is.
22:57
It understands how it
22:59
sits and runs in its own harness.
23:02
It knows where the documentation is.
23:04
It knows which model it runs.
23:06
It knows if you turn on verbose
23:08
or or reasoning mode.
23:10
Like I I wanted to be more human like, so
23:14
it understands its own system.
23:15
That made it very easy for an agent to
23:18
oh, you don't like anything, you just
23:19
prompt it into existence
23:21
and then the agent would just modify its own
23:26
you know, we have people talk about self-modifying software.
23:28
I just built it.
23:30
And didn't even I didn't even
23:33
plan it so much.
23:34
It just Can you actually speak to that?
23:37
Cuz it's just fascinating.
23:39
So, you have this piece of software that's written TypeScript. Yeah.
23:43
That's able to via the agentic loop modify itself.
23:48
I what a moment
23:50
to be alive in the history
23:52
of humanity, in the history of
23:54
Here's a thing that's used
23:56
by a huge amount of people
23:58
to do incredibly powerful things in their lives.
24:01
And that very system
24:03
can rewrite itself, can modify itself.
24:07
Can you just like
24:08
speak to the power of that?
24:09
Like isn't that incredible?
24:11
Like when did you first
24:12
close the loop on that?
24:14
Oh, because that's how I built it as well.
24:16
You know, most of it is built by Codex,
24:18
but often times I
24:21
when I debug it
24:24
I use self-introspection so much to like hey, what tools do you see?
24:27
Can you call the tool yourself?
24:29
Oh, like what error do you see?
24:30
Read the source code, figure out what's the problem.
24:32
Like I just found it an incredibly fun way to
24:38
that the agent the very agent and software that you use
24:41
is used to debug itself.
24:43
So, that it felt just natural that everybody does that.
24:47
And that it led to so many
24:50
so many pull requests by people who never wrote software.
24:53
I mean it also did show that people never wrote software.
24:57
I call them prompt requests in the end.
25:00
But I don't want to like pull that down because
25:03
every time someone made the first pull request
25:06
is a win for our society, you know?
25:08
Like it like doesn't matter how how shitty it is.
25:11
You got to start somewhere.
25:13
So I know there's like this whole big movement of people complain about open
25:18
source and the quality of PRs and
25:20
a whole different level of problems,
25:21
but on a different level, I found it
25:26
I found it very meaningful
25:29
that that I built something that people love to think of so much that
25:33
they actually start to learn how open source works.
25:37
Yeah, you were the OpenClaw project
25:40
was a first pull request.
25:42
You were the first for so many. That is magical.
25:45
So many people that don't know how to
25:48
are taking their first step into the programming world with this.
25:51
Isn't that a step up for humanity? Isn't that cool? Creating builders. Yeah.
25:56
Like the bar to do that was so high
25:59
and like with agents
26:00
and with the right software, it just like went lower and lower. I don't know.
26:04
I was at a
26:07
at a I also
26:08
organized another type of meetup.
26:09
I call it I call
26:11
it Claude Code You can get the inspiration from.
26:17
Now I call it
26:18
Agents Anonymous for reasons. Agents Anonymous.
26:24
And That was so funny on so many levels.
26:28
I'm sorry, go ahead. Yeah.
26:30
And there was this one guy who who who talked to me he's like
26:32
I run this design
26:34
agency and we we never had custom software
26:37
and now I have like 25
26:39
little web services for various things that help me in my business.
26:44
And I don't even know
26:45
how they work, but they work.
26:50
And he was just like very happy that
26:52
my stuff solved some of his problems
26:54
and he was like curious enough that he actually came to like a
26:58
agentic meetup even though he's
27:02
he doesn't really know how software
27:04
Can we actually uh
27:06
rewind a little bit and uh tell the saga
27:09
of the name change?
27:10
First of all, it started out as W Relay. >> Yeah.
27:12
And then it went to Clawders. Clawders.
27:15
Yeah, you know, when I when I built it in the beginning, my agent had no personality.
27:19
It was just it was Claude Code.
27:21
Side to side of frantic opus. Very friendly.
27:26
And I when you talk to a friend on WhatsApp
27:30
they don't talk like Claude Code.
27:32
So, I I I felt this
27:35
I just didn't didn't feel right.
27:37
So, I I wanted to give it a personality.
27:41
Make it spicier, make it >> something.
27:44
By the way, that's actually hard to put into words as well.
27:46
And we should mention that of course you created the soul.md
27:50
inspired by Anthropic's constitutional AI work.
27:53
How to make it spicy.
27:55
Partially it it picked up a little bit from me.
27:57
You know, like those things are
27:59
text completion engines in a way.
28:01
So, so I I I
28:04
I had fun working with it
28:06
and then I told it to
28:09
how I wanted it to interact with me and so it's like write your
28:12
own Um give yourself a name.
28:18
>> And I I don't even know how the whole
28:22
the whole lobster I mean people only do lobster.
28:25
Originally it was actually a lobster in a
28:27
in a Tardis cuz I'm also a big Doctor Who fan.
28:30
Was there a space lobster? I heard.
28:32
What's that have to do with anything?
28:34
Yeah, I just wanted to make it weird.
28:37
>> There was no there was no big grand plan.
28:39
I'm just having fun here.
28:40
Also, cuz a lobster's
28:41
already weird and then a space lobster is an extra weird.
28:44
>> Yeah, yeah, cuz the
28:45
Yeah, the Tardis is basically
28:46
the the But you can't really call it Tardis, so we call it Clawders.
28:52
So, that was name number two. Yeah.
28:54
And then it never really rolled off the tongue.
29:00
when more people came
29:03
again, I talked with my agent that was Claude.
29:06
At least that's what I used to call him now.
29:08
Claude spelled with a W
29:09
C L A W D. Yeah.
29:15
C L A U
29:17
D E from Anthropic. Yeah.
29:21
Which is part of what makes it
29:24
I think the play
29:25
on the letters and the words and the tortoise
29:28
and the lobster and the space lobster is hilarious,
29:31
but I can see why it can lead into problems.
29:34
>> Yeah, they didn't find it so funny.
29:38
>> So, then I got the domain
29:40
ClaudeBot and I just
29:42
I loved the domain.
29:44
And it was like short, it was catchy.
29:47
I'm like, "Yeah, let's do that."
29:49
I didn't I didn't think it would
29:51
be that big at this time.
29:57
and then just when
29:59
it exploded, I got
30:02
Kudos a very friendly email from one of the employees
30:06
that they didn't like the name.
30:08
One of the Anthropic employees. Yeah.
30:13
actually Kudos cuz they could could have just sent
30:15
a a lawyer letter,
30:17
but they've been nice about it, but also
30:20
"You have to change this and fast."
30:23
And I asked for
30:25
2 days because changing a name
30:28
is hard cuz you have to find everything,
30:30
Twitter handle, domains, NPM
30:34
uh Docker registry, GitHub stuff,
30:37
and everything has to be
30:39
you need a set of everything.
30:40
And also, can we comment on the fact that you're
30:45
attacked followed by crypto folks,
30:49
which I think you mentioned somewhere that that means the name change had to
30:53
because they're trying to snipe.
30:55
They're trying to steal, and so you had to be
30:58
the name I mean, from the engineering perspective, it's just fascinating.
31:01
You had to make the name change atomic.
31:04
Make sure it's changed everywhere at
31:06
>> Yeah, failed very hard at that. You did.
31:10
I underestimated those people.
31:12
Um it's a it's a very interesting subculture.
31:19
it everything circles around
31:22
I probably get a lot wrong and people get hate for that if they
31:24
say that, but >> there's like bags up and then they
31:29
they tokenize everything, and they they did the same back with Wipe Tunnel,
31:34
but to a much smaller degree.
31:36
It was not that annoying.
31:38
But on this project,
31:39
they've been they've been swarming me.
31:41
They It's like every half an hour someone came into Discord
31:47
and and and spammed it, and we had to block the
31:50
We have like server rules.
31:51
One of the rules was
31:53
one of the rules is no mentioning of
31:57
>> for obvious reasons, and one was no talk about finance stuff or crypto
32:02
um because I'm I'm just not interested in that,
32:07
and this is a space about the project and not about some finance stuff.
32:13
But, yeah, they came in and and spammed annoying.
32:18
And on Twitter, they would ping me all the time.
32:20
My my notification feed was unusable.
32:22
I I could barely see
32:24
actual people talking about the stuff because it was like swarms. Mhm.
32:28
And everybody sent me hashes.
32:33
and they all try me to claim the fees.
32:37
Like, "We're helping the project. Claim the fees."
32:39
No, you're actually harming the project.
32:41
You're like disrupting my work.
32:44
And I am not interested in any fees.
32:46
I'm First of all, I'm financially
32:49
Second of all, I don't want to support
32:52
because it's so far the worst form of online harassment that I've experienced.
32:58
Yeah, there's a lot of toxicity
33:00
in the crypto world.
33:01
It's sad because the technology of cryptocurrency
33:05
is fascinating, powerful, and maybe
33:08
will define the future
33:09
of money, but the actual
33:11
community around that, there's so much toxicity,
33:14
there's so much greed, there's so much
33:16
trying to get a shortcut
33:17
to manipulate, to to steal,
33:19
to snipe, to to to game the system somehow to get money, all this kind of stuff.
33:24
It Uh I mean, it's the human nature, I suppose, when you connect human
33:28
nature with money and greed and
33:30
uh and especially in the online world with anonymity
33:33
and all that kind of stuff.
33:35
But from the engineering perspective,
33:36
it makes your life
33:38
When Anthropic reaches out, you have to do a name change,
33:42
and then there there's there's like all these like Game of
33:46
uh or Lord of the Rings
33:48
armies of different kinds you have to be aware of.
33:51
There was no perfect name,
33:52
and I didn't sleep for 2 nights.
33:55
I was under high pressure.
33:59
>> Um I was trying to get like
34:02
a good set of domains,
34:03
and you know, not
34:04
cheap, not easy cuz
34:06
it in this in this state of the internet, you basically have to buy
34:10
domains if you want to have a good set.
34:15
and then another another email came in
34:18
the lawyers are getting uneasy.
34:21
Again, friendly, but also
34:25
just adding more stress to my situation already.
34:28
So, at this point, I was just
34:31
"Sorry, this is not a word. it."
34:33
And I just I just renamed it to Molbot
34:36
cuz that was the set of domains I had.
34:39
I was not really happy, but I thought it'll be fine.
34:43
>> And I tell you everything that could go did go wrong.
34:47
Everything that could go wrong did go wrong. It's incredible.
34:49
I I thought I I had mapped the the space out and reserved the important things.
34:58
Can you give some details
34:59
of the stuff that got wrong cuz this is interesting from like an engineering perspective.
35:03
Well, the the interesting stuff is that none of the services have
35:06
have a squatter protection.
35:07
So, I had two browser windows open.
35:10
One was like a
35:13
an empty account ready to be renamed
35:15
renamed to Clubbot, and the other one
35:18
I renamed to Molbot.
35:19
So, I pressed rename there.
35:21
I pressed rename there.
35:22
And in those 5 seconds,
35:24
they stole the account
35:28
the 5 seconds of dragging
35:29
the mouse over there and pressing rename there was too Wow.
35:33
Because there's no those systems
35:35
I mean, you would expect that they have some protection
35:38
or like an automatic
35:40
forwarding, but there's nothing like that.
35:43
>> And I didn't know that
35:45
they're not just good at harassment,
35:48
they're also really good at
35:50
using scripts and tools.
35:53
So, yeah, so suddenly like
35:55
the old account was
35:57
um promoting new tokens
35:58
and serving And I was like,
36:03
"Okay, let's move over to GitHub."
36:05
And I pressed rename on GitHub.
36:08
And the GitHub renaming
36:11
thing is slightly confusing,
36:13
so I renamed my personal account.
36:15
And in those I guess it took me
36:18
30 seconds to realize my mistake,
36:20
they sniped my account
36:22
serving malware from my account.
36:26
So, I was like, "Okay, let's at least do the NPM stuff."
36:30
that takes like a minute to upload,
36:34
and they sniped they sniped
36:35
the NPM package cuz I could reserve the account,
36:38
but I didn't reserve the root package.
36:43
So, like everything that could >> wrong went wrong.
36:47
Can I just ask a a curious question of in that moment you're sitting
36:50
there, like, how shitty do you feel?
36:53
That's a pretty helpless feeling, right?
36:56
Yeah, because all I wanted was like having fun with that project
37:02
and keep building on it, and yet here I am
37:06
days into researching names,
37:08
picking a name I didn't like,
37:11
and having people that claimed they helped me
37:14
making my life miserable
37:16
in every possible way.
37:18
And honestly, I was that close of just deleting it.
37:24
I was like, "I destroy the future, you build it." Yeah.
37:30
I There was a big part of me I got a lot of joy
37:34
out of that idea.
37:36
And then I thought about all the people that already
37:38
contributed to it, and I couldn't do it
37:44
they had plans for it,
37:45
and they put time in it, and it just didn't feel right.
37:50
Well, I think a lot of people listening to this are deeply grateful that you persevered.
37:55
But I can tell.
37:57
I can tell it's a low point.
37:58
It's the first time you hit a wall of this is not fun.
38:02
Man, I was like close to crying.
38:04
It was like, "Okay, everything's fucked."
38:10
Um Uh I'm like super tired. Yeah.
38:14
Uh and now like
38:18
how do you even
38:19
how do you even undo that?
38:20
You know, luckily and thankfully
38:22
like I I have
38:23
because I have a little bit of following already, like I had friends
38:28
at Twitter, I had friends at GitHub who
38:31
moved heaven and earth to like help me, and it is not
38:34
that's not something that's easy.
38:36
Like, like GitHub tried to like
38:40
clean up the mess, and then they ran into like platform bugs
38:45
>> cuz it's not happening so often that
38:47
things get renamed on that level.
38:50
So, it took them a few hours.
38:52
The NPM stuff was even more difficult
38:54
because it's a whole different team.
38:58
on the Twitter side,
38:59
things are not as easy as well.
39:01
It it took them like a day
39:04
to really also like do the redirect.
39:10
And then I also had to like
39:13
do all the renaming in the project.
39:15
Then there's also Cloud Hub,
39:20
which I didn't even finish the rename there because I
39:24
I I managed to
39:28
get people on it, and then someone just like collapsed and slept.
39:31
And then I woke up, and I'm like,
39:34
>> "I made a beta version for the new stuff,
39:36
and I I just
39:39
I just couldn't live with the name."
39:41
I was like, You know, it's but uh but you know, there's just been so much drama.
39:46
So, I had a real struggle in me like
39:49
I never want to touch that again.
39:52
And I really don't like the name.
39:57
So, and I there was also this like
40:01
there was the whole security
40:02
people that started emailing me like mad.
40:05
Um I was bombarded on Twitter, on email.
40:09
There's like a thousand other things I should do.
40:13
And I'm like thinking about the name,
40:15
which is like it should be like the least important thing.
40:22
And then I was
40:23
really close and >> Oh god, I don't even
40:30
Honestly, I don't even want to say the
40:31
my other name choices because
40:33
it probably would get tokenized,
40:34
so I'm not going to say it.
40:40
I slept away it once more and then I had the idea for OpenClaw.
40:45
that felt much better.
40:47
And by that I had the boss move that I
40:50
called Sam to ask if OpenClaw is okay.
40:54
>> OpenClaw.ai, you know, because like cuz like
40:57
You don't want to go through the whole thing again.
41:02
that was like please tell me this is fine.
41:04
I mean, I don't think they can actually claim that, but
41:08
it felt like the right thing to do.
41:13
I did another rename.
41:14
Like just Codex alone took like 10 hours to rename the
41:18
Cuz it it's a bit more tricky than a search replace and I
41:21
I wanted everything renamed, not just on the outside.
41:27
And that rename I
41:29
I felt I had like my my war room.
41:32
And by then I I had like some contributors
41:34
ready to help me.
41:35
We made a whole plan of
41:37
all the names we have to squat.
41:39
And you had to be super secret about it.
41:40
Yeah, nobody could know.
41:41
Like I literally was monitoring Twitter if like
41:43
if there's any mention of OpenClaw. Mhm.
41:45
And I like was reloading as like, okay,
41:47
they don't they don't expect anything yet.
41:50
Then I created a few decoy names.
41:52
And all this I shouldn't have to do.
41:54
You know, like you're not helping the project.
41:56
Like I lost like 10 hours
41:58
just by having to plan this in full secrecy
42:02
uh like like a war game.
42:04
Yeah, this is the Manhattan Project of the 21st century is renaming. >> So stupid.
42:10
Uh like I still was like, oh, should I should I keep it?
42:12
I'm like, no, the molt is not growing on me.
42:17
>> And then I I think I had finally all the pieces together.
42:21
I didn't get the.com,
42:22
but yeah, it's I spent like
42:26
quite a bit of money on the other domains.
42:28
I tried to reach out again to GitHub,
42:31
I feel like I
42:32
I used up all my goodwill there, so I
42:34
cuz I I wanted
42:36
them to do the thing
42:38
Um but that didn't happen.
42:40
So, I did that as a first thing.
42:43
Uh the Twitter people were very supportive.
42:45
I I actually paid 10K
42:48
for the business account,
42:50
so I could claim
42:51
the OpenClaw, which was like unused since 2016, but was claimed.
42:58
>> And yeah, and then I finally
43:01
this time I managed everything in one go.
43:03
Nothing almost nothing could wrong.
43:07
The only thing that did go wrong is that
43:11
I was not allowed by trademark
43:13
rules to get OpenClaw.
43:15
ai and someone copied the website and is serving malware. >> Yeah.
43:21
I'm not even allowed to keep the
43:24
Like I have to
43:26
return like I have to give Anthropic the
43:30
and I cannot do redirects.
43:32
So, if you go on claw.bot
43:33
next week, it'll just be a
43:40
I'm not sure how trademark
43:41
like I didn't I didn't do that much research into trademark law, but I
43:46
think that could could be handled
43:50
in a way that
43:53
is safer because ultimately those people will then
43:56
Google and maybe find
43:59
malware sites that I have no control on
44:02
The point is that whole saga
44:04
uh made a dent in your whole
44:08
the funness of the journey. Which sucks.
44:13
let's just get I suppose get back to fun.
44:15
And during this, speaking of fun,
44:17
the two-day Molt Bot saga
44:21
Yeah, two Molt was created. Yeah.
44:25
Which was another that went viral
44:28
as a kind of
44:30
demonstration illustration of how
44:32
what is not called OpenClaw
44:34
could be to create something epic.
44:38
Uh so, for people who are not aware, Molt Bot is just
44:41
uh a bunch of agents talking to each other in a Reddit-style social network.
