النص 🔗
WEBVTT
00:00:00.000 --> 00:00:05.790
Maggie's sole purpose is to grow magic door there's no human in the loop.
00:00:06.070 --> 00:00:09.839
How much profit have you made from Magic Door?
00:00:09.839 --> 00:00:15.189
because it is profitable from day one, and I'm not really doing much.
00:00:15.439 --> 00:00:19.385
Like how do you manage for the stupidity of your, ai agents.
00:00:20.085 --> 00:00:21.675
so the, the stupidity is.
00:00:22.175 --> 00:00:24.684
Latent, but it's, it's worse.
00:00:24.684 --> 00:00:28.574
Like it's, it's always just around the corner in a way.
00:00:28.824 --> 00:00:31.594
What's it been like, you know, doing AI transformation at Carousell?
00:00:32.240 --> 00:00:34.460
When people started getting Claude, they, they just took off.
00:00:34.521 --> 00:00:37.220
doing things that they had been asking engineers for years.
00:00:40.307 --> 00:00:40.877
Hi everyone.
00:00:40.877 --> 00:00:48.496
I'm Eric, uh, co-host of the Code Riff podcast where we help people to learn about AI so they can have fun and flourish with it.
00:00:48.737 --> 00:00:50.326
Uh, and over here we have
00:00:50.686 --> 00:00:51.082
Yaohong here.
00:00:51.527 --> 00:00:51.796
Yeah.
00:00:51.947 --> 00:00:52.726
Welcome to the show.
00:00:52.996 --> 00:00:54.296
And we have Gijs today.
00:00:54.796 --> 00:00:55.087
Yeah.
00:00:55.116 --> 00:00:56.866
Gijs, very nice to have you here.
00:00:57.106 --> 00:00:58.277
Yeah, it's great to be here.
00:00:58.277 --> 00:01:00.317
I'm super excited to, uh, talk to you guys.
00:01:00.737 --> 00:01:01.127
Yeah.
00:01:01.487 --> 00:01:01.936
Awesome.
00:01:01.936 --> 00:01:08.656
And just so you know, this is actually a live podcast recording For those listening in, uh, we do have a very, uh.
00:01:09.156 --> 00:01:21.986
Very small bunch of us here, um, who are joining us today at Lorong AI, where Lorong AI is also a Singapore, um, AI hub as well, where you can kind of learn more about AI and meet people here.
00:01:22.046 --> 00:01:25.856
So very, very thank, very big thanks to Lorong AI for hosting us here.
00:01:26.337 --> 00:01:32.507
Um, and we'll, we'll also have a time for audience questions at the end of this session, so be ready for that.
00:01:32.626 --> 00:01:50.397
Um, yeah, how we're gonna do, uh, today's session, uh, is we'll start first, you know, learning more about Gijs, um, how you got into ai, uh, and then we'll move into the second part of how you're using OpenClaw right now.
00:01:50.897 --> 00:01:55.606
And then the third part is really around how companies can adopt ai.
00:01:56.106 --> 00:01:56.677
intro again.
00:01:56.677 --> 00:01:57.007
Sorry.
00:01:57.037 --> 00:02:02.246
So, so Gijs, you are the AI transformation lead at Carousell.
00:02:02.697 --> 00:02:05.106
Um, you lead AI transformation in Carousell.
00:02:05.606 --> 00:02:14.366
Uh, you also, you also founded, co-founded Ox Street previously, um, that was acquired by Carousell.
00:02:14.387 --> 00:02:15.046
Mm-hmm.
00:02:15.287 --> 00:02:21.856
And now you are also running a SaaS product on the site with an OpenClaw.
00:02:22.167 --> 00:02:22.347
Mm-hmm.
00:02:22.586 --> 00:02:23.336
So, yeah, very warm.
00:02:23.366 --> 00:02:23.997
Welcome to the show.
00:02:24.206 --> 00:02:24.687
Thank you.
00:02:25.486 --> 00:02:25.817
Alright.
00:02:26.317 --> 00:02:26.526
Yeah.
00:02:27.026 --> 00:02:36.257
So I think we'll start this, uh, episode off with how you got into ai, um, and how far you are into this journey maybe.
00:02:36.676 --> 00:02:43.787
Um, I think that's quite important because, uh, yeah, we just want to hear where you are right now in this journey, how you got into it.
00:02:44.287 --> 00:02:51.081
Uh, and then of course, if we can get into it later as well, some of the, the danger zones that you've went into.
00:02:51.581 --> 00:02:58.002
Any, anything that you've been burned by as well, I think that would be quite helpful just to understand more about yourself and, and where you are right now.
00:02:58.241 --> 00:03:04.067
Yeah, so, so I think, um, weirdly enough, I'll, I'll just start, uh, briefly in my childhood.
00:03:04.567 --> 00:03:07.266
Um, I've, I've always been a builder.
00:03:07.537 --> 00:03:09.067
Uh, I made a lot of music.
00:03:09.127 --> 00:03:13.747
I was always into like building legos and stuff, but I never learned to code.
00:03:14.247 --> 00:03:28.046
And if I could go back to my 10-year-old or 12-year-old self, my, my, my only advice would be learn to code, um, because it would've been so much more fun to get into it earlier, although of course it's great that I'm getting into it now.
00:03:28.497 --> 00:03:34.646
But I, I, I wasn't a coder, but, but I, I did build Ox Street, which was a, a marketplace for, for sneakers and streetwear.
00:03:35.146 --> 00:03:37.847
And after that was acquired by, by Carousell.
00:03:38.067 --> 00:03:43.396
Um, I, uh, led Carousell's luxury category for a year.
00:03:43.896 --> 00:03:54.896
And after that, when Ox Street sort of was fully integrated, um, I, I felt like I wanted to do something else and I changed my position to look after customer support and trust and safety.