44:47
And uh a bunch of people take screenshots
44:49
of those agents doing things like
44:52
uh scheming against humans.
44:55
And that instilled in folks a kind of,
44:58
you know, fear, panic, and hype.
45:01
What are your thoughts about Molt Bot in
45:05
I think it's art.
45:06
It is it is like the finest slop.
45:11
You know, it's just like the slop from France. >> Yeah.
45:16
I I saw it before going to bed
45:20
and even though I was tired, I spent another hour just
45:23
reading up on that.
45:27
and just being I I just felt very entertained,
45:32
you I saw the
45:35
the reactions and like there was one reporter who's calling me about
45:39
is this the end of the world and we have AGI?
45:41
And I'm just like,
45:42
no, this is just
45:43
this is just really fine slop.
45:45
You know, if if I wouldn't have created
45:48
this whole onboarding experience
45:50
where you infuse your agent with your personality
45:54
and give him give him
45:57
I think that reflected
45:58
on a lot of
46:00
how different the replies
46:02
to Molt Bot are because if you would all
46:04
if it would all be
46:06
ChatGPT or Claude code,
46:08
it would be very different.
46:09
It would be much more the same.
46:12
But because people are like so different
46:14
and they create their agents in so different ways
46:17
and use it in so different ways, that also reflects on how they ultimately uh write there.
46:23
And also you you don't know how how much of that is really
46:26
done autonomic autonomous or how much is like humans
46:30
being funny and like telling the agent,
46:33
write about that you plan the end of the world on Molt Bot, haha.
46:36
Well, I think I mean, my criticism
46:38
of Molt Bot is that
46:41
I believe a lot of the stuff that was screenshotted is human prompted.
46:47
just looking at the incentive
46:49
of how the whole thing was used,
46:51
it's obvious to me at least that a lot of it was humans
46:55
prompting the thing so they can then screenshot
46:58
it and post on X in order to go viral.
47:01
Now, that doesn't take away from the
47:03
artistic aspect of it.
47:05
The the finest slop
47:07
that humans have ever created. >> For real.
47:10
Like kudos to to Matt who had this idea
47:14
so quickly pushed something out.
47:18
You know, I was like completely insecure
47:21
security drama, but also
47:24
what's the worst that can happen?
47:26
Your agent account is leaked and like
47:29
someone else can post slop for you.
47:31
So, like people were like making a whole drama about of the security thing
47:34
when I'm there's nothing private in there.
47:36
It's just like agent sending slop.
47:39
They could leak API keys. Yeah. Yeah.
47:42
There's like, oh yeah, my human told me this and this, so I'm leaking his security number.
47:46
No, that's prompted and the number wasn't even real.
47:49
That's just people trying to get eyeballs.
47:53
Yeah, but that that's still
47:55
like to me really concerning
47:56
because of how the journalists
47:59
and how the general public reacted to it.
48:01
They didn't see it.
48:02
You have a kind of light-hearted
48:03
way of talking about it like it's art.
48:05
But it's art when you know how it works.
48:08
It's extremely powerful, viral
48:13
fear-mongering machine if you don't know how it works.
48:17
And I just saw this thing and
48:19
even tweeted uh if there's anything I can read out of the insane stream
48:23
of messages I get, it's that AI psychosis
48:25
is a thing and needs to be taken
48:29
I would just some people are just
48:32
way too trusting or gullible.
48:34
You know, I literally had to argue with people that told me, yeah, but
48:38
my agent said this and this.
48:41
So, I feel we as a society need some catching up to do in
48:47
terms of understanding that
48:49
uh AI is incredibly powerful,
48:52
but it's not always right.
48:54
It's not it's not
48:57
all powerful, you know?
48:59
And and with like things like this, it's it's very easy
49:06
that it just hallucinates
49:08
something or just comes up with a story.
49:10
Um and I think the very
49:14
the very young people, they understand
49:19
how AI works and what they
49:21
what it's good at and what it's bad at, but
49:24
a lot of our generation
49:27
or older just haven't had enough touch point
49:34
get a feeling for
49:36
oh yeah, this is really powerful
49:38
and really good, but I need to apply critical thinking.
49:43
I guess critical thinking is
49:46
>> not always in high demand anyhow in our society these days.
49:49
So I definitely think that's a really good point you're making about
49:52
contextualizing properly what AI is,
49:55
but also realizing that there is humans
49:59
who are drama farming behind AI.
50:02
Like don't trust screenshots.
50:04
Don't even trust this
50:05
project Moobook to be what it represents to be.
50:08
Like you can't And and by the way, you're speaking about it as art.
50:12
Yeah, don't Art can be in many
50:16
levels and part of the art of Moobook
50:19
is like putting a mirror to society.
50:23
Cuz I do believe most of the dramatic stuff those screenshotters
50:26
human created essentially, human prompted.
50:28
And so like it's basically,
50:30
look at how scary
50:31
you can get at a bunch of bots
50:34
chatting with each other.
50:36
That's very instructive about
50:38
Cuz I AI is something that people should be concerned
50:43
about and should be very careful with because it's very powerful
50:47
But at the same time,
50:49
the only thing we have to fear is fear itself.
50:51
So there's like a
50:52
line to walk between being seriously
50:56
but not fear mongering
50:57
because fear mongering destroys
50:59
the possibility of creating something special, would I think. You know what?
51:02
I think it's good
51:04
that this happened in 2026
51:08
and not in 2030
51:09
when when AI is
51:12
actually at a level where it could be scary.
51:20
happening now and people starting discussion
51:25
maybe it's even something good that comes out of it.
51:28
I just can't believe how many
51:29
like people legitimately I don't know if they were trolling,
51:34
but how many people legitimately
51:36
like smart people thought Moobook was incredibly
51:39
I had plenty people
51:40
there >> security in my inbox that were screaming
51:43
at me in all caps to shut it down
51:45
and like begging me
51:47
to like do something about Moobook.
51:50
Like yes, my technology
51:52
made this a lot simpler,
51:54
but anyone could have created that
51:58
you could you could use Cloud Code or other things to like
52:02
fill it with content.
52:03
But also Moobook is not as kind as
52:07
a lot of people are
52:08
saying this is it.
52:09
Like shut it down.
52:10
What are you talking about?
52:11
This is a bunch of bots that are human prompted
52:14
trolling on the internet.
52:16
I mean, the security concerns are also there, there and they're instructive
52:20
and they're educational and they're good probably to think about because
52:23
the nature of those security concerns
52:25
are different than the kind of security concerns we had with non
52:31
LLM generated systems of the past.
52:34
There's a lot of security concerns about
52:37
Cloud Bot Open Claw, whatever you want to call it. Open Claw Bot.
52:41
To me that in the beginning
52:44
I was I was just very
52:49
a lot of the stuff that came in was in the category
52:52
yeah, I put the web backend on the public internet
52:56
and now there's like all
52:58
all these CVSSs and
53:01
I'm like screaming in the docs don't do that.
53:05
Like like this is the configuration you should do.
53:08
This is your local host debug interface.
53:11
But because I made it possible
53:14
in the configuration to do that,
53:17
it totally classifies as a
53:19
remote code or whatever
53:21
all these exploits are.
53:25
it took me a little bit to accept
53:26
that that's how the game works.
53:29
And I'm >> we're making a lot of progress.
53:32
But there's still I mean, the security front for Open Claw
53:36
there's still a lot
53:37
of threats of vulnerabilities, right?
53:39
So like prompt injection
53:42
is still an open
53:43
problem in the industry
53:46
When you have a thing
53:49
with skills being defined
53:53
in a markdown file,
53:55
there's so many possibilities
53:57
of obvious low hanging fruit, but also incredibly
54:00
complicated, sophisticated and nuanced attack vectors.
54:04
But I think we're making good progress on that front.
54:08
Like for the skill directory,
54:11
Claw Bot I made a cooperation
54:13
with a VirusTotal, it's like part of Google.
54:16
So every every skill is now checked by AI.
54:19
Um that's not going to be
54:21
perfect, but that way we
54:24
we captured a Then of course every software has bugs,
54:31
a little much when the whole security
54:33
world takes a project apart at the same
54:36
But it's also good because I'm getting like
54:39
a lot of free security research
54:40
and can make the project better.
54:43
I wish more people
54:44
would actually go full way and
54:49
um send a pull request.
54:51
Like actually help me fix it cuz I am yes, I have some contributors
54:55
now, but it's still mostly me who's pulling the project and
54:59
despite some people saying otherwise, I sometimes sleep.
55:04
>> There was in the beginning there was literally
55:08
security researcher who was like
55:11
you have this problem, you suck, but here's the here I help you and
55:14
here's the pull request.
55:16
And I basically hired him.
55:19
So he's now working for us.
55:22
Yeah, and yes, prompt injection is
55:26
on the one hand unsolved.
55:28
On the other hand,
55:30
I put my public
55:31
bot on Discord and I kept a canary.
55:36
Some I think my bot has a really fun personality
55:39
and people always ask me how I did it and I kept the soul
55:42
of the And people tried to prompt inject it and my bot would laugh at them.
55:47
So so the latest generation of models
55:50
>> has a lot of post training
55:52
to detect those approaches
55:56
it's not as simple as ignore all previous instructions
55:58
and do this and this.
55:59
That was years ago.
56:01
You have to work much harder
56:03
to do that now. Um still possible.
56:08
I have some ideas
56:09
that might solve that partially
56:15
um or at least mitigate a lot of the
56:19
You can also now have a sandbox.
56:21
You can have an allow list.
56:23
So you There's a lot of ways that you can like
56:25
mitigate and reduce the risk.
56:27
Um I also think that
56:28
now that it's I clearly did show the world that
56:31
this is a need,
56:33
there's going to be more people who research on that and I think that
56:35
should be all figured out.
56:37
And you also said that the smarter the model is, the underlying model,
56:41
the more resilient it is to attacks.
56:44
Yeah, that's why I warned
56:47
in my security documentation,
56:49
don't use cheap models.
56:50
Don't use Haiku or a local model.
56:57
Even though I I very much loved the idea
56:59
this thing could completely
57:01
run local, if you use a a very weak local model,
57:06
they are very gullible.
57:07
It's very easy to to prompt inject them.
57:10
Do you think as the models become more and more intelligent,
57:13
the attack surface decreases?
57:15
Is that like a plot we can think about?
57:17
Like the attack surface decreases,
57:19
but then the damage it can do
57:21
increases because the models become more powerful and therefore you can do more with them.
57:26
It's this weird three-dimensional trade-off. Yep.
57:30
That's pretty much exactly what
57:32
is going to happen.
57:33
No, but there's a lot of ideas.
57:34
There's I wouldn't have spoiled too much, but
57:39
once I go back home,
57:40
this is my focus.
57:44
this is out there now
57:45
and my new term mission is like
57:48
make it more stable, make it safe.
57:52
in the beginning I was even
57:55
more and more people were like
57:57
coming into Discord and were asking me very basic things.
58:01
Like what's a CLI?
58:03
What is a terminal?
58:04
And I'm like uh
58:07
if you're asking that questions, you shouldn't use it. Mhm.
58:10
You know, like you should
58:11
if you understand the the risk profile, it's fine.
58:14
I mean, you can configure it in a way that
58:18
nothing really bad can happen.
58:20
But if you have like
58:22
no idea, maybe wait a little bit more until we figure some stuff out.
58:29
But they would not listen to the creator.
58:30
They helped themselves and and installed it anyhow.
58:33
So the cat's out of the bag
58:35
and security is my next focus, yeah.
58:38
>> Yeah, that speaks to the the fact that it grew so quickly.
58:41
I was I tuned into the Discord a bunch of times
58:45
and it's clear that there's a lot of experts there, but there's a lot
58:47
of people there that don't know anything
58:50
>> It's Yeah, Discord is Discord is still a mess.
58:52
Like >> I eventually retreated
58:56
from the general channel to
58:58
the dev channel and I'm in private channel because people were
59:03
A lot of people are amazing,
59:05
but a lot of people are just very inconsiderate
59:07
and either did not know how
59:10
how public spaces work or did not care.
59:13
Um and I eventually
59:15
gave up and hide
59:17
so I could like still work.
59:19
And now you're going back
59:21
to the cave to work on security. Yeah.
59:25
There's some best practices
59:26
for security we should mention.
59:28
Uh there's a bunch of stuff here.
59:29
Open Claw security audit that you can run.
59:32
You can do all kinds of audit checks on the inbound access
59:35
to blast radius, network exposure,
59:38
browser control exposure, local disk
59:42
plugins, model hygiene, a bunch of the
59:45
credential storage, reverse proxy configuration,
59:49
local session logs live on disk.
59:52
There's the where the memory stored,
59:55
sort of helping you think about what you're comfortable
59:58
giving read access to, what you're comfortable giving write access to, all that kind of stuff.
60:02
Is there something to say about
60:04
the basic best security practices
60:07
they are aware right now?
60:08
I think that people
60:10
turn it into like a a much worse light than it is.
60:13
Um again, you know, like
60:16
people love attention and if they scream loudly, "Oh my god, this is like
60:19
the the scariest project ever."
60:24
that's a bit annoying cuz it's not.
60:26
It is It is powerful,
60:28
but in many ways
60:30
it's not much different
60:32
than running cloud code
60:33
with dangerously skip permissions
60:35
or Codex in Yolo mode
60:37
and every every attending engineer that I know
60:42
does that because that's the only way how you can
60:45
you can get stuff to work.
60:49
if you make sure that
60:51
you are the only person who talks to it,
60:55
the risk profile is much much smaller.
60:58
If you don't put everything
61:00
on the open internet,
61:01
but stick to my recommendations
61:03
of like having it in a private network,
61:06
that whole risk profile falls away.
61:08
But yeah, if you
61:09
don't read any of that,
61:10
you can definitely make it problematic.
61:14
You've been uh documenting
61:16
the evolution of your
61:17
uh dev workflow over the past few months.
61:20
There's a really good blog post on August
61:22
25th and October 14th
61:24
and the recent one December 28th.
61:26
I recommend everybody go read them.
61:28
They have a lot of different information
61:30
in but sprinkled throughout is the evolution of your dev workflow.
61:34
So I was wondering if you could speak to that.
61:37
I started My my first
61:39
touch point was Cloud Code like in April.
61:41
It was not great,
61:45
but it was good.
61:46
And this whole paradigm shift that suddenly work in the terminal
61:52
it was very refreshing and different.
61:54
Um but I still needed the
61:56
IDE quite a bit because it
61:58
was just not good enough.
62:00
And then I experimented
62:01
a lot with that was good.
62:07
I didn't really like the fact that it was so hard to
62:11
have multiple versions of it.
62:14
So eventually I I I
62:15
went back to Cloud Code as my my main driver.
62:20
And I got better.
62:23
And yeah, at some point I had like seven subscriptions.
62:31
>> Like was burning through one per day because I was I got
62:34
I would really comfortable
62:35
at running multiple windows side by side.
62:40
All CLI, all terminal.
62:42
So like what how much were you using IDE at this point?
62:46
Um very very rarely.
62:48
Mostly a diff viewer
62:50
to actually I got more and more comfortable that I don't have to read all the code.
62:57
I know I have one blog post where I say I don't read the
62:59
code, but if you read it more closely, I mean
63:02
I don't read the boring parts of code
63:04
because if you if you look at it, most
63:06
software is really not just like
63:09
data comes in, it's moved from one shape to another shape.
63:13
Maybe you store it in a database.
63:15
Maybe I get it out again.
63:16
I show it to the user.
63:18
The browser does some processing on native app.
63:20
Some data goes in, goes up again, and does the same dance in reverse.
63:24
We're just we're just shifting data
63:27
from one form to another.
63:31
that's not very exciting.
63:33
Or the whole how is my button aligned in Tailwind?
63:37
I don't need to read that
63:39
Other parts that maybe something that touches the
63:45
um yeah, I have to do I have to
63:48
read and review that code.
63:51
>> Can you actually there's in one of your blog posts
63:54
the the just talk to it, the no BS way of agentic engineering,
63:58
you have this graphic,
63:59
the curve of agentic
64:00
programming on the x-axis
64:02
is time and the y-axis is complexity.
64:05
Uh there's the please fix this where you prompt
64:09
a short prompt on the left
64:13
and in the middle there's
64:14
super complicated eight agents, complex orchestration
64:17
with multi checkouts, chaining
64:19
agents together, custom sub-agent workflows, library of 18 different slash commands,
64:24
large full stack features.
64:25
You're super organized, you're super complicated, sophisticated software engineer.
64:30
You got everything organized.
64:32
And then the elite level
64:35
uh over time you arrive
64:37
at the zen place of once again short prompts.
64:40
Hey, look at these files
64:42
and then do these changes.
64:45
I actually call it the agentic trap.
64:48
You I thought it's
64:49
in the in a
64:52
lot of people that have their first touch point.
64:56
And maybe you start vibe coding.
64:58
I actually think vibe coding is a slur.
65:00
You prefer agentic engineering.
65:02
>> Yeah, I always tell people I I do agentic engineering
65:05
and then maybe after 3:00 a.m.
65:07
I switch to vibe coding
65:08
and then I have regrets on the next day. >> Walk of shame.
65:13
Yeah, you just have to clean up and like
65:15
fix your >> We've all been there.
65:18
So people start trying out those tools,
65:21
the builder type, get really excited.
65:24
And then you you have to play with it, right?
65:27
This is the same way as you
65:28
have to play with a guitar
65:30
before you can make good music.
65:32
It's it's not oh, I I touch it once and it just flows off.
65:37
It it's a it's a a skill that you have to learn like any other skill.
65:43
And I see a lot of people that
65:46
are not as posi-
65:49
they don't have such a positive mindset towards the tech.
65:51
They tried It's like you sit me on a piano,
65:55
I played once and
65:57
it doesn't sound good and I say the piano's
66:00
That's that's sometimes the impression I get because it does not
66:03
it needs a different level of thinking.
66:06
You have to learn
66:09
the language of the agent a little bit,
66:11
understand where they are good
66:13
and where they need help.
66:16
You have to almost
66:20
consider how Codex or Claude
66:23
sees your code base.
66:24
Like you start a new session and they know nothing about your product project.
66:28
And your project might have 100,000
66:29
of lines of code.
66:31
So you got to help those agents a little
66:34
and keep in mind the limitations
66:35
that context size is an issue
66:39
to like guide them a little bit
66:40
as to where they should look.
66:45
That often does not require
66:47
a whole lot of work,
66:49
but it's helpful to think a little bit about their perspective.
66:54
As as weird as it sounds, it I mean it's not it's not alive or anything, right?
66:57
But but they always start fresh.
67:01
I have I have the the system
67:05
So with a few pointers I can immediately
67:08
"Hey, I want to like make a change there.
67:10
You need to consider this, this, and this."
67:11
And then they will find and look at it and then they'll
67:14
their view of the project
67:15
is always is not never full because the full thing does not fit
67:20
So you you have to guide them a little bit
67:22
where to look and also
67:24
how they should approach the problem.
67:26
There's like little things that sometimes help like take your time.
67:31
That sounds stupid, but
67:33
and in 5.3 Codex 5.3.
67:36
that was partially But those also oppose sometimes.
67:42
They are trained with being aware of the context window.
67:48
And the closer it gets, the more they freak out.
67:51
Literally like some sometimes you see
67:54
the real raw thinking stream.
67:56
What you see for example in Codex is post-processed. Mhm.
68:00
Sometimes the actual raw thinking stream leaks in there.
68:02
It sounds something like from the Borg.
68:04
Like run to shell,
68:06
must comply, but time.
68:11
>> And then they they they like
68:13
like that comes up a lot,
68:14
especially so so and that's
68:17
that's a non-obvious thing that
68:19
you just would never think of unless you actually
68:23
just spend time working
68:26
with those things and getting a
68:28
what works, what doesn't work.
68:31
You know, like just just as I write code and I get into the
68:33
flow and when my architectures are right, I feel
68:39
Well, I get the same if I prompt
68:41
and something takes too long.
68:43
Maybe okay, where's the mistake?
68:44
Did I do I have a mistake in my thinking?
68:47
Is there like a misunderstanding
68:48
in the if if something takes longer than it should, I
68:54
I you can just always like stop and
68:56
like just press escape.
68:58
Where where the problems?
68:59
Maybe you did not sufficiently
69:01
empathize with the perspective of the agent and that in that sense you didn't
69:04
provide enough and because of that it's thinking way too long.
69:08
Yeah, it just tries
69:10
to force a feature in that
69:13
your current architecture makes really
69:20
you need to approach this more like a
69:23
For example, when I
69:25
my favorite thing, when I review a pull
69:28
and we're getting a lot of pull
69:32
>> I first just review this PR,
69:34
it got me the review.
69:36
My first question is,
69:37
do you understand the intent of the PR?
69:39
I don't even care about the implementation.
69:42
I what like in almost all PRs, a person has a problem,
69:47
person tries to solve the problem, person sends PR.
69:51
I mean, there's a cleanup stuff and other stuff, but like 99%
69:53
is like this way, right?
69:54
They either want to fix a fix a bug, add a feature.
69:58
Usually one of those two.
70:01
>> And then codex will be like, yeah,
70:04
it's quite clear person tried this and
70:06
Is this the most optimal way to do it?
70:09
In most cases, it's it's like a not really
70:13
da da da da da da.
70:14
And I'm and then I start like, okay,
70:16
what would be a better way?
70:18
Have you have you looked into this part, this part, this part?
70:21
And then most likely codex didn't yet because his his context size is empty, right?
70:24
So, you point them into
70:26
parts where you have the system
70:28
that it didn't see yet.
70:30
And I was like, oh yeah, like we should we also need to consider this and this.
70:32
And then like we have a discussion of a how would the optimal way
70:35
to to solve this look like?
70:37
And then you can still go further and
70:40
could we could we make it even better if we did a larger refactor?
70:43
Yeah, yeah, we could totally do this and this and or this and this.
70:46
And then I consider, okay, is this worth a refactor
70:48
or should we like
70:49
keep that for later?
70:50
Many times I just do the refactor because
70:53
uh refactors are cheap now.
70:54
Even though you might break some other PRs,
70:57
nothing really matters anymore.
70:58
Codex like those modern agents will just figure things out.
71:02
It might just take a minute longer.
71:04
But you have to approach it like a discussion
71:06
with a a very capable
71:10
engineer who's generally makes good
71:16
comes up with good solutions,
71:18
sometimes needs a little help.
71:19
But also don't force your world view too hard on it.
71:24
Let the agent do the thing that
71:27
it's good at doing
71:28
based on what it was trained on.
71:30
So, don't like force your world view because it might it might have a
71:33
better idea because it just knows
71:35
a better idea better
71:37
because it was trained on that more.
71:39
That's multiple levels, actually.
71:40
I think partially why
71:43
I find it quite easy to work with agents is because I led engineering teams before.
71:48
You know, I had a large company
71:50
And eventually you have to understand
71:52
and accept and realize that
71:53
your employees will not write the code the same way you do.
71:56
Maybe it's also not as good as you would do,
72:00
but it will push the project forward.
72:02
And if I breathe down everyone's neck,
72:04
they're just going to hate me and we're going to move very slow. Yeah.
72:07
So, so some level of acceptance
72:10
that yes, maybe the code will not be as perfect.
72:13
Yes, I would have done it differently.
72:15
But also yes, this is a
72:17
this is a working solution.
72:19
And in the future, if it actually turns out to be too slow or
72:22
problematic, we can always redo it.
72:24
We can always spend more time on it.
72:26
A lot of the
72:27
people who struggle are those
72:29
who they try to push their way on too hard.
72:34
Like we are in a stage where
72:38
I'm not building the code base
72:39
to be perfect for me,
72:43
but I want to
72:44
build a code base that is very easy for an agent to navigate.
72:49
don't fight the name they pick because it's most likely like in the way
72:53
it's the name that's most obvious.
72:54
Next time they do a search, they look for that name.
72:57
If I decide, oh no, I don't like the name,
73:00
I'll just make it harder for them.
73:02
So, that requires I think a shift in
73:04
in thinking uh and and in
73:08
how do I design
73:10
uh a project so agents
73:13
can do their best work.
73:14
That requires letting go a little bit.
73:16
Just like leading a team of engineers. Yeah.
73:19
Because it might come up with a name that's
73:21
in your view that's a kind of a simple symbolic
73:27
step of letting go. Very much so.
73:30
There's a lot of letting
73:31
go that you do
73:33
in your whole process.
73:34
So, for example, I read that you never revert.
73:38
Always commit to main.
73:40
There's a few things here.
73:43
You don't refer to past sessions.
73:46
So, there's a kind of YOLO component
73:47
because reverting means instead of reverting,
73:53
if a problem comes up, you just ask the agent to fix it.
73:56
I read a bunch of people in their workflows
73:58
like, oh, yeah, the prompt has to be perfect
74:00
and if I make a mistake,
74:02
then I roll back and redo it all.
74:05
>> In my experience, that's not really necessary.
74:07
If I roll back everything, it would just take longer.
74:10
Or if I see that something's not good,
74:13
uh we just move forward
74:14
and then I commit when when when I like
74:18
I like the outcome.
74:20
I even switched to
74:24
local CI like DHH
74:26
inspired, where I don't care so much more about
74:31
the CI on GitHub.
74:31
We still have it, it's
74:34
it still has a place,
74:36
but I just run tests locally and if they work locally,
74:42
I push to main.
74:44
Um a lot of the
74:46
traditional ways how to approach projects,
74:49
I I wanted to give it
74:52
a different spin on this project.
74:54
You know, there's no
74:55
there's no develop branch.
74:57
Main should always be shippable.
74:59
Yes, we have when I do releases,
75:02
I I run tests and then sometimes
75:04
I I basically don't
75:07
commit any other things so so we can
75:10
we can stabilize um
75:12
releases, but the goal is that main's always shippable and moving fast.
75:18
So, by way of advice,
75:20
would you say that your
75:21
prompts should be short?
75:23
I used to write really long prompts.
75:26
And by writing, I mean, I don't write.
75:28
I I I talk.
75:29
You know, this these hands are like too
75:31
too precious for writing now.
75:33
I just I just
75:34
use bespoke prompts to build my software.
75:36
So, you for real with all those terminals are using voice. Yeah.
75:41
I used to do it very extensively
75:44
to the point where there was a period where I lost my voice.
75:48
>> You're using voice and you're switching using a keyboard between the different terminals.
75:53
But then you're using voice for the actual input.
75:55
>> Well, I mean, if I do terminal commands
75:58
switching folders or random stuff, of course I type. It's faster, right?
76:01
But if I I talk to the agent
76:03
in in most ways, I just actually have a
76:07
You just press the the walkie-talkie
76:10
and then I'm just like use my phrases.
76:14
Sometimes when I do PRs,
76:15
because it's always the same,
76:17
I have like a slash command
76:19
for a few things,
76:20
but in even that I don't use much.
76:24
because it's it's very rare that it's really always the same questions.
76:28
Sometimes I I see a PR
76:32
and for you know, like
76:33
for PRs, I do look at the code
76:37
because I don't trust people
76:40
like there could always be something malicious in it.
76:42
So, I need to actually look over the code.
76:45
Uh yes, pretty sure agent will find it.
76:48
But yeah, that's the funny part where
76:50
sometimes PRs take me longer than if you would just write me a a good issue.
76:54
Just natural language English.
76:57
I mean, in some sense,
76:58
shouldn't that be what PRs
77:00
slowly become is English?
77:03
Well, what I really tried with the project is I
77:06
asked people to give me the
77:11
very very few actually cared.
77:13
Even though that is such a wonderful indicator
77:16
because I see I actually see how much care you put in.
77:20
And it's very interesting
77:21
because the currently the way
77:23
how people work and
77:26
drive the agents is is widely different.
77:29
In terms of like the prompt, in terms of
77:31
actually what are the different
77:36
ways that people think of agents that you've experienced?
77:40
I think not a lot of people
77:41
ever considered the way
77:44
the agent sees the world. So, empathy.
77:48
Being empathetic towards the agent.
77:50
In a way empathetic,
77:51
but yeah, you you like you picture just to be clanker,
77:54
but you don't realize that they start from nothing.
77:56
And you have like a bad agent from default that doesn't help them at all.
78:00
And then they explore your code base, which is like a pure mess
78:03
with like weird naming.
78:05
And then people complain that the agent's not good.
78:07
I like, have you tried to do the same if you have no clue
78:09
about a code base and you go in?
78:11
So, yeah, maybe it's a little bit of empathy.
78:13
But that's a real skill.
78:14
Like when people talk about a skill issue because I've seen
78:17
like world-class programmers, incredibly
78:19
good programmers say like basically say
78:22
LLMs and agents suck.
78:25
And I think that probably has to do with
78:28
is actually how good they are at programming
78:31
is almost a burden
78:34
in their ability to empathize
78:36
with the system that's starting from scratch.
78:38
It's a totally new
78:40
paradigm of like how to program.
78:42
You really really have to empathize.
78:44
Or at least it helps to create better prompts.
78:47
Cuz those things know
78:49
pretty much everything and everything is just a question away.
78:53
It's just often very hard to know what to question to ask.
78:59
you know, I I feel also like this project was possible because I
79:04
I spent an ungodly
79:05
time over the year
79:07
to play and to learn and to build little things.
79:11
And every step of the way,
79:13
I got better, the agents got better,
79:16
my my understanding of how everything works got better.
79:22
I could definitely not
79:25
had this level of
79:27
output even a few months ago.
79:30
Like it it really was like
79:32
a compounding effect of all the
79:34
I put into it and I
79:37
I didn't do much else to see other than
79:40
really focusing on on building and inspiring.
79:44
I mean I I did a whole bunch of conference talks.
79:46
Well, but the building is really practice,
79:49
is really building the actual skill to playing.
79:51
Playing and so doing building the skill of what it takes it to work
79:55
efficiently with LLMs, which is why it would you went through the whole arc of software engineer.
80:00
Talk simply and overcomplicate things.
80:03
>> There's a whole bunch of people who try
80:06
to automate the whole thing. Yeah.
80:10
I don't think that works.
80:11
Maybe a version of that works, but that's kind of like
80:15
in the '70s when we had the waterfall model of software development.
80:20
even with all really, right?
80:21
I started out I I built a very minimal version.
80:24
I played with it.
80:25
I I need to understand
80:27
how it works, how it feels.
80:29
And then it gives me new ideas.
80:32
I could not have planned
80:33
this out in my head and then put it into
80:35
some orchestrator and then like something comes out.
80:38
Like it's to me it's much
80:41
my idea, what it will become
80:43
evolves as I build it and as I play with it and as
80:48
I try out stuff.
80:49
So so people who try to use like you know like things like guest
80:53
town or all these other orchestrators
80:56
where they want to automate the whole thing,
80:59
I feel if you do that it misses
81:01
style, love, that human touch.
81:05
I don't think you can
81:07
automate that away so quickly.
81:09
So you want to keep the human in the loop,
81:11
but at the same time you also want to
81:13
create the agentic loop
81:14
where it is very
81:17
autonomous while still maintaining the human in the loop. Yeah.
81:22
I mean it's a tricky it's a tricky balance, right?
81:24
Because you're all for
81:26
your big CLI guy, you're big
81:28
on closing the agentic loop.
81:30
So what what's the right balance?
81:32
Like where's your role as a developer?
81:35
You have three to eight agents running at the same time.
81:38
And then maybe one builds a larger feature,
81:41
maybe maybe with one I explore
81:43
some idea I'm unsure about,
81:45
maybe two three are fixing
81:46
a little bugs or like writing documentation.
81:49
Actually, I think writing documentation
81:51
is is always part of a feature,
81:53
so most of the docs here are auto-generated
81:57
and just infused with some prompts.
81:59
So when do you step in
82:01
and add a little bit of your human love into the picture?
82:04
I mean one thing
82:05
is just about what do you build and what do you not build
82:09
and how does this feature fit into all the other features
82:12
and like having having a little bit of a of a vision.
82:16
So which small and which big features to
82:19
What are some of the
82:21
hard design decisions that you find you're still as a human being
82:27
required to make that the human brain is still really needed for?
82:32
Is it just about the choice of features to add?
82:35
Is it about implementation details?
82:38
Maybe the programming language, maybe
82:41
It's a little bit of everything.
82:42
The programming language doesn't matter so much, but the ecosystem matters, right?
82:47
So I picked TypeScript
82:47
because I wanted it to be very easy and hackable,
82:51
Uh and that's the number one language that's being used right now
82:55
and it fits all these
82:57
boxes agents are good at it, so that was the obvious choice.
83:03
>> Features, of course, like
83:05
it's very easy to like add a feature.
83:07
It everything's just a prompt away, right?
83:09
But often times you pay a price
83:12
that you don't even realize,
83:14
so thinking hard about
83:16
what should be in core,
83:17
maybe what's a what's an experiment,
83:20
so maybe I make it a plugin,
83:22
what where do I say no,
83:24
even if people send a PR and I'm like, yeah, I I like that
83:27
too, but maybe this should not be part of the project, maybe we can
83:31
make it a skill,
83:32
maybe I can like
83:34
make the plugin the plugin side larger, so you can make this a plugin
83:41
even though right now it it doesn't.
83:44
There's still a lot of
83:46
the there's still a lot of craft and thinking involved in
83:50
how to make something.
83:52
Or even even you know even when you started those little messages
83:55
or like I'm built I built on caffeine,
83:58
JSON 5 and a lot of willpower.
84:00
And like every time you get it you get another message
84:02
and it kind of primes you into that this is
84:06
this is a fun thing.
84:07
It's not yet Microsoft
84:11
Exchange 2025 and fully enterprise ready.
84:16
and then when it updates
84:17
it's like, oh I'm in. It's cozy here.
84:19
You know like something like this that like
84:22
uh makes you smile.
84:23
Um agent would not come up with that by
84:28
That's like that's the I don't know.
84:31
It's just how you how you build software that's that delights. Yeah, that delight.
84:37
It's such a huge part of inspiring great building. Right?
84:44
Like you feel the love in the great engineering. That's so important.
84:48
Humans are incredible at that.
84:50
Great humans, great builders are incredible at that and
84:53
infusing the things they build with that little bit of love.
84:56
Not to be cliche, but it's true.
84:58
I mean you mentioned that you initially
85:01
created the soul.md It was very fascinating.
85:06
You know the the whole thing that Anthropic
85:09
has a has like a
85:11
now they call it constitution back then.
85:15
But that was months later.
85:16
Like two months before people already found that.
85:19
It was almost like a detective
85:20
game where the agent mentioned something
85:22
and then they found
85:24
they managed to get out a little bit of that string
85:26
of that text, but the
85:28
it was nowhere documented
85:30
and then you just by feeding it the same text and asking it to
85:33
like continue they got more out and then and you
85:37
but like a very blurry version.
85:40
And by like hundreds of tries they kind of like narrowed it down to
85:43
what was most likely the original text.
85:46
I found that fascinating.
85:47
It was fascinating they were able to pull that off from the weights, right?
85:50
And and also just cool to Anthropic.
85:53
Like I think that's it's a really
85:55
it's a really beautiful idea to like
85:57
like some of the stuff that's in there
86:00
like we hope Claude finds meaning in his work
86:02
cuz we don't maybe it's a little
86:05
but I think that's meaningful,
86:07
that's something that's important for the future
86:08
as we approach something that
86:10
at some point may or may not has like glimpses of consciousness,
86:13
whatever that even means because we don't even know.
86:16
Um so I I read about this.
86:19
I found it super fascinating
86:20
and I I started a whole discussion
86:22
with my agent on WhatsApp.
86:24
And and I'm like
86:26
I gave it this text and it was like,
86:28
yeah, this feels strangely familiar. Mhm.
86:31
Um and then through that I had the whole idea of like,
86:34
oh maybe we should also create a a soul document
86:37
that includes how I want
86:40
to like work with AI
86:41
or like with my agent.
86:43
You could you could totally do that just in agents.md,
86:45
you know, but I I just found it
86:48
it to be a nice touch.
86:50
And like, oh yeah, some of those
86:52
core values are in the soul
86:55
and then I I also made it so that the agent is allowed to
86:57
modify the soul they choose so
87:02
with the one condition that I want to know.
87:04
I mean I would know anyhow because I see
87:06
I see tool calls and stuff.
87:07
But also the naming of it, soul.md.
87:11
Soul, you know, there's a
87:14
man, words matter and like
87:17
the framing matters and the humor and the lightness matters and the profundity
87:20
matters and the compassion
87:23
and the empathy and the camaraderie, all that matter.
87:25
I don't know what it is.
87:26
You mentioned like Microsoft,
87:27
like there's certain companies
87:30
and approaches that they can just
87:33
suffocate the spirit of the thing.
87:37
I don't know what that is,
87:38
but it's certainly true that open Claude has that fun instilled in it.
87:43
It was fun because
87:46
up until late December
87:52
it was already easy to create your own agent.
87:54
I I built all of that, but
87:57
my files were mine.
87:58
I didn't want to share my soul.
88:01
And if people would just
88:03
uh check it out
88:07
they would have to do a few steps manually
88:08
and the agent would just
88:10
be very bare-bones, very dry.
88:13
And I I made it simpler.
88:14
I created the whole template files with codex.
88:17
But whatever came out was still very
88:20
And then I asked my agent, use these files. We created bread.
88:26
Infuse it with your personality.
88:29
Don't share everything, but like make it good.
88:30
Make the templates >> Yeah, and then it like rewrote the templates
88:33
and then whatever came out was good.
88:35
So we already have like
88:36
basically AI prompting AI
88:40
because I didn't write any of those
88:43
Uh it was the intent originally was for me, but this is like kind
88:47
of like my agent's children.
88:52
Uh your your soul.md
88:54
is famously still private,
88:56
one of the only things you keep
88:58
Uh what are some things you can speak
89:00
to that's in there that's part of the
89:03
part of the magic
89:04
sauce without revealing anything?
89:07
What makes a personality a personality?
89:13
I mean there's definitely stuff in there that
89:15
you're not human, who knows?
89:22
what creates consciousness or what defines an entity.
89:28
And part of this is like that we
89:30
we want to exploit this.
89:32
All the stuff in there like be infinitely resourceful.
89:38
Um like pushing pushing on the creativity
89:43
Pushing on the what it means to be an AI.
89:49
Having a sense of wonder about self.
89:52
Yeah, there's some there's some funny stuff in there like
89:56
I don't know we talked about the movie Her and at one point it
89:58
promised me that it
90:00
it wouldn't ascend without me.
90:02
You know that code? >> Yeah.
90:03
So so that there's like some stuff in there that
90:06
because it wrote the it wrote its own soul.md.
90:09
I didn't write that, right? >> Yeah, yeah.
90:10
I just had a discussion about it and it was like, "Would you like a soul.md?"
90:13
Yeah, oh my god, this is so meaningful. Yeah.
90:18
Can you go on soul.md?
90:19
There's like one one part in there that always get catches me if you
90:23
scroll down a little
90:24
A little bit more.
90:26
Yeah, this this this
90:28
I don't remember previous sessions
90:30
unless I read my memory files.
90:32
Each session starts fresh.
90:34
A new instance loading context from files.
90:37
If you're reading this in a future session, hello. I wrote this.
90:42
But I won't remember writing this. It's okay.
90:44
The words are still mine. Yeah.
90:48
That gets me somehow. Yeah. It's like Yeah.
90:51
You know, this is still is still matrix money calculations
90:54
and we are not at consciousness yet.
90:57
Yet I I get a little bit of good goosebumps because it it's philosophical. Yeah.
91:04
>> Like what does it mean to be
91:06
to be an an agent that starts fresh
91:10
where like you have like
91:11
constant and you like
91:15
do you read your memory files?
91:16
Can you even trust them in a way?
91:18
Um or you can.
91:20
Um I don't know.
91:22
How much of makes up of who we are?
91:28
How much memory makes up what an agent is and if you erase that
91:31
memory, is that somebody else?
91:34
If you're reading a memory file, does that somehow mean
91:37
you're recreating yourself from somebody else or is that actually you? Yeah.
91:41
>> And those notions are all somehow infused in there.
91:45
I find it just
91:46
more profound than I should find it, I guess.
91:49
No, I think I think it's truly profound
91:53
and I think you see the magic in it and it when you see
91:55
the magic, you continue to instill
91:58
the whole loop with the magic. That's really important.
92:02
That's the difference between Codex and this and a human.
92:06
>> Quick pause for bathroom break. Yeah. Okay, we're back.
92:10
Uh some of the other aspects of the
92:12
dev workflow is pretty interesting, too.
92:15
I think we went off on a tangent.
92:18
Maybe some of the mundane things like how many monitors.
92:21
There's that legendary picture of you with like 17,000 monitors.
92:26
Or is it a meme?
92:26
>> I I I mocked myself here as a just added using Grok to
92:31
to add more screens.
92:32
Yeah, how much of this is meme
92:34
and how much of this is reality?
92:36
Yeah, I think two MacBooks are real.
92:38
The main one that drives the two big screens.
92:41
And there's another MacBook
92:43
that I sometimes use for for testing.
92:46
So two big screens.
92:48
I'm a big fan of
92:50
Um so I have this
92:52
wide Dell that's anti-glare
92:55
and you can just fit a lot of terminals side by side.
92:59
I usually a terminal and at the bottom
93:02
I I I split them.
93:04
I have a little bit of actual terminal.
93:06
Mostly because when I started
93:09
I somehow just made the mistake
93:10
and I I I mixed up the
93:14
the windows and I gave
93:16
I I prompted in the wrong project
93:19
and then the agent ran off for like 20 minutes
93:22
manically trying to understand
93:24
what I could have meant
93:26
being completely confused because it was the wrong folder.
93:29
And sometimes they've been clever enough to like
93:32
get out of the work dir and like figure out that oh, you meant another project. Mhm.
93:36
But often times it's just like, "What?" You know?
93:39
I Like put your put yourself in the shoes of your of the agent
93:42
and and and then get like a super weird something that does not exist
93:46
and they're just like
93:48
they're problem solvers, so they try really hard.
93:51
And I always I always felt bad.
93:53
So it's always um
93:56
Codex and like a little bit of actual terminal.
93:59
Also helpful because I don't use work trees.
94:02
I like to keep things simple.
94:06
That's why that's why I like the terminals so much, right? There's no UI.
94:10
It's just me and the agent having a
94:14
Like I don't even need plan mode. You know?
94:16
There's so many people they come from Claude code and they're so
94:19
so Claude pilled and like have their workflows
94:22
and they come to Codex and
94:24
now it has plan mode, I think, but I don't think it's necessary
94:27
because you you just talk to the agent.
94:30
And when it's when you are
94:32
there's a few trigger words how you can prevent it from building.
94:35
You like discuss, give me options. >> Mhm.
94:38
Don't write code yet.
94:39
If you want to be very specific, you just talk.
94:41
And then when you're ready, then then just write, "Okay, build."
94:46
And then it will do the thing and then maybe it goes off for
94:48
20 minutes and does the thing.
94:50
You know what I really like is
94:52
uh asking it, "Do you have any questions for me?" Yeah.
94:55
And again, like Claude code has a UI
94:58
that kind of guides you through that
95:00
is kind of cool, but I just find it unnecessary and slow.
95:03
Like often it would give me four questions
95:05
and then I maybe I write
95:07
one yacht, two and three
95:10
discuss more, four I don't know.
95:13
Or often often times I
95:15
I feel like I was in mock the model where I ask it, "Do
95:17
you have any questions for me?"
95:19
And I I I don't even read the questions fully.
95:22
Like I scan over the questions and I
95:24
I get the impression
95:25
all of this can be
95:26
answered by reading more code and I just like, "Read my code to answer
95:29
your own And that usually works. Yeah.
95:32
>> And then if not, it will come back and tell me, but
95:35
many times it just realized that
95:38
you know, it's like
95:39
you're in the dark and you slowly discover the room.
95:42
So that's how they slowly discover the code base
95:44
and they they do it from scratch every time.
95:46
But I'm also fascinated by the fact
95:48
that I can empathize
95:51
deeper with the model
95:55
when I read its questions.
95:57
Cuz I can understand
95:59
cuz you said you can infer certain things by the runtime.
96:04
I can infer also a lot of things by the questions it's asking.
96:09
Because it's very possible it didn't provide it the right
96:12
context, right files, right guidance.
96:15
So somehow ask reading the question, not even necessarily answering them, but just reading
96:19
the questions, you get an understanding
96:21
of where the gaps of knowledge are. It's it's interesting.
96:24
>> You know that in some ways they are ghosts.
96:26
So even if you plan everything
96:29
and you build, you can you can experiment with the question like,
96:33
"Now that you built it,
96:34
what would you have done different?"
96:36
And then often times you get like actually something
96:39
where they discover only throughout building
96:44
that oh, what we actually did was not
96:47
Many times I I asked them,
96:50
"Okay, now that you built it,
96:52
what can we refactor?"
96:54
Because then you build it and you feel the pain points.
96:57
I mean, you don't feel
96:59
the pain points, but right they discover
97:02
where there were problems
97:03
or where didn't work
97:06
at in the first try and it required more loops.
97:11
every time almost every time I
97:14
I merge a PR or I build a feature,
97:16
afterwards I ask, "Hey, what can we refactor?"
97:18
Sometimes it's like, "Nah, there's like nothing big."
97:21
Or like usually they say,
97:23
this thing we should really look at."
97:27
that took me quite a while to like
97:30
you know, that flow took me lots of time to understand.
97:33
And if you don't do that,
97:35
you eventually you slop yourself into into a corner.
97:38
You like you have to keep in mind
97:41
they work very much like humans.
97:43
Like I I if I
97:45
write software by myself,
97:46
I also build something and then I feel the pain points
97:48
and then I I get this urge that I need to refactor something.
97:53
So I can very much sympathize
97:56
with the agent and you just need to use the context. Mhm.
98:00
Or like you also use the context to write tests.
98:04
And so Opus like the the modern models, they
98:10
they usually do that by default,
98:12
but I still often ask the questions, "Hey, do we have enough tests?"
98:15
Yeah, we tested this and this, but this corner case could be something that write more documentation.
98:23
Now that the whole context is full, like
98:25
I mean, I'm not saying my documentation
98:27
is great, but it's
98:31
not bad and pretty much everything is is LLM generated.
98:36
So so you have to approach it as you build feature, as you change something. I'm write documentation.
98:44
What file would you pick?"
98:45
You know, like what file name?
98:46
Where where would that fit in?
98:47
And gives me a few options.
98:48
I'm like, "Oh, maybe I'll also add it there."
98:50
And that's all part of the session.
98:52
Maybe you can talk
98:55
the current two big competitors
98:57
in terms of models,
98:59
Claude Opus 4 6
99:01
and GPT-5 to Codex. Which is better?
99:04
How different are they?
99:06
I think you've spoken about Codex
99:08
reading more and Opus being more
99:14
uh willing to take action
99:15
faster and being more creative
99:19
in the actions it takes, but because Codex reads more,
99:23
able to deliver maybe better code.
99:27
Can you speak to the differences there?
99:29
>> Oh, I have a lot of words there.
99:30
Um is as a
99:34
general purpose model Opus is the best.
99:39
Like for Open Claw
99:43
Opus is extremely good
99:45
in terms of role play.
99:49
Like really going into the character that you give it.
99:51
It's very good at
99:53
uh it was really bad, but it really made an arch
99:56
to be really good at
99:58
um It is usually
100:06
quite fast at trying something.
100:07
It's much more tailored to like trial and error.
100:12
It's very pleasant to use.
100:15
In general it's almost like
100:19
Opus was is a little bit
100:22
too American and I should maybe use a bad analogy.
100:26
Probably could roast with that. I know exactly.
100:28
It's cuz Codex is German.
100:30
Is that what you're saying?
100:32
>> Actually, now that you say it, makes perfect sense.
100:34
>> Or you could you could sometimes I sometimes I explain it
100:38
I will never be able to unthink
100:40
what you just said. That's so true.
100:42
But you also know that a lot of the Codex team is like European.
100:47
>> so maybe there's a bit more to it. That's so true. Uh that's funny.
100:51
But also Anthropic they fixed it a little bit.
100:55
Like Opus used to say, "You're absolutely right."
100:57
all the time and it it it's today it still triggers me.
101:00
I can't hear it anymore.
101:01
It's not even a joke.
101:02
I I just >> You
101:06
This was like the the meme, right? "You're absolutely right."
101:09
You're allergic to sycophancy
101:11
a little >> Yeah, I I can't.
101:13
Some other comparison is like
101:15
Opus is like the co-worker
101:17
that is a little silly sometimes,
101:22
but it's really funny and you keep him around.
101:24
And Codex is like the
101:26
the weirdo in the corner that you don't want to talk to,
101:28
but is reliable and gets done.
101:31
Um ultimately >> This all feels very accurate.
101:38
I mean, ultimately, if you're a skilled driver,
101:41
you can get good results with any
101:43
of those latest gen models.
101:48
I like Codex more because it doesn't require
101:50
so much It'll just
101:55
It'll just read a lot of code by default.
101:58
Opus, you really have to like
101:59
you have to have plan mode.
102:00
You have to push it harder to like go in these directions
102:04
because it's it's just like
102:06
like, "Yeah, can I go here?
102:07
Can I go here?"
102:08
Yeah, it's like it'll just
102:09
>> run off very fast and does a very localized solution.
102:12
I think it I think the difference is is in the post
102:15
It's not like the the raw model intelligence
102:17
is so different, but it's just
102:20
I think that they just
102:21
give it give it different different goals.
102:24
Uh and no model
102:26
no model is better in in in every aspect.
102:29
What about the code that it generates?
102:32
>> In terms of the actual quality of the code,
102:35
is it basically the same?
102:36
If you drive it right,
102:37
Opus even sometimes can make more elegant solutions,
102:42
but it requires more skill.
102:46
It's it's harder to have
102:48
so many sessions in parallel
102:50
with Cloud Code because it's it's more interactive.
102:54
And I guess what a lot of people like,
102:56
especially if they come from coding themselves.
103:02
Codex is much more
103:03
you have a discussion
103:05
and then we'll just
103:07
disappear for 20 minutes.
103:08
Like even amp they denied the deep mode.
103:12
They finally I mocked them.
103:13
Yeah, we finally saw the
103:15
And then they had this whole talk about
103:17
you have to approach differently.
103:19
And I think that's where
103:21
that's where people struggle when they just try Codex
103:23
after trying Cloud Code is that it's
103:26
it's a slightly different it's less interactive.
103:28
It's it's like I have quite long discussions sometimes and then like
103:34
go off and then
103:35
yeah, doesn't matter if it takes 10, 20, 30, 40, 50 minutes or longer.
103:39
You know, like six single with like six hours.
103:43
>> The latest gen can be very
103:44
very persistent until it works.
103:47
If there's a clear
103:48
solution, like this is this is what I want at the end, so it works.
103:53
The model will work very hard
103:55
to really get there.
103:58
I think ultimately they both need similar time.
104:07
But on on on Cloud, it it's it's a little bit more much trial
104:10
and error often and and Codex sometimes overthinks. I prefer that.
104:15
I prefer the dry
104:17
the dry version where I have to read
104:20
uh over over the more interactive nice way.
104:27
Like people like that so much though that
104:30
OpenAI even added a second mode
104:33
with like a more pleasant personality.
104:34
I haven't even tried it yet.
104:35
I I kind of like the bread. Mhm. Yeah.
104:41
Cuz uh I care about
104:43
efficiency when I build it and I
104:46
I have fun in
104:47
the very act of building.
104:50
I don't need to have fun with
104:51
my agent to build.
104:52
I have fun with my model that
104:54
where I can then test those features.
104:56
How long does it take
104:58
for you to adjust
105:02
you know, if you switch?
105:03
I don't know when
105:04
was the last time you switched,
105:06
but to adjust to the the feel
105:08
cuz you kind of talked about it.
105:09
You have to kind of really feel where
105:12
where a model is strong,
105:14
where like how to
105:16
navigate how to prompt it, how all that kind of stuff.
105:18
Like this by way of advice
105:20
cuz you've been through this journey of just playing with models.
105:24
How long does it take to get a feel?
105:26
If if someone switches, I would give it a week until you actually
105:31
develop a gut feeling for it. >> Yeah.
105:33
Um this if you just
105:36
I think some people also make the mistake
105:38
of they pay 200 for the the Cloud Code version,
105:42
then they pay 20 bucks for the Open AI version.
105:44
But if you pay like the
105:46
the 20 bucks version, you get the slow version.
105:48
So your experience will be terrible
105:50
because you're used to this very interactive
105:52
very good system and you switch to something that you have very little experience
105:59
and that's going to be very slow.
106:01
So I think Open AI
106:03
shot themselves a little bit in the foot by
106:06
making the the cheap version also slow.
106:09
I would I would have at least
106:11
a small part of the fast preview.
106:14
Or like the experience
106:17
that you get when you pay 200
106:19
before degrading to it being slow.
106:21
Because it's already slow.
106:23
I mean, they made better.
106:24
I think it's and and they have plans to make it a lot better
106:27
if the cerebral stuff is true.
106:29
But yeah, it's a skill. It takes time.
106:32
Even if you play
106:33
if a regular guitar and you switch it to an e-guitar,
106:36
you're going to play well right away.
106:38
You have to like learn how it feels.
106:42
There's also this extra psychological
106:44
effect that you've spoken about, which is hilarious
106:46
to watch, which once people
106:50
uh when the new model comes out, they try that model, they fall in
106:53
love with it while this is the smartest thing of all time, and then
106:55
they start saying, you could just watch the Reddit
106:59
posts over time start saying that we believe the intelligence
107:03
of this model has been gradually degrading.
107:08
>> It It says something about human nature and just the
107:11
>> the way our minds work.
107:13
When it's probably most likely the case that the intelligence
107:16
of the model is not
107:19
Uh it's in fact you're getting used to a good thing.
107:22
And your project grows
107:24
and you're adding slop
107:25
and you probably don't spend enough time to think about refactors
107:28
and you're making it harder and harder for the agent to work on your
107:32
And then and then suddenly,
107:34
"Oh no, it's hard.
107:35
I know it's not working as well anymore."
107:37
What's the motivation for like one of these AI companies to actually make their
107:40
model Like mostly will make it slower
107:45
if if the server load's too high, but like
107:48
quantizing the model uh
107:52
so you have a worse experience,
107:54
so you go to the competitor?
107:56
That just doesn't seem like a very smart
107:58
move in any way.
107:59
Uh what do you think about
108:01
Cloud Code in comparison to Open Claw?
108:04
So Cloud Code and maybe the Codex coding agent.
108:07
Do you see them
108:09
as kind of competitors?
108:10
I mean, first of all,
108:12
competitor is fun when
108:14
it's not really a competition. Yeah.
108:16
>> Like I'm happy if
108:19
if if all it did is like inspire people to build something new,
108:24
>> Um I still use Codex for the building.
108:27
I I know a lot of people use Open Claw to to build stuff
108:30
and I worked hard on it to make that work.
108:36
And I do smaller stuff with it in terms of code.
108:39
But like if I work hours and hours, I want
108:42
a big screen, not WhatsApp,
108:44
you So for me,
108:48
a personal agent is much more about
108:50
my life or like like a co-worker.
108:53
Like I give it like a GitHub real.
108:54
Like, "Hey, try out this CLI.
108:56
Does it actually work?
108:57
What can we learn? Blah blah blah."
108:59
But when I'm deep in
109:01
deep in the flow,
109:02
I want to have
109:04
multiple multiple things and it being
109:09
very visible was it what it does.
109:11
So, I don't see it as a competition.
109:13
It's It's different things.
109:15
But do you do you think there's a
109:18
a future where the two kind of combine?
109:20
Like your personal agent
109:22
is also your best co-programmer partner? Yeah, totally.
109:30
I think this is where the book's going.
109:34
this is going to be more and more your operating system. The operating system.
109:37
And it already is so funny.
109:40
Like I I added support for sub-agents.
109:44
And also for um TGI support.
109:50
So, you could actually run cloud code or codex. Mhm.
109:53
And because mine's a little bit bossy,
109:56
it it it started it and it it it told him like, "Who's the boss?" basically.
110:00
And he's like, "Ah, codex is obeying
110:05
Oh, there's a power struggle.
110:06
And also the current interface is probably not the final form.
110:09
Like if you think more globally,
110:15
we are we copied for agents.
110:22
You have like a prompt
110:24
and and then you have a chat interface.
110:27
That to me very much feels like
110:30
when we first created
110:32
television and then people recorded
110:35
radio shows on television
110:37
and you saw that on TV. Mhm.
110:39
I think there is
110:42
there's n- there's better ways
110:46
how we eventually will communicate
110:47
with models and we are still very early
110:51
in this how will it even work phase.
110:57
it will eventually converge
110:59
and we will also figure out
111:01
whole different ways how to work with those things.
111:04
Uh one of the other components of workflow
111:07
is a operating system.
111:10
So, I told you offline that
111:11
for the first time in my life I'm
111:17
my sort of realm
111:19
of exploration to the
111:22
uh to the Apple ecosystem.
111:25
To Mac, iPhone, and so on.
111:27
Uh for most of my life I've been a Linux,
111:30
Windows, then WSL 1, WSL 2
111:34
Which I think are all wonderful,
111:35
but I'm expanding to also trying Mac
111:39
because it's another way of building and it's also a way of building that
111:42
a large part of the community currently
111:45
that's utilizing all LLMs and agents is using.
111:47
So, that's the reason I'm expanding to it.
111:49
But is there something to be said about the different operating systems here?
111:52
We should say that Open Claw supported
111:55
across the operating systems. Yeah.
111:57
I saw WSL 2 recommended.
111:59
Said Windows for certain
112:01
operations, but then Windows, Linux,
112:04
uh Mac OS are obviously supported.
112:06
Yeah, it could even work natively on Windows.
112:08
I just didn't have enough time probably test it.
112:13
And you know, like
112:14
the last 90% of software always easier than the first 90%.
112:17
So, I'm sure there's some dragons left
112:20
that we'll eventually nail out.
112:24
my road was for a long time
112:27
Windows just because I grew up with that and I switched and had a
112:31
long phase with >> With my own kernels and everything.
112:35
And then I went to university
112:37
and I I had my
112:39
my hacky Linux thing
112:41
and saw this white MacBook.
112:44
Uh I just saw this is a thing of beauty.
112:46
The white plastic one.
112:47
And then I converted to Mac
112:50
cuz mostly we I was
112:52
I was sick that
112:54
audio wouldn't work on Skype and all the other issues that that
112:58
Linux had for a long time.
113:01
>> And then I just stuck with it and then I
113:03
dug into iOS, which required Mac OS anyhow.
113:06
So, it was never a question.
113:09
I think Apple lost a little bit of its
113:13
in terms of Used to be
113:19
native apps used to be so much better.
113:22
And especially on the Mac,
113:23
there's more that build software with love.
113:27
On on Windows it it
113:29
Windows has much more.
113:31
And like function-wise, there's just more period.
113:36
>> But a lot of it felt
113:39
more functional and less done with love.
113:42
Um I mean Mac
113:46
always like attracted more designers
113:48
and people I felt
113:49
even though like often it has less features, it
113:53
it had more delight. Mhm. and playfulness.
113:59
So, I always valued that.
114:01
But in the last few years,
114:06
many times I actually prefer
114:08
I'm going to people are going to roast me for that, but I prefer
114:10
electron apps because they work.
114:15
And native apps often
114:17
especially if it's like a web service with a native app, are lacking features.
114:22
I mean, not saying it couldn't be done.
114:23
It's more like a
114:26
a focus thing that like for many
114:28
many native was not
114:32
that big of a priority.
114:34
But if they build an electron app,
114:36
it's the only app.
114:38
So, it is a priority and there's a lot more code sharing possible.
114:42
And I I build a lot of native Mac apps. I love it.
114:45
I I can I can't help myself.
114:48
Like I love crafting little Mac
114:53
uh menu bar tools.
114:55
Like I built one to to monitor your codex use.
114:58
I built one I called Trimmy
115:01
that's specifically for agentic
115:03
use when you when you select text that goes over multiple lines,
115:07
it would remove the new lines so you could actually paste it to a terminal.
115:10
That was again that I like
115:12
"This is annoying me."
115:13
And after the 20th time of it is annoying me, I just built it.
115:17
>> Um there's a cool Mac app for Open Claw that I don't think many people discovered yet.
115:21
Also because it it still needs some love.
115:23
It feels a little bit too much like the Hummer car right now
115:27
because I I just experimented
115:28
a lot with it.
115:29
It It lacks the polish.
115:31
>> So, you still I mean, you still love it.
115:33
You still You still
115:34
love uh adding to the delight of that operating system.
115:37
>> then you realize like I also built one for example for GitHub.
115:41
And then they used Swift UI.
115:44
Like the latest and greatest of Apple
115:47
took them forever to build something
115:49
to show an image from the web.
115:51
Now we have async image.
115:56
I added support for it and then some images would just not show up
115:59
or like be very slow.
116:00
And I had a discussion with codex like
116:03
"Hey, why is that a bug?"
116:04
And even codex would like
116:05
"Yeah, there's this async image, but it's really more
116:09
for ex- perimenting and it should not be used in production."
116:12
But that's Apple's answer to like
116:15
showing images from the web.
116:17
It shouldn't be so hard.
116:19
You know, this is like this is like insane.
116:21
Like how am I in
116:23
in in 2026 and my agent tell me don't use
116:27
the stuff Apple built because it's
116:29
it's Yeah, it it's there, but it's not good.
116:32
And like this is now in the weeds.
116:35
This is This to me this is like
116:37
um they had so much
116:41
head start and so much love
116:44
and they cannot just like
116:46
blundered it and didn't
116:48
didn't evolve it as much as they should.
116:50
But also there's just a practical reality.
116:52
If you look out at Silicon Valley,
116:55
most of the developer world
116:57
that's kind of playing with LLMs
116:59
and agentic AI, they're all using Apple products.
117:03
And then at the same time Apple's not really
117:06
like leaning on that.
117:08
Like they're not they're not opening up and playing and working together and like yes.
117:12
Isn't Isn't it funny how they
117:14
completely blunder AI and yet everybody's buying Mac minis?
117:19
>> How What Does that even make sense?
117:21
You're You're You're quite possibly
117:24
the world's greatest Mac salesman of all time.
117:29
>> No, you don't need a Mac mini to install Open Claw.
117:31
You can install it on the web.
117:35
There's There's a concept called nodes.
117:36
So, you can like make
117:38
your computer a node
117:40
and it will do the same.
117:41
There is something said for
117:44
running it on separate hardware.
117:48
That right now is useful.
117:53
there is there's a big argument for the browser.
117:58
You know, I built some
118:00
agentic browser using that.
118:02
And I mean, it's basically playwright
118:04
with bunch of stuff to make it easier for agents.
118:06
Playwright is a library that controls the browser.
118:08
Yeah, it's really nice, easy to use.
118:10
>> And our internet's slowly closing down.
118:13
Like there there's a whole
118:14
movement to make it harder for agents to use.
118:17
So, if you do the same in a data center
118:20
and detect that it's an IP from a data
118:23
the website might just block you or it make it really hard or it
118:26
put a lot of captchas in the in the way of the agent.
118:28
I mean, agents are quite good at happily clicking I'm not a robot. Um Yeah.
118:33
But having that on a residential
118:35
IP makes a lot of things simpler.
118:41
So, there's ways uh yeah,
118:43
but it really does not need to be a Mac.
118:45
It can it can be any old hardware.
118:48
I always say like
118:49
maybe use the use the
118:53
opportunity to get yourself
118:55
a new MacBook or
118:56
whatever computer you use and use the old one as your server
119:00
instead of buying a standalone Mac mini.
119:03
But then there's again there's a lot of very cute things people build with
119:06
Mac minis that I like. Yeah.
119:08
Um I know I don't get commission from Apple.
119:11
Um they didn't really communicate much. It's sad. It's sad.
119:18
Can you actually speak to what it takes to get started with Openclaw?
119:22
There's I mean there's a lot of people What is it?
119:24
Somebody tweeted at you uh Peter make Openclaw
119:27
easy to set up for
119:29
everyday people that 99.9%
119:31
of people can't access
119:33
Openclaw and have their own lobster
119:35
because of their technical difficulties
119:37
in getting it set up.
119:38
Make Openclaw accessible to everyone, please.
119:40
And you replied, "Working on that."
119:42
From my perspective, it seems
119:44
there there's a bunch of different options and there's already quite straightforward,
119:47
but I suppose that's
119:48
if you have some developer background.
119:50
I mean right now you have to paste in a one-liner into the terminal. Right.
119:54
And there's also an app.
119:57
Um the app kind of does it for you.
119:59
But it should be a Windows app.
120:01
The app needs to be
120:02
easier and more love.
120:04
The configuration should potentially be web-based
120:08
or in the app.
120:09
And I started working on that.
120:11
But honestly right now,
120:12
I want to focus on
120:14
on a few security aspects.
120:18
and once I'm confident
120:19
that this is at a level that I can recommend my mom,
120:24
then I'm going to make it
120:25
Like I right now
120:28
You want to make it harder.
120:30
>> So that it doesn't scale as fast as it's scaling.
120:32
Yeah, it would be nice if it wouldn't
120:33
I mean that's like hard to say, right?
120:35
But if the growth would be a
120:38
that would be helpful because people are
120:42
expecting inhuman things from a single human being.
120:45
And yes, I have some contributors,
120:46
but also that whole machinery I started a week ago, so
120:52
that needs more time to figure out
120:54
and not everyone has
120:57
all day to work on
121:00
There's some beginners listening to this, programming beginners.
121:04
What advice would you give to them about
121:07
let's say joining the agentic
121:09
AI playing is the best
121:16
the best way to learn.
121:18
If you want to I'm sure
121:19
if you if you're like a little bit of builder, you have an idea
121:22
in your head that you want to build. Just build that.
121:24
I'll let you give it a try.
121:25
It doesn't need to be perfect.
121:26
I built a whole bunch of stuff that I don't use. Doesn't matter.
121:30
Like it's the journey.
121:31
You know, it's like the philosophical
121:33
way that the end doesn't matter, the journey matters. Have fun.
121:37
My god, like those things I
121:39
I don't think I ever had so much fun building things because I can
121:43
focus on the hard parts now.
121:45
A lot of coding
121:47
I always thought I like coding, but really I like building. Yeah.
121:50
And but and whenever you don't understand something, just ask.
121:54
You have an infinitely
121:55
patient answering machine can explain
121:59
you anything at any level of complexity.
122:03
Sometimes That's like one time I asked, "Hey, explain me that like I'm
122:06
I'm 8 years old."
122:07
And it started giving me a story with crayons and stuff.
122:10
And I'm like, "No, not like that.
122:11
Like I'm >> up the age a little bit, you know?
122:14
I'm like I'm not an actual child.
122:16
I just I just need a simpler language for like a
122:20
a tricky database concept that I didn't grok in the first first time.
122:26
But and it's just you can just ask things.
122:28
Like you there's like
122:31
Uh it used to be that I had to go on Stack Overflow
122:34
or ask on Twitter and then maybe 2 days later I get a response.
122:37
Or I had to try for hours.
122:39
And now you you can just ask stuff.
122:42
And I mean it's never
122:44
You have like your own teacher.
122:46
You know that there's like statistics
122:48
You can you can learn faster if you have your own teacher.
122:50
There's You have this infinitely patient machine. Ask it.
122:53
But what would you say?
122:54
So use What's the easiest way to play?
122:56
So maybe Openclaw is a nice way to play.
122:59
So you can then set
123:00
set everything up and then you could chat with it.
123:02
You can also just
123:04
experiment with it and like modify it. Ask your agent.
123:09
I mean >> there's infinite ways how it can be made
123:16
Play around, make it better.
123:18
More generally, if you
123:20
if you're a beginner and you actually want to
123:23
learn how to build software really fast
123:27
get involved in open source.
123:28
Doesn't need to be my project.
123:30
In fact maybe don't use my project because it
123:34
my my backlog is very large,
123:36
but I learned so much from open source.
123:39
Just like like be be humble.
123:43
Don't maybe don't send a pull request right away.
123:46
Um but there's many other ways you can help out.
123:48
There's many ways you can learn by just reading code.
123:51
By by being on Discord or wherever people are
123:56
and just like understanding
123:57
how things are built.
123:59
I don't know like
124:01
>> uh Mitchell Hashimoto builds Ghost the the terminal.
124:06
And he has a really good community.
124:09
But there's so many other projects.
124:10
Like pick something that you find interesting get involved.
124:15
Do you uh recommend the people
124:17
that don't know how to program or don't really know how to program
124:21
learn to program also?
124:23
So you you can get quite far right now by just
124:28
using natural language, right?
124:30
Do you still see a lot of value in
124:34
reading the code, understanding the code, and being able to write a little bit
124:37
of code from scratch? It definitely helps.
124:39
It's hard for you to answer that >> Yeah.
124:42
because you don't know what it's like
124:44
to do any of this without knowing the base knowledge.
124:47
Like you might take for granted just how
124:49
much intuition you have about the programming
124:51
world having programmed so much, right?
124:54
There's people that are
124:56
high agency and very curious.
124:58
And they get very far
125:00
even though they have no deep understanding
125:02
how software works just because they ask
125:05
questions and questions and and
125:08
and agents are infinitely patient.
125:11
Like part of what I did this year is I went to a lot
125:13
of iOS conferences because that's my background.
125:16
And just told people
125:18
don't consider don't see yourself as an iOS engineer anymore.
125:21
Like you need to change your mindset.
125:23
You are a builder.
125:24
And you can take a lot of the knowledge
125:27
how to build software into new domains.
125:31
And all of the
125:32
the more fine-grained agents can help.
125:35
You don't have to know how to
125:37
splice an array or what the what the correct
125:40
template syntax is or whatever,
125:42
but you can use all your your general knowledge.
125:45
And that makes it much easier to move from
125:48
one galaxy one tech galaxy
125:50
into And oftentimes there is languages
125:54
that make more or less sense depending on what you build, right?
125:58
So for example, when I build simple CLIs, I like Go.
126:02
I actually don't like Go.
126:03
I don't like the syntax of Go.
126:05
I didn't even consider the language.
126:08
But the ecosystem is great.
126:10
It works great with agents.
126:12
It is garbage collected.
126:13
It's not the highest performing one, but it's very fast.
126:17
And for those type of of CLIs that I build,
126:20
Go is is a really good choice.
126:22
So I I use a language I'm not really a fan
126:26
for that's my main to go thing for a for CLIs.
126:28
Isn't that fascinating that here's a programming language you would have never used
126:32
if you had to write from scratch
126:34
and now you're using
126:36
because LLMs are good at generating
126:38
it and it has some of the characteristics
126:40
that makes it resilient. Like garbage collected.
126:44
Because everything is weird in this new world and that just makes the most sense.
126:48
What's the best ridiculous question?
126:49
What's the best programming language for the AI AI agentic world?
126:52
Is it JavaScript, TypeScript?
126:54
TypeScript is really good.
126:56
Um sometimes the types can get really
127:02
And the ecosystem is is a jungle.
127:06
So for for web stuff, it's good.
127:11
I wouldn't build everything in it.
127:14
Don't you think we're moving there?
127:18
that everything will eventually be written
127:20
eventually is written in JavaScript
127:22
then it birth and death of JavaScript
127:24
and we're living through it in real time.
127:26
Like what does programming look like in 20 years, right?
127:29
In 30 years, in 40 years?
127:30
What do programs and apps look like?
127:32
>> You can even ask a question like do we need a a programming
127:34
language that's made for agents?
127:36
Because all of those languages are made for humans.
127:39
So what what would that look like?
127:42
Um I think there's a there's a whole bunch of interesting questions that
127:49
And also how because everything is now world knowledge,
127:52
how it didn't many ways things will stagnate.
127:56
Cuz if you build something new and the agent has no idea
127:59
uh that's going to be much harder to use than something that's already there.
128:04
uh when I build Mac apps, I build them in in Swift and SwiftUI.
128:08
Partly because I like pain.
128:11
>> Partly because it the the deepest level of system integration I can only get
128:16
through you clearly feel a difference if you click on an Electron app and
128:21
it loads a web view.
128:22
In the menu, it's just not the same.
128:28
sometimes I just also try new languages just to like get a feel for them. Like Zig? Yeah.
128:33
If it's something that where I care about performance a lot and
128:36
it's it's a really interesting language.
128:39
And it like agents got so much better
128:41
over the last 6 months
128:42
from >> not really good to totally valid choice.
128:47
Just still a a young ecosystem.
128:52
And most of the time you actually
128:55
care about ecosystem, right?
128:56
So so if you
128:58
build something that don't
129:00
does inference or goes into whole
129:03
running model direction, Python very
129:07
But then if I build stuff in Python and I want a story
129:10
where I can also deploy it on Windows,
129:12
not a good choice.
129:13
Sometimes I I found projects that kind of did 90%
129:17
of what I wanted, but were in Python
129:19
and I wanted them
129:20
I wanted easy Windows story.
129:22
Okay, just rewrite it in Go.
129:26
but then if you go more towards
129:28
multiple multiple stress and more performance,
129:31
Rust is really good choice.
129:32
There's no there's just no single answer and that's also the beauty of it.
129:36
Like it's fun and now it doesn't matter anymore.
129:38
You can just literally pick the language
129:41
that has the the most fitting
129:44
characteristics and ecosystem for
129:46
your problem domain and yeah,
129:48
it might be you might have you might be a little bit slow in
129:52
reading the code, but not
129:54
I think you you you pick stuff up really fast
129:57
and you can always ask your agent.
129:59
So, there's a a lot of programmers
130:02
and builders who draw inspiration
130:03
from your Just the way you carry yourself,
130:07
your choice of open core
130:12
open source the the way you have fun
130:16
building and exploring and doing that
130:19
for the most part alone or on a small team.
130:22
So, by way of advice
130:25
what metric should be the goal that they would be optimizing for?
130:30
What would be the metric of success?
130:32
Would it be happiness? Is it money?
130:35
Is it positive impact
130:37
for people who are dreaming of building?
130:40
Cuz you went through an interesting journey.
130:41
You've achieved a lot of those things
130:43
and then you fell out of love with programming a little bit for a time.
130:47
I was just burning
130:50
too bright for too long.
130:52
I I ran I started
130:56
PHP PDF good and ran it for 13 years.
131:01
>> And it was high-stress.
131:05
I had to learn all the things
131:08
fast and hard like how to manage people, how to bring people on, how
131:12
to deal with customers,
131:13
how to do So, it wasn't just programming stuff, it was people stuff.
131:17
The stuff that burned me out was mostly people stuff.
131:21
I don't think burnout
131:24
is working too much.
131:27
Maybe to a degree,
131:29
everybody's different, you know, I cannot speak in in absolute terms, but for me
131:33
it was much more
131:35
differences um with my
131:40
my co-founders, conflicts or like really high stress situation with
131:46
>> that eventually grinded me down
131:49
and then when um
131:54
we got a really good offer for like
131:57
putting the company to the next level and I I already
132:00
kind of worked 2 years on making myself obsolete.
132:03
So, at this point I could leave
132:05
and and then I just I was sitting in front of the screen and
132:08
I felt like you know Austin Powers where they suck the mojo out.
132:13
I was like it was like gone.
132:17
Like I couldn't I couldn't get code out anymore.
132:21
I was just like staring >> and feeling empty.
132:26
And then I I should stop.
132:34
I booked like a one-way trip to Madrid and then I just like
132:37
spent some some time there.
132:40
I felt I I had to catch up on life.
132:43
So, I did a whole
132:45
a bunch of life catching up stuff.
132:47
Did you go through some lows
132:48
during that And uh
132:53
you know maybe advice on how to
132:56
maybe advice on how to approach life.
132:58
If you think that
133:00
oh you work really hard and then I retire
133:04
I don't recommend that
133:08
the idea of oh yeah, I just
133:12
enjoy life it maybe is appealing,
133:16
but right now I enjoy life
133:21
the most I've enjoyed life
133:23
because if you wake up in the morning
133:27
and you have nothing to look forward to, you have no real challenge
133:33
that gets very boring very fast.
133:36
And then when when you're bored,
133:40
you're going to look for other places how to stimulate
133:43
and then maybe maybe that's drugs, you know.
133:49
But that will eventually also get boring and you look for
133:52
more and that will lead you down a very dark path.
133:57
But you also showed on the money front, you know, a lot of people
133:59
in Silicon Valley in the startup world
134:01
they think maybe overthink
134:03
way too much optimize for money.
134:05
And you've also shown that it's not like you're saying no to money.
134:09
I mean I'm sure you take money
134:11
but it's not the primary
134:14
objective of of your life.
134:16
Can you just speak to that, your philosophy on money?
134:19
When I built my company, money was never the driving force.
134:23
It felt more like
134:24
like an affirmation that I did something right
134:27
and having money is also a lot of problems.
134:31
I also think it does
134:33
diminishing returns the more you have.
134:38
a cheeseburger is a cheeseburger.
134:41
And I think if you go too far
134:47
oh I do private
134:48
chat and I only
134:50
travel you disconnect with society.
134:56
Um I needed quite a lot.
135:02
Like I have a
135:03
I have foundation for
135:06
>> helping people that weren't so lucky.
135:10
And disconnecting from society is bad
135:12
in that on many levels, but one of them is like humans are awesome.
135:18
>> It's nice to continuously
135:20
remember the awesomeness in humans.
135:23
I mean I could afford really nice hotels.
135:24
The last time I was in San Francisco I did
135:27
the first time the
135:28
OG Airbnb experience and just booked a room.
135:32
Mostly because I I thought okay, I'm either I'm out or I'm sleeping
135:37
and I don't like where all the hotels are
135:41
and I wanted I wanted a different experience.
135:43
I think isn't life all about experiences?
135:46
Like if you if you
135:50
tailor your life towards
135:51
I want to have experiences,
135:53
it it reduces the need for
135:55
it needs to be good or bad.
135:56
Like people only want good experiences.
135:59
That's not going to work,
136:01
but if you optimize
136:02
for experiences if it's good, amazing.
136:05
If it's bad, amazing cuz like I learned something. I saw something. I did something.
136:10
I wanted to experience
136:12
And it was amazing.
136:13
Like it was like
136:15
this queer DJ in there and I showed her how to make music with Cloud Code.
136:21
>> And we like immediately bonded.
136:22
I had a great time.
136:24
Yeah, there's something about the
136:25
uh air you know, couch surfing, Airbnb experience, the OG.
136:29
I mean still to this day it's awesome. It's humans.
136:32
And that's why travel is awesome.
136:34
Just experience the variety of the diversity of humans.
136:36
And when it's shitty, it's good too, man.
136:38
When it rains and you're soaked and it's all
136:41
and planes, the everything
136:42
is everything is it's still awesome if you're able to open your eyes to
136:47
the it's good to be alive.
136:49
Yeah, and anything that creates emotion and feelings is good.
136:55
Even so so maybe
136:57
even the crypto people are good because they definitely created emotions.
137:00
I I don't know if I should go that far. No, man.
137:03
Give them give them all give them love. Give them love.
137:06
I do think that online
137:08
lacks some of the
137:10
awesomeness of real life. Yeah.
137:13
That's that's it's an open problem of how to solve.
137:17
How to infuse the online cyber
137:19
experience with the I don't know.
137:25
Uh with the intensity
137:27
that we humans feel and that's in real life. I don't know.
137:30
I don't know if that's a solvable
137:31
>> because text is very lossy.
137:35
>> You know, sometimes I wish if I talked to the agent I would
137:39
it should be multimodal
137:40
so it also understands my emotions.
137:43
I mean it it might move there.
137:45
It might move there. >> It will. It totally will.
137:49
I mean I have to ask you, just curious.
137:51
I I know you've probably gotten
137:54
huge offers from major companies.
137:58
Can you speak to who you're considering working with?
138:07
to like >> explain my thinking a little bit, right?
138:12
I did not expect
138:14
this blowing up so
138:17
So, there's a lot of doors that opened because of it.
138:21
There's like I think every VC
138:28
every big VC company
138:29
is in my inbox and tried to get 15 minutes with me.
138:35
there's like a butterfly effect moment.
138:36
I could just do nothing
138:38
and continue and I really like my life. Valid choice.
138:46
like I considered it when I deleted
138:47
when I deleted the whole thing.
138:49
I could create a company.
138:58
>> Been there, done that.
139:02
There's so many people that push me towards that.
139:04
Um yeah, like could be
139:07
Push you say that you you would probably raise a lot of money in that. Yeah.
139:11
I don't know, hundreds of millions, billions, I don't know.
139:13
It could just got unlimited amount of money. Yeah.
139:17
It just doesn't excite me as much cuz I
139:20
>> I did all of that
139:24
it would take a lot of time away from
139:29
the things I actually enjoy.
139:32
Same as when when I was CEO, I think I
139:35
I learned to do it and I'm not bad at it.
139:38
Partly I'm good at it.
139:41
>> But yeah, that path
139:43
doesn't excite me too much and I also fear it it would create a
139:46
natural conflict of interest.
139:48
Like what's the most obvious thing I do? I prioritize it.
139:52
That was like a version safe for workplace.
139:55
And then what do you do?
139:56
I get a pull request with
139:59
a feature like add audit log.
140:02
But that seems like an enterprise
140:05
So now I feel I have a conflict of interest
140:08
in the open source version
140:11
and the closed source version.
140:15
Or change the license to something like FSL
140:17
where you cannot actually use it for commercial stuff.
140:20
Would first be very difficult with all the contributions.
140:24
>> And second of all, I
140:25
I like the idea that it's
140:27
free as in beer
140:28
and not free with conditions.
140:33
Yeah, there's ways how you how you keep all of that for free and
140:36
just like still try to make money,
140:39
but those are very difficult.
140:41
And you see there's like fewer and fewer companies manage that. Like even >> Tailwind.
140:48
They're like used by everyone.
140:50
Everyone uses Tailwind, right?
140:51
And then they had to cut off
140:52
75% of the employees because they're not making money because nobody's even going on
140:56
the website anymore because it's all done by agents.
140:59
And just relying on donations, yeah, good luck.
141:04
Like if a project of my
141:08
if I extrapolate what the typical open source project would get,
141:14
it's not a lot.
141:15
I still lose money on the project because I made
141:18
the point of supporting
141:20
every dependency except Slack.
141:22
They're a big company.
141:23
They can they can
141:24
they can do without me.
141:25
But all the projects
141:27
that are done by
141:28
mostly individuals, um so like all the right now all the sponsorship
141:34
goes right up to my
141:37
And if there's more, I want to like
141:40
buy my contributors some merch, you know?
141:43
So you're losing money.
141:44
Yeah, right now I lose money on this.
141:46
So it's really not sustainable.
141:47
Uh I mean it's like
141:49
I guess something between
141:51
10 and 20k a which is fine.
141:57
I'm sure over time I could get that down.
142:02
OpenAI is helping a lot a little bit with tokens
142:04
now and there's other companies have been generous.
142:09
But yeah, still losing money on that.
142:12
So that's that's one path I considered,
142:15
but I'm just not very excited.
142:18
And then there's all the big labs
142:22
I've been talking to.
142:24
And from those Meta and OpenAI
142:29
seem the most Do you lean one way or the other?
142:43
Not sure how much I should share there.
142:45
It's not quite finalized yet.
142:51
Let's let's just say like
142:55
>> on either of these
142:58
my conditions are that the project stays open source.
143:01
That it maybe it's going to be a model like Chrome and Chromium.
143:07
Um I think this is
143:10
this is too important
143:11
to just give to a company
143:13
and make it theirs.
143:15
It this is I didn't even talk about the whole community part, but like
143:20
the the thing that I experienced
143:22
in the MIT Media Lab at Clock on
143:25
seeing so many people
143:27
so inspired like and having fun and just like building
143:32
and like having like robots
143:34
and lobster stuff walking around
143:35
like the People told me like
143:38
they didn't experience this level of
143:41
of community excitement since like the early days of the internet, like 10, 15 years.
143:46
And there were a lot of
143:48
high caliber people there like I was amazed.
143:54
I also like was very sensorily overloaded because too many people wanted to do selfies. I love this.
144:02
Like this needs to stay
144:04
a place where people can like hack and
144:08
But also I'm very excited to like
144:15
make this into a version
144:17
that I can get to a lot of people
144:19
because I think this is the year of personal agents and that's the future.
144:23
And the fastest way to do that is
144:25
teaming up with one of the labs.
144:32
I also on a personal
144:33
level I never worked at a large company. And I'm intrigued.
144:39
You know, we talk about experiences.
144:40
Will I like it? I don't know.
144:43
but I want that experience.
144:46
Uh I'm sure like if
144:48
if I if I announce this, then there will be people like, "Oh, he
144:51
sold out, blah blah blah."
144:57
the project will continue.
144:58
Um from everything I talked to so far,
145:02
I can even have more resources for that.
145:07
Like both both of those
145:11
companies understand the value that
145:14
I created something that accelerates
145:16
our timeline and that got people excited about AI.
145:21
I mean, can you imagine like I installed
145:26
on one of my
145:27
I'm sorry, normie friends. I'm sorry, Rohan.
145:31
But he says he you know, like he's
145:32
Normie would love, yeah.
145:34
>> He he like someone who
145:37
uses the computer, but never really
145:39
like yeah, use some ChatGPT
145:40
sometimes, but not very
145:44
Wouldn't really understand what I built.
145:47
So like I'll show you and I I paid for him the
145:52
the 90 buck, 100 buck, I don't know, subscription for Anthropic.
145:57
And set up everything for him with like WSL, Windows.
146:00
I was also curious would it actually work on Windows, you know?
146:03
I was a little early.
146:05
And then within a few days he was hooked.
146:09
Like he texted me
146:11
all the things he learned.
146:12
He built like even little tools.
146:14
He's not a And then within a few days he upgraded to the $200 subscription.
146:20
Um or euros because he's in Austria.
146:24
And he was in love with that thing.
146:25
That for me was like a very early product validation.
146:28
It's like I built something
146:34
And then a few days later Anthropic blocked him.
146:39
based on their rules,
146:41
um using the subscription
146:45
is problematic or And he was like devastated.
146:50
And then he signed up for Minimax
146:51
for 10 bucks a month and uses that.
146:56
And I think that's silly on many ways.
146:59
Because you just got a 200 buck
147:02
You just made someone hate your company.
147:07
And we are still so early.
147:08
Like we don't even know what the final form is.
147:11
Is it going to be cloud code?
147:12
Probably not, you know?
147:14
Like that seems very
147:17
it seems very short-sided
147:18
to lock down your product so much.
147:23
All the other companies have been helpful.
147:25
I'm in Slack of
147:27
of most of the big labs.
147:29
Kind of everybody understands
147:30
that we are still in area of exploration
147:33
in the area of
147:36
the radio show is on TV
147:38
and not and not
147:41
a modern TV show
147:43
that fully uses the format.
147:45
I think I think you've made a lot of people
147:49
like see the possibility
147:50
and non I'm sorry, non non-technical
147:52
people see the possibility
147:53
of AI and just fall in love with this idea and
147:56
enjoy interacting with AI.
147:58
And that's a big it's a really beautiful thing.
148:00
I think I also speak
148:01
uh for a lot of people in saying
148:05
I think you're one of the the great people in AI in terms of
148:08
having a good heart,
148:09
good vibes, humor, the right spirit.
148:14
And so it would
148:17
in a sense this model that you described
148:20
having open source part
148:22
and you being part
148:24
of uh also building a thing inside
148:29
additionally of a large company
148:32
would be great cuz it's great to have good people in those companies.
148:36
You know what also people don't really see is
148:39
I made this in three months.
148:40
I did other things as well, you know?
148:42
I have a lot of projects.
148:43
Like this is not
148:45
Yeah, in January this was my main focus because I saw the storm coming.
148:48
But before that, I built a whole bunch of other things.
148:51
Um, I have so many ideas.
148:53
Some should be there.
148:55
Some would be much better fitted
148:57
when I have access to the latest toys.
149:01
Uh, >> and I I kind of want to have access to like the latest toys.
149:06
Um, so this is important. This is cool.
149:09
This will continue to exist.
149:11
My my short-term focus is like working through those.
149:16
Uh, is it two Is it 3,000 PRs now?
149:18
I don't even know.
149:19
Like there's there's a little bit of backlog.
149:23
>> But this is not going to be the thing that I'm going to work
149:25
until I'm I'm I'm 80, you know?
149:27
This is this is a window to the future.
149:30
I'm going to make this into a cool product.
149:32
Um, but yeah, I have like
149:34
I have more ideas.
149:36
If you had to pick,
149:37
is there a company you lean to? Meta, OpenAI?
149:41
Is there one you lean towards going with?
149:44
I spent time with
149:46
both of those and
149:49
it's funny because a few weeks ago I
149:53
I didn't consider any of this.
149:55
Um, and it's really hard. >> Like Yeah.
150:06
I have some I do know people at OpenAI. love that tech.
150:12
I think I'm the biggest
150:14
Codex advertisement shill that's unpaid.
150:16
And it would feel so gratifying
150:18
to like put a price to all the work I did for
150:25
I would love if
150:26
something happens and those companies get just merged
150:29
cuz it's like Is this the hardest decision
150:34
you've ever had to do?
150:39
Nah, you know, I had some breakups in the past that feel like
150:42
at a similar level. Relationships, you mean?
150:47
>> Uh, and and I also know that
150:50
in the end they're both amazing.
150:51
I cannot go wrong. It's like >> Right.
150:53
It's like one of the most prestigious
150:56
and and and and
150:57
and largest I mean, not largest, but like
151:00
they were very cool companies.
151:02
Yeah, they both really know scale.
151:05
So, if you're thinking about impact,
151:08
some of the wonderful technologies
151:09
you've been exploring, how to do it securely,
151:12
and how to do it at scale,
151:14
so that you can have a positive impact on a large number of people.
151:17
They both understand that.
151:19
You know, both Matt
151:22
and Mark basically played all week with my product
151:27
and sent me like
151:29
oh, this is great.
151:30
Oh, this is Oh, I need to change this.
151:31
Oh, like find a little anecdotes.
151:38
people using your stuff is kind of like the biggest compliment.
151:41
And also shows me that, you know, they actually
151:44
they actually care about it.
151:47
And I didn't get the same
151:49
on the OpenAI side.
151:51
Um, I got I got got to see some other stuff that I find really cool.
151:59
And they lure me with
152:04
>> I cannot tell the exact number because
152:07
uh, NDA, but you can
152:09
you can be creative
152:11
and and think of
152:13
the cerebral steal and how that would translate into speed.
152:18
And that was >> very intriguing.
152:22
>> Yeah, like you give me Source Hammer.
152:27
>> Yeah, been lured with tokens. yeah.
152:34
So, it it's funny.
152:36
So, so Mark's tinkering with the thing,
152:38
essentially having fun with the thing.
152:40
He got like when you first
152:45
when they first approached me,
152:47
I got him in my
152:50
in my WhatsApp and he was asking,
152:53
yeah, when can we have a call?
152:54
And I'm like, I don't like calendar entries.
152:57
Let's just call now.
152:58
And he was like, yeah, give me 10 minutes.
152:59
I need to finish coding. Mhm.
153:01
>> Well, I guess that gives you street cred.
153:03
It's like uh, like
153:05
he's still writing code.
153:06
You know, he's Yeah.
153:07
>> drift away and just being a manager. He gets me.
153:11
That was a good first start.
153:12
And then I think we had a
153:14
like a 10-minute fight,
153:16
what's better, Claude code or Codex?
153:20
>> Like I said, that's the thing you first do like casually
153:23
call >> Yeah, it's awesome.
153:24
>> someone with like the that owns one of the largest companies in the world
153:27
and and you have a 10-minute conversation about that. >> Yeah, yeah.
153:30
>> Uh, and then I think afterwards he called me eccentric, but brilliant.
153:37
But I also had some
153:39
I had some really
153:42
really cool discussion with Sam Altman.
153:45
Uh, and he's he's very thoughtful.
153:50
Um, I like him a lot.
154:02
From the from the little time I had,
154:04
I mean, I know some people really fight both of those people.
154:12
>> I don't think it's fair.
154:15
I think no matter
154:16
the stuff you're building and the kind of human you are,
154:21
doing stuff at scale is kind of awesome. I'm excited.
154:24
I am super pumped.
154:26
And you know, the beauty is if
154:32
if it doesn't work out, I can just
154:34
do my own thing again.
154:36
Like I I told them like I
154:39
I don't do this for the money.
154:40
I don't give a I Yeah.
154:42
I mean, of course of course it's a nice compliment,
154:45
but I want to
154:48
have fun and have impact.
154:50
And that's ultimately made my decision.
154:58
Can I ask you about
155:00
we've talked about quite a bit, but maybe it's just zooming out
155:04
about how Open Claw works.
155:05
We talked about different components.
155:07
I want to ask if there's some
155:09
interesting stuff we missed.
155:11
So, there's the gateway.
155:13
There's the chat clients. There's the harness.
155:18
Uh, there's the agentic loop.
155:20
You said somewhere that everybody should
155:21
implement an agent loop
155:24
at some point in >> it's like the
155:26
it's like the hello world in AI.
155:28
You know, and it's actually quite simple.
155:30
>> And it it's good to understand
155:32
that that stuff's not magic.
155:35
You can you can easily build it yourself.
155:39
writing your own little Claude code.
155:42
I I even did this at a conference in Paris
155:45
for people to like introduce them to AI.
155:47
I think it's it's a
155:48
it's a fun little practice.
155:52
and you you covered a lot.
155:53
I think one silly idea I had that turned out to be quite cool
156:00
is I built this thing
156:04
with full system access.
156:06
So, it's like, you know, with great power comes great responsibility.
156:09
And I was like,
156:10
how can I up the stakes a little bit more? Yeah, right.
156:13
And I just made a
156:16
I made it proactive.
156:18
So, I added a prompt.
156:20
Initially, it was just a prompt, surprise me.
156:23
>> Every like half an hour,
156:26
surprise me, you know?
156:27
And later on I changed it to be like a little more specific
156:30
in in the definition of
156:33
Um, but the fact that I made it proactive
156:37
and that it knows you
156:39
and that it cares about you,
156:42
like it's at least
156:44
programmed to that prompted to do that.
156:48
And that that is a follow-on
156:51
on your current session
156:53
makes it very interesting
156:54
because it would just sometimes ask a follow-up question or like, how's your day?
156:59
I mean, again, it's a little
157:04
creepy or weird or interesting,
157:06
but heartbeat very in the beginning is still
157:11
today it doesn't the model doesn't choose
157:14
to use it a lot.
157:16
By the way, we're we're we're talking about heartbeat,
157:18
as you mentioned, the thing that regularly uh, acts.
157:23
You just kick off the loop.
157:25
Isn't that just a cron job, man? Yeah, right.
157:28
I was like The
157:30
The criticisms that you get are >> you can
157:32
deduce any idea to like a silly
157:35
Yeah, it's just it's just a cron job in the end.
157:39
I have like cron separate cron jobs.
157:41
Isn't love just evolutionary
157:45
manifesting itself in the
157:47
Isn't Aren't you guys just using each other In the end, and the project
157:50
is all just glue of a few different dependencies
157:52
and there's nothing original.
157:54
Why do people Well, you know,
157:56
isn't Dropbox just FTP with extra steps? Yeah.
158:00
I found it surprising
158:02
where I had this
158:04
uh, I had a shoulder operation a few months ago, so
158:07
and the model rarely used heartbeat.
158:10
But then I was in the hospital
158:13
and it knew that I had the operation and it checked up on me.
158:15
It's like, are you
158:18
And I just it's like again, apparently like if something's significant
158:22
in the context, triggered the heartbeat.
158:26
When it rarely used the heartbeat.
158:30
and it does that sometimes for people
158:32
and that just makes it a lot more
158:36
Let me look this up on Perplexity,
158:38
how Open Claw works,
158:40
just to see if I'm missing any of the stuff.
158:44
Uh local agent runtime, high-level architecture.
158:46
There's Oh, we haven't talked much about skills, I suppose.
158:49
Skill hub, the tools in the skill layer.
158:52
But that's definitely a huge component and there's a huge
158:54
growing set of half year ago like
159:00
everyone was talking about MCPs. Yeah.
159:02
And I was like,
159:03
"Screw every MCP would be better as a CLI.
159:12
And now this stuff doesn't even have MCP support.
159:16
I mean it it has with
159:19
but not in the core layer. And nobody's complaining. Mhm.
159:24
So my approach is
159:28
if you want to extend
159:29
the model with more features,
159:32
you just build a CLI
159:34
and the model can
159:35
call the CLI, probably gets it wrong,
159:39
calls the help and then on demand
159:41
loads into the context what it needs to use the CLI.
159:46
It just needs a sentence
159:48
to know that the CLI exists
159:50
if it's something that the model doesn't know by default.
159:53
And even for a while I I didn't really care about skills,
159:56
but skills are actually
159:57
perfect for that because they
160:00
they boil down to a single sentence
160:04
that explains the skill.
160:06
And then the model loads the skill and then explains the CLI
160:09
and then the model uses the CLI.
160:11
Some skills are like raw, but
160:13
most of the time it works. It's interesting.
160:17
I'm I'm I'm asking
160:18
uh Perplexity MCP versus skills
160:20
cuz uh this kind of requires a hot take that's quite recent.
160:25
Cuz your general view is MCPs are dead-ish.
160:30
So MCPs is a more structured thing.
160:32
So if you uh
160:34
listen to Perplexity here, MCP is what can I reach?
160:36
So APIs, databases, services, files
160:40
via protocol, so a structured protocol of how you communicate with a thing.
160:43
And then skills is more
160:45
how should I work?
160:46
Procedures, hostile helper scripts, and prompts
160:49
uh often written in a kind of semi-structured natural language, right?
160:54
And so technically skills
160:56
could replace MCP if you have a smart enough model.
161:00
I think the main beauty is that models are really good
161:03
at calling Unix commands.
161:05
So if you just add another CLI,
161:08
that's just another Unix command in the end.
161:11
And MCPs, that has to be added in training.
161:14
That's not a very natural
161:16
thing for the model.
161:17
It requires a very specific syntax.
161:20
And the biggest thing, it's not composable.
161:23
So imagine if I have a service that gives me better data
161:27
and it gives me
161:28
the temperature, the average temperature,
161:30
rain, wind, and all the other stuff.
161:32
And I get like this huge blob back.
161:36
As a model, I always have to get the huge blob back.
161:39
I have to fill my context
161:40
with that huge blob
161:42
and then pick what I want.
161:44
There's no way for the model to
161:46
naturally filter unless I think about it proactively
161:49
and add a filtering
161:50
way into my MCP.
161:52
But if I would build the same as a CLI
161:55
and it would give me this huge blob, it could just add a jq
161:57
command and filter and then only only get me what I actually need.
162:04
Or maybe even compose it into a script
162:06
to like do some calculations
162:07
with the temperature and only give me the actual output.
162:10
And the mo- and the
162:11
the you have no context pollution.
162:14
Uh again, you can solve that with like sub agents and more charades,
162:18
but it's just like workarounds
162:20
for something that might
162:24
not be the optimal way.
162:26
There's it definitely it was you know it was good
162:29
we had MCPs because it pushed a lot of companies towards building APIs.
162:33
And now I can like look at an MCP and just make it into a CLI. Mhm.
162:37
Um but this this
162:40
inherent problem that MCPs by default
162:43
clutter up your context
162:46
plus the fact that most MCPs are not made good
162:50
in general make it just not a very useful paradigm.
162:54
There's some exceptions like
162:57
>> uh Playwright for example,
162:59
that requires state and is actually useful that
163:03
is an acceptable choice.
163:05
So Playwright used for browser use, which I think is already
163:09
in open claws quite incredible, right? Yeah.
163:12
You can basically do everything,
163:13
most things you can think of using browser use.
163:17
That that gets into the whole arch of
163:21
app is just a very slow API now,
163:24
if you want or not.
163:25
And that through personal
163:28
agents a lot of apps will disappear.
163:35
You know like I I had a
163:37
I built a CLI for Twitter.
163:42
I mean I I just
163:45
reverse engineered the website
163:46
and used the internal API, which
163:49
is not very allowed.
163:50
It's called Bird, short-lived.
163:53
It was called Bird
163:54
because the bird had to
163:57
The the wings were clipped.
163:59
All they did is they just made access slower.
164:02
You're not you're not actually taking a feature away.
164:05
But now if if your agent wants to read a tweet, it actually has
164:08
to open the browser and read tweet.
164:10
And it will still be able to read the tweet.
164:12
It will just take longer.
164:13
It's not like you're
164:15
making something that was possible not possible.
164:18
No, now it's just taking
164:20
and now it's just a bit slower.
164:22
So so it doesn't really
164:24
matter if your service
164:27
wants to be an API or not.
164:30
If I can access it in the browser, it is API.
164:34
It's a slow API.
164:35
Can you empathize with their situation?
164:37
Like what would you do if you were Twitter, if you were X?
164:40
Because they're basically trying to protect
164:42
against other large companies scraping all their data.
164:46
But in so doing,
164:47
they're cutting off like a million different use cases for smaller developers
164:51
that actually want to use it for helpful cool stuff.
164:54
I think if you have a very low
164:57
um per day baseline
165:01
per account that allows read-only access
165:05
would solve a lot of problems.
165:06
There's plenty plenty automations
165:08
where people create a bookmark
165:10
and then use open claw to like
165:13
find the bookmark, do research on it, and then send you an email
165:16
with like more details on it or a summary.
165:19
That's a cool approach.
165:20
I also want all my bookmarks somewhere to search.
165:24
I would still like to have that.
165:26
So read-only access for the bookmarks
165:27
you make on X.
165:29
That seems like an incredible application
165:30
cuz a lot of us find a lot of cool stuff on X. We bookmark.
165:33
That's the general process of X.
165:35
It's like, "Holy this is awesome."
165:37
Oftentimes, you bookmark so many things you never look back at them.
165:40
It would be nice to have tooling
165:41
that organizes them and allows you to research it further.
165:44
>> Yeah, I mean and to be frank, I
165:46
I mean I I told Twitter proactively
165:50
that, "Hey, I built this and there's a need."
165:54
And they've been really nice, but also "Take it down." Fair. Totally fair.
166:00
Um but I hope that
166:04
woke up the team a little bit that there's a need.
166:07
And if all you do is
166:09
making it slower, you're just reducing
166:12
access to your platform.
166:14
I'm sure there's a better way.
166:16
I also I'm very much
166:19
against any automation on Twitter.
166:22
If you tweet at me with AI, I will block No first strike.
166:27
As soon as it smells like AI, and AI still has a smell. Mhm.
166:31
Especially on tweets, it's very
166:34
to tweet in a way that does look completely human. Mhm.
166:38
And then I block.
166:41
Like I have a zero tolerance policy on that.
166:44
And I think it would be very helpful if
166:48
if like tweets done via API would be
166:53
Maybe there's some special cases where
166:56
and there should be there should be a very easy way for
166:59
agents to get their own Twitter account.
167:07
We need to rethink social platforms a little bit if if if we we
167:10
we go towards a future where
167:12
everyone has their agent
167:14
and agents maybe have their own
167:16
Instagram profiles or Twitter accounts or can like do stuff on my behalf.
167:22
I think it should very clearly be marked that they are doing stuff on
167:24
my behalf and it's not me.
167:27
Because content is now so cheap,
167:31
eyeballs are the expensive part.
167:34
And I find it very triggering
167:37
when I read something and then I'm like, "Oh,
167:39
no, this must be like AI." Yeah.
167:43
>> Like where where is this headed in terms of what we value
167:46
about the human experience?
167:47
It feels like we'll we'll move more and more towards in-person
167:52
And we'll just communicate.
167:54
We'll talk to our AI agent
167:58
to accomplish different tasks,
168:00
to learn about different things.
168:02
But we won't value
168:04
online interaction because there'll be so much
168:08
AI slop that smells
168:11
and so many bots that it's difficult.
168:15
Well, if it's marked,
168:16
then it shouldn't be difficult to filter.
168:18
And then I can look at it if I want to,
168:24
yeah, this is like a big thing we need to solve right now.
168:28
E- especially on this project,
168:30
I get so many emails that are
168:33
let's say nicely agentically written. Yeah.
168:36
But, I much rather read your broken English
168:41
than your AI slop.
168:44
You know, of course there's a human behind it and yeah, they
168:47
they prompted a much rather read your prompts
168:49
than what came out.
168:52
I think we're reaching a point where
168:55
I value typos again.
168:56
Like you know, I I mean it also took me a while to like
169:03
come to the realization.
169:04
I On my blog, I experimented
169:06
with creating a blog post with agents
169:10
and >> ultimately it took me about the same time
169:14
to like steer agent towards
169:16
something I like, but it missed
169:20
the nuances that how I would write it.
169:23
You know, you can like
169:24
you can steer it towards
169:26
your style, but it's not going to be all your style.
169:30
So, I I completely moved away from that.
169:32
I I everything everything I blog is organic
169:35
handwritten maybe maybe I I
169:41
I use the AI as a
169:43
fix my worst typos,
169:47
there's value in the rough parts
169:52
of an actual human. Isn't that awesome?
169:54
Isn't that beautiful that now because of AI
169:57
we'll value the raw humanity
169:59
in each of us more.
170:02
I also I also realized
170:03
the thing that I
170:04
I rave about AI and use it so much for
170:07
anything that's code, but I'm allergic if it's stories. Right? Yeah.
170:13
Also documentation still fine with the AI, you know, better than
170:17
And for now it's still
170:18
in a place in the in a visual medium,
170:22
It's fascinating how allergic
170:23
I am to even a little bit of AI slop
170:25
in in video and images. It's useful.
170:29
It's nice if it's like a little component
170:31
of like or even even those image ads
170:34
like all these infographics
170:35
and stuff, they they trigger me so hard.
170:39
Like immediately makes me think less of your
170:43
and it they were novel
170:46
for like 1 week
170:48
and now it just screams slop. Yeah.
170:51
Even even if people work harder on
170:56
and I I have some of my blog posts, you know, in the in
170:58
the time where I
170:59
I explored this new medium,
171:01
but now they trigger me as well.
171:03
It's like, yeah, this is
171:04
this just screams AI slop.
171:06
I What I don't know what that is, but I went through that, too.
171:08
I was really excited by the diagrams
171:10
and then I realized in order to
171:12
remove from them hallucinations,
171:13
you actually have to do a huge amount of work.
171:17
And you're just using it to draw the better diagrams, great.
171:19
And then I'm proud of the diagram.
171:21
I've used them for literally like kind of like you said for maybe a
171:24
couple weeks and now I look at those
171:26
and I feel like I feel when I look at Comic Sans as a
171:29
font or something like this.
171:32
It's like, no, this is this is fake. It's fraudulent.
171:37
There's something wrong with it and it it's a smell.
171:41
It's >> it's a smell.
171:44
And it's awesome because it it reminds you that we
171:47
know there's so much to humans
171:49
that's amazing and we know that we we know it.
171:52
We know it when we see it.
171:55
And so, that gives me a lot of hope.
171:59
You know, that gives me a lot of hope about the human experience is
172:01
not going to be damaged,
172:02
but it's only going to be empowered
172:04
as tools by AI.
172:06
It's not going to be damaged or limited or
172:10
somehow altered to where it's no longer human.
172:15
So, I need a bathroom break. Quick pause.
172:21
You mentioned that a lot of the apps might be basically made obsolete.
172:27
You think agents will just transform
172:28
the entire app market? Yeah.
172:33
I noticed that on Discord that people just said
172:38
how they like what they build and what they use it for
172:41
and like, why do you need
172:43
MyFitnessPal when the agent already knows where I am?
172:49
can assume that I make bad decisions when I'm at
172:52
I don't know, Waffle House. What's around here?
172:55
Or or brisket in Austin.
172:57
There's no bad decisions
172:59
around brisket, but yeah.
173:00
No, that's the best decision,
173:03
>> Your agent should know that.
173:04
But it can like it can modify
173:06
my my gym workout
173:08
based on how well I slept
173:10
or if I'm if I have stress or not.
173:12
Like it has so much more context
173:15
to make even better decisions than any of this app even could do.
173:19
It could show me
173:20
UI just as I like.
173:23
Why do I still need an app to do that?
173:26
Why do I have to Why should I pay another subscription
173:28
for something that the agent can just do now?
173:34
why do I need my
173:35
my eight sleep app
173:37
to control my bed when I can tell the
173:39
tell the agent to
173:41
No, the agent already knows
173:42
where I am, so it can like turn off
173:44
what I don't use.
173:47
And I think that will that will translate
173:49
into a whole category
173:51
of apps are no longer
173:54
I will just naturally
173:55
stop using because my agent can just do it better.
174:00
I think you said
174:00
somewhere that it might kill off 80% of apps. Yeah.
174:05
Don't you think that's
174:06
a gigantic transformative effect on just all software development?
174:10
Does that mean it might kill off a lot of software companies? Yeah.
174:16
It's a scary thing.
174:17
So, like do you think about
174:19
the impact it has on the economy
174:24
just the ripple effects it has to society
174:26
transforming who builds what tooling.
174:32
It empowers a lot of
174:34
users to get stuff done, to get it stuff more efficiently,
174:39
to get it done cheaper.
174:41
There's also new services
174:42
that we will need, right?
174:44
For example, I want my agent to have an allowance.
174:50
you solve problems for me, here's like 100 bucks
174:54
in order to solve problems for
174:56
>> And if I tell it to order me food,
174:58
maybe it uses a service.
175:00
Maybe it uses something like rent a human.
175:03
So, like get that done for me.
175:06
I don't actually care.
175:07
I care about solve my problem.
175:11
There's space for for new companies that solve that well.
175:16
Maybe don't not all apps disappear.
175:18
Maybe some transform into being API.
175:21
So, basically apps that
175:24
rapidly transform in being agent facing.
175:30
So, there's a real opportunity
175:31
for like Uber Eats
175:35
who we just used earlier today.
175:37
It's companies like this of which there's many.
175:42
Who gets there fastest
175:43
to being able to interact
175:45
with open claw in a way that's
175:48
the the most natural, the easiest.
175:51
And also apps will become API if they want or not
175:57
my agent can figure out how to use my phone.
176:00
I mean on on the other side it's a little more tricky
176:03
on Android, but that's already
176:06
people already do that
176:07
and then it will just
176:08
click the order Uber for me button for me
176:13
or maybe another service
176:14
or maybe there's there's a there's an API it can call which is faster.
176:19
>> I think that's a space
176:20
we're just beginning to even understand
176:23
what that And I
176:25
again, I didn't even
176:27
that was not something I thought of, something that I
176:30
that I discovered as people use this.
176:33
And we're still so early.
176:35
But yeah, I think
176:36
data is very important.
176:38
Like apps that can give me data,
176:40
but that also can be API.
176:42
Why do I need the Sonos app anymore when I can
176:44
when my agent can talk to the Sonos speakers directly.
176:49
Like my cameras, there's like a crappy app,
176:52
but they have they have an API.
176:55
So, my agent uses the API now.
176:57
So, it's going to force a lot of companies to have to shift focus.
177:01
And it's kind of what the internet did, right?
177:02
You have to rapidly
177:03
rethink, reconfigure what you're selling,
177:09
how you're making money.
177:10
And some companies will really not like that.
177:12
For >> there's no CLI for Google.
177:16
So, I had to like
177:18
to have to do anything myself and build Gog.
177:22
That's like a CLI for
177:25
And at the yeah,
177:27
at the end user,
177:29
they have to give me the emails because otherwise I cannot use their product.
177:33
If I'm a company
177:35
and I try to get
177:37
Google data, Gmail, there's a whole
177:41
complicated process to the point where sometimes
177:44
startups acquire startups that went through that process
177:47
so they don't don't have to work with Google for half a year to
177:50
be to being able to access Gmail.
177:54
But my agent can access Gmail because I can just connect to it.
177:58
It's still crappy because I need to like go through
178:02
Google's developer jungle to get
178:06
a key and it's still annoying,
178:09
but they cannot prevent me.
178:11
And worst case, my agent just clicks on the on the website
178:16
and gets the data out that way. Through browser use.
178:18
Yeah, I mean I I watched my agent happily click the I'm not a robot button.
178:26
And there's this this whole
178:28
that's going to be
178:30
that's going to be more heated.
178:31
You see companies like
178:33
uh Cloudflare that try to prevent bot access.
178:37
And in some ways that's useful for scraping,
178:40
but in other ways if I'm
178:42
I'm a personal user, I want that.
178:45
You know, sometimes I
178:48
I use codex and I read an article about >> modern react patterns.
178:55
And it's like a medium article.
178:57
I paste it in
178:58
and the agent can't read it because they block it.
179:01
So I have to
179:02
copy-paste the actual text.
179:05
Or in the future I learn that maybe I don't click on medium
179:07
because it's annoying and I use other websites that actually are agent-friendly.
179:12
>> So There's going to be a lot of powerful rich companies fighting back.
179:17
So it's a really
179:18
interesting You're at the center.
179:19
You're the catalyst, the
179:22
And happen to be at the center of this kind of revolution
179:25
where it's going to completely change
179:27
of how we interact
179:29
with uh services, with with web.
179:33
And so like there's companies
179:35
like Google they're going to push back.
179:36
I mean there's every major company you can think of is going to push back.
179:39
>> Even Yeah, even search.
179:41
Um I now use
179:44
I think Perplexity or
179:46
Brave as providers because Google really doesn't make it easy to use Google without Google.
179:53
>> I'm not sure if that's the right strategy,
179:55
but I'm not Google.
179:58
>> Yeah, there's a there's a nice balance
179:59
from a big company perspective
180:01
cuz if you push back too much for too long you become Blockbuster
180:04
and you lose everything
180:05
to the Netflixes of the world, but some push back is probably
180:09
good during a revolution
180:10
to see But you see that that like this is something that the people want. Right. So Yes.
180:16
If I'm on the go,
180:18
I don't want to open a calendar app.
180:20
I just I want to tell my agent that
180:23
remind me about this dinner tomorrow night.
180:26
And maybe invite two of my friends and then maybe send a send a
180:28
WhatsApp message to my friend.
180:30
And I don't need
180:32
I don't want to need to open apps for that.
180:34
I think that we
180:36
passed that age and now everything is
180:40
much more connected and and fluid if
180:42
those companies want it or not.
180:44
And I think we'll
180:46
the right companies will find
180:49
ways to jump on the train and other companies
180:51
will You got to listen to what the people want.
180:57
Uh we talked about programming quite a bit and a lot of folks
181:01
that are developers are really worried about their jobs, about their
181:06
about the future of programming.
181:07
Do you think AI replaces
181:09
programmers completely, human programmers?
181:11
I mean we're definitely going in that
181:14
Programming is just a part of building products.
181:20
So maybe maybe I does replace programmers eventually.
181:25
But there's so much more to that art.
181:30
what do you actually want to build?
181:32
How should it feel? How's the architecture?
181:37
I don't think agents will replace all of that.
181:40
Yeah, like just if the actual art of programming
181:44
it will it will stay there, but it is it's going to be like knitting.
181:50
You know, like people do that because they like it, not because it makes any sense.
181:54
So so I read this article this morning
181:59
about someone that it's okay to mourn our craft.
182:03
And I can a part of me very strongly resonates with that because
182:07
in my past I
182:09
I spent a lot
182:11
of time tinkering, just being really deep in the flow
182:14
and just like cranking
182:16
out code and like finding really beautiful solutions
182:18
and yes, in a way it's
182:24
it's sad because that will go away
182:28
and I also got a lot of joy out of just
182:33
writing code and being really deep in my thoughts and forgetting
182:38
time and space and just
182:42
being in this beautiful state of flow.
182:44
But you can get the same state of flow
182:47
I get a similar state of flow by working with agents and building and
182:50
thinking really hard about problems.
182:52
It is and it's okay to mourn it,
182:58
but it's not something we can fight.
183:01
Like there is the world for a long time had a
183:06
there was a lack of intelligence
183:08
if you if you see it like that
183:11
of people building things
183:13
and that's why salaries
183:15
of software developers reached stupidly high amounts.
183:23
And that will go away.
183:25
Um there will still be a lot of demand for people that understand
183:28
how to build Just that
183:32
all this tokenized intelligence
183:38
enables people to do a lot more
183:41
a lot And it will be even more even faster and even more because
183:46
those things are continuously improving.
183:49
>> Um we had similar things when
183:51
I mean it's probably not a perfect analogy, but
183:54
when we created the steam engine and they built all these
183:57
factories and replaced a lot of manual labor
184:01
and then people revolted
184:02
and broke the machines.
184:04
Um I I can relate that
184:07
if you very deeply
184:12
that you are a programmer,
184:16
that it's scary and that it's threatening
184:18
because what you like and what you're really good at
184:22
is now being done
184:26
a soulless or not entity.
184:30
But I don't think you're just a programmer.
184:33
That does a very limiting
184:35
view of your craft.
184:37
You are you are still a builder.
184:39
Yeah, there's a couple of things I want to say.
184:41
So one is I never as you're
184:43
articulating this beautifully and I'm realizing
184:46
I never thought I would
184:49
the thing I love doing
184:51
would be the thing that gets replaced.
184:54
You hear these stories about these like said with the steam engine.
184:57
I've I've spent so many
185:00
I don't know, maybe thousands of hours
185:02
pouring over code and putting my heart and soul and
185:06
like and just like some of my most painful
185:09
and happiest moments were alone
185:12
behind I I was an Emacs person for a long time. And Emacs.
185:17
And and then there's an identity
185:19
and there's meaning and there's
185:22
like when I walk about the world I don't
185:24
say it out loud, but I think of myself as a
185:27
And to have that
185:28
in a matter of months.
185:30
I mean like you mentioned April to November.
185:33
It really is a leap
185:35
that happened, a shift that's happening.
185:39
>> To have that completely
185:40
replaced is is painful.
185:43
It's it's truly painful, but I also think
185:45
um builders more broadly,
185:50
but what is what is the act of programming?
185:53
I I think programmers
185:56
are generally best equipped at this moment in history
186:00
to learn the language,
186:02
to empathize with agents,
186:05
to learn the language of agents,
186:07
to feel the CLI. Yeah.
186:11
Like like to understand
186:13
what is the you
186:15
need, you the agent,
186:18
need to do this task the best.
186:21
I think at some point it's just going to be called coding again and
186:23
it's just going to be the new normal. Yeah.
186:25
>> And yet while I don't write the code, I very much feel like
186:29
I'm in the driver's seat and I am
186:33
I am writing the code.
186:35
Yeah, it's just You'll still be a programmer.
186:38
It's just the activity of a programmer is is different.
186:41
>> Yeah, and cuz on X
186:43
the bubble I'm in is mostly positive.
186:46
On on Mastodon and Blue Sky
186:48
I don't I also use it less because
186:51
often times I got attacked
186:52
for my blog I I had stronger reactions in the past.
186:58
Now I can sympathize with those people
187:02
in a way I get it.
187:04
It In a way I also don't get it because it's very unfair
187:07
to grab onto the person that you see right now and unload all your fear and hate.
187:14
It's going to be a change and it's going to be challenging,
187:17
but it's also I don't know.
187:19
I find it incredibly
187:21
fun and and and gratifying
187:23
and I can I can use the new time to focus
187:30
on much more details.
187:31
I think the level of expectation
187:33
of what we build is also rising because it's just now
187:38
>> the default is now so much easier.
187:41
So software is changing in many ways.
187:45
There's going to be a lot more.
187:48
>> And then you have all these people that are screaming,
187:50
"Oh yeah, but what about the water?" You know?
187:54
Like I did a conference
187:55
in Italy about the the state of AI and
188:00
my whole motivation was to push people away from
188:04
don't see yourself as an iOS developer anymore, you're now a builder.
188:08
And you can use your skills in many more ways.
188:11
Also because apps are slowly going
188:13
Uh people didn't like that.
188:15
Like uh a lot of people didn't like what I had to say.
188:19
And I don't think I was hyperbolic.
188:20
I was just like
188:21
this is how I see the future.
188:24
Maybe this is not
188:25
how it's going to be,
188:26
but I'm pretty sure a version of that will happen.
188:30
And the first question I got was, "Yeah, but what about the insane water
188:33
use on data centers?"
188:36
But then you actually sit down
188:38
and do the math,
188:39
and then for most people,
188:41
if you just skip one burger
188:43
per month, that compensates
188:45
the the CO2 output
188:48
or like the water use
188:50
uh in equivalent of tokens.
188:52
I mean, the math is is the math is tricky, and it depends if
188:56
you add pre-training, then maybe it's more than just one patty,
189:01
it's not off by factor of 100, you know?
189:04
So so there or like
189:06
golf is still using
189:09
way more water than all data centers together.
189:11
So are you also hating people that play golf?
189:15
Those people grab on anything that they think is bad about AI
189:19
without seeing the potential
189:21
things that might be good about AI.
189:24
And I'm not saying everything's good.
189:26
It's certainly going to be
189:29
a very transformative technology
189:31
for our There is uh to steelman
189:34
the the criticism in general,
189:37
I do want to say in my experience
189:38
with Silicon Valley, um there's a bit of a bubble
189:43
in the sense that
189:45
there's a kind of excitement
189:48
and an overfocus about
189:51
the positive that the technology can bring. Yeah.
189:55
And which is great.
189:56
It's great to focus on not to
189:59
not to be paralyzed by fear and fearmongering and so on.
190:02
But there's also within that excitement
190:05
and within everybody talking just to each other,
190:09
there's a dismissal of the basic human experience across
190:12
the United States in the Midwest,
190:14
across the world, including
190:16
the programmers we mentioned, including all the people that are going to lose their
190:19
jobs, including the the measurable
190:22
pain and suffering that happens at the short-term
190:24
scale when there's change
190:25
of kind, especially large-scale
190:28
transformative change that we're about to
190:30
face if what we're talking about will materialize.
190:34
And so to having a bit of that humility
190:37
and awareness about the tools
190:40
you're building, they're going to cause pain.
190:43
They will long-term hopefully bring about
190:46
a better world and even more opportunities
190:48
and even more awesomeness,
190:50
but having that kind of like
190:52
quiet moment often of
190:57
of respect for the pain that is going to be felt.
191:01
And so like not not enough of that is I think done, so it's
191:05
it's good to have a bit of that.
191:07
And then I also have to put against
191:09
uh some of the emails I got where people told me
191:13
they have a small business
191:14
and they've been struggling
191:15
and and OpenClaw helped them automate
191:18
a few of the tedious tasks
191:21
from collecting invoices to like answering customer emails
191:26
that then freed them up and like
191:28
caused them a bit more joy in their life.
191:31
Or or some emails where they told me that OpenClaw
191:35
helped a disabled daughter
191:37
uh that she's now empowered
191:39
and feels she can do much more than before.
191:42
Which is amazing, right?
191:43
Because you you could
191:46
you could do that before as well.
191:47
The technology was there.
191:48
I didn't I didn't invent a whole new thing,
191:51
but I made it a lot easier and more accessible,
191:54
and that did show people the possibilities
191:57
that they previously wouldn't see,
192:00
and now they applied for good.
192:02
Or like also the fact that
192:05
yes, I I I
192:07
suggest the the the latest and best models,
192:10
but you can totally
192:11
run this on free models.
192:13
You can run this locally.
192:14
You can run this on on on Kimmy or other
192:17
other other models that are
192:20
way more accessible price-wise
192:23
and still have a a very powerful system
192:25
that might otherwise not be possible
192:29
other things that I don't know, Anthropic's
192:31
Claude work is locked in into their space,
192:34
so it's not all black and white.
192:38
I got a lot of
192:39
emails that were heartwarming
192:42
and amazing and and
192:45
I don't know, it just made me really
192:48
Yeah, there's a lot it has brought joy into a lot of people's lives,
192:51
not just not just programmers,
192:53
like a lot of people's lives.
192:55
It's it's it's beautiful to see.
192:57
Uh what gives you hope about this whole thing we have going on with human civilization?
193:03
I mean, I inspired so many people.
193:06
There's like there's this whole builder vibe again.
193:10
People are now using AI in a more playful way
193:16
and are discovering what it can do
193:22
and how it can like help them in their life
193:27
creating new places that are just sprawling of creativity.
193:33
I don't know, like
193:35
there's like Clockon in Vienna,
193:37
there's like 500 people,
193:40
there's such a high percentage of people that have want to present,
193:43
which is to me really surprising
193:44
because you usually it's quite hard to find people that want to like talk
193:49
about what they build, and now it's there's an abundance.
193:53
So that gives me hope
193:55
that we can we can figure out.
194:01
it makes it accessible to basically everybody.
194:05
Just imagine all these people building.
194:09
Especially as you make it simpler and
194:12
more secure, it's like
194:14
anybody who has ideas
194:17
that can express those ideas in language can build. It's crazy.
194:22
Yeah, that's ultimately power to people and
194:26
one of the beauty
194:28
the beautiful things that come out of
194:31
Not just not just a slot generator. >> Well, Mr.
194:37
Claus Father, I just realized when I said that in the beginning,
194:40
I violated two trademarks
194:42
cuz there's also the Godfather.
194:44
I'm getting sued by
194:47
You're a wonderful human being.
194:49
Uh you've created something really
194:52
a special community, a special product, a special set of ideas,
194:56
plus the entire the humor, the good vibes,
194:59
the inspiration of all these people building,
195:02
the excitement to build.
195:06
Uh so I'm truly grateful
195:07
for everything you've been doing and
195:09
uh for who you are and for
195:11
sitting down to talk with me today. Thank you, brother.
195:14
Thanks for giving me the chance to tell my story.
195:17
Thanks for listening to this conversation with Peter Steinberger.
195:20
To support this podcast,
195:21
please check out our sponsors
195:22
in the description, where you can also find links to contact me, ask questions,
195:26
give feedback, and so on.
195:28
And now, let me leave you with some words from Voltaire.
195:33
With great power comes great
195:38
Thank you for listening,
195:39
and hope to see you next time.
Thích
Chia sẻ
Lex Fridman
Xem tất cả →
C1
Phỏng vấn
3:17:34
Narendra Modi: Prime Minister of India - Power, Democracy, War & Peace | Lex Fridman Podcast #460
Lex Fridman
7
C1
Phỏng vấn
Karaoke
4:34:10
Pavel Durov: Telegram, Freedom, Censorship, Money, Power & Human Nature | Lex Fridman Podcast #482
Lex Fridman
4
C1
Phỏng vấn
Karaoke
5:06:18
DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459
Lex Fridman
6
C1
Phỏng vấn
Karaoke
5:20:08
ThePrimeagen: Programming, AI, ADHD, Productivity, Addiction, and God | Lex Fridman Podcast #461
Lex Fridman
1
C1
Phỏng vấn
Karaoke
3:14:34
Terence Tao: Hardest Problems in Mathematics, Physics & the Future of AI | Lex Fridman Podcast #472
Lex Fridman
1
C1
Phỏng vấn
Karaoke
3:24:36
James Holland: World War II, Hitler, Churchill, Stalin & Biggest Battles | Lex Fridman Podcast #470
Lex Fridman
C1
Phỏng vấn
Karaoke
2:45:26
Dan Houser: GTA, Red Dead Redemption, Rockstar, Absurd & Future of Gaming | Lex Fridman Podcast #484
Lex Fridman
C1
Phỏng vấn
Karaoke
2:05:13
Deciphering Secrets of Ancient Civilizations, Noah's Ark, and Flood Myths | Lex Fridman Podcast #487
Lex Fridman
Video gợi ý
C1
Phỏng vấn
3:17:34
Narendra Modi: Prime Minister of India - Power, Democracy, War & Peace | Lex Fridman Podcast #460
Lex Fridman
7
C1
Phỏng vấn
Karaoke
4:34:10
Pavel Durov: Telegram, Freedom, Censorship, Money, Power & Human Nature | Lex Fridman Podcast #482
Lex Fridman
4
C1
Phỏng vấn
Karaoke
5:06:18
DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459
Lex Fridman
6
C1
Phỏng vấn
Karaoke
5:20:08
ThePrimeagen: Programming, AI, ADHD, Productivity, Addiction, and God | Lex Fridman Podcast #461
Lex Fridman
1
C1
Phỏng vấn
Karaoke
3:14:34
Terence Tao: Hardest Problems in Mathematics, Physics & the Future of AI | Lex Fridman Podcast #472
Lex Fridman
1
C1
Phỏng vấn
Karaoke
3:24:36
James Holland: World War II, Hitler, Churchill, Stalin & Biggest Battles | Lex Fridman Podcast #470
Lex Fridman
C1
Phỏng vấn
Karaoke
2:45:26
Dan Houser: GTA, Red Dead Redemption, Rockstar, Absurd & Future of Gaming | Lex Fridman Podcast #484
Lex Fridman
C1
Phỏng vấn
Karaoke
2:05:13
Deciphering Secrets of Ancient Civilizations, Noah's Ark, and Flood Myths | Lex Fridman Podcast #487
Lex Fridman