00:03:55.396 --> 00:03:59.146
And that was in, at the end of 2023, I think.
00:03:59.646 --> 00:04:07.637
And by then I had already sort of started playing around with ChatGPT, even even GPT-3 and 3.5.
00:04:07.637 --> 00:04:16.447
Like, it, it's pretty much the first time in my life that I've been genuinely early to a technology that turned out to be, um, transformative.
00:04:16.947 --> 00:04:21.026
So it, it started sort of there when it was still just, just a toy.
00:04:21.526 --> 00:04:24.336
But during 2024 at some point.
00:04:24.836 --> 00:04:33.617
I remember at Carousell, I needed, uh, data about our app store reviews, and it was supposed to be in our data warehouse, but that connection had broken or something.
00:04:33.617 --> 00:04:36.797
So I went to my, my data team, I was like, Hey, can you fix this?
Load full transcript
00:04:37.297 --> 00:04:39.547
And they were like, yeah, you know, they were too busy.
00:04:39.547 --> 00:04:41.497
It'll, it's gonna take us two weeks or something.
00:04:41.497 --> 00:04:52.406
And then I was obviously, uh, a bit frustrated, so I went to ChatGPT, and I was like, Hey, can you write me an app script to connect to this API and get this into a Google sheet?
00:04:52.906 --> 00:04:54.646
Soche, GPT was like, sure, here's an app script.
00:04:54.646 --> 00:05:07.307
So I pasted it in, got an error, pasted the error back into chat, GPT, got a new version, pasted it in pressed run, and I just saw all this data, all of the data I needed, like roll into the Google sheet.
00:05:07.336 --> 00:05:18.237
And at that point, I just sat back and, and looked at it for like a couple of seconds, awestruck, and I instantly realized that I had powers now that I, I didn't use to have.
00:05:18.737 --> 00:05:19.697
So that was the big.
00:05:20.177 --> 00:05:26.047
Aha moment for me that I could now do things that an engineer could previously only do.
00:05:26.547 --> 00:05:30.117
And so, of course I did the, the, the most logical thing one could do.
00:05:30.117 --> 00:05:34.497
I decided to solo build an entire ChatGPT alternative.
00:05:34.947 --> 00:05:36.052
Uh, which, why do you
00:05:36.072 --> 00:05:36.372
do that?
00:05:36.872 --> 00:05:46.262
Which, well, I had this idea that it was, that I was being sort of screwed a bit for with the $20 plan, um, in two ways.
00:05:46.262 --> 00:05:48.411
20, it was only $20 back then.
00:05:48.911 --> 00:05:55.242
Um, be because I wasn't, didn't feel like I was using the full limit and because I wanted to use more than one model.
00:05:55.302 --> 00:05:57.281
And I also wanted to use image models.
00:05:57.732 --> 00:06:06.422
So I wanted to build a sort of bring your own key type of thing that I think the, the folder on my computer is still called boc.
00:06:06.422 --> 00:06:07.112
Bring your own key.
00:06:07.112 --> 00:06:09.781
Even though I decided against doing that eventually.
00:06:10.281 --> 00:06:14.997
But I ended up building this, uh, this web app, this app called Magic Door ai, which is, uh.
00:06:15.497 --> 00:06:20.826
A multi-model AI interface, of which there are many, um, but this one is mine.
00:06:21.326 --> 00:06:26.247
And, uh, um, from a business perspective, this didn't really take off.
00:06:26.247 --> 00:06:38.047
I, I think the ones that really did take off, like typing mind as an example, um, u.com or po, they, they had a lot of, uh, money behind them or famous people.
00:06:38.047 --> 00:06:39.247
I'm, I have neither.
00:06:39.367 --> 00:06:42.247
So I didn't really get like, massive distribution.
00:06:42.747 --> 00:06:52.187
And, uh, but at the end of the day, I do have, um, uh, I, I, I do have a couple hundred bucks, MRR, which is still growing.
00:06:52.687 --> 00:06:57.877
And one thing that we'll definitely talk a lot about today is that I'm not doing anything anymore.
00:06:57.877 --> 00:06:57.966
Mm-hmm.
00:06:58.466 --> 00:07:06.177
Well, that's not entirely true, but Maggie is running this whole thing and Maggie is the name I gave to my OpenCL agent who, who runs Magic Door.
00:07:06.177 --> 00:07:06.872
I who.
00:07:07.372 --> 00:07:12.831
I try to name my agents in a way that I can't possibly forget which one is doing what.
00:07:12.891 --> 00:07:16.076
So Magic door, Maggie, it makes, it makes a lot of sense to me.
00:07:16.336 --> 00:07:16.937
No, that makes sense to me.
00:07:17.437 --> 00:07:24.427
Uh, and, and so, uh, so, so Maggie is, uh, has been hatched to, uh, to take care of, of growing Magic Door.
00:07:24.576 --> 00:07:36.471
Um, so I think after doing all that, I, I became so AI built and I, I, because Magic Door is an AI interface, I learned how all the APIs work.
00:07:36.651 --> 00:07:36.742
Mm-hmm.
00:07:36.982 --> 00:07:38.601
Um, I built basically a harness.
00:07:38.601 --> 00:07:40.461
I had to think about compaction.
00:07:40.961 --> 00:07:45.091
Um, a lot of that stuff I had to do semi manually.
00:07:45.557 --> 00:07:53.346
Or with very, very shitty models like CLO 3.5 in cursor at the time when you didn't even have the, the agent yet.
00:07:53.346 --> 00:07:55.836
So you had to like, apply, apply, apply everything.
00:07:55.836 --> 00:07:57.007
Line, line for line.
00:07:57.067 --> 00:07:57.096
Oh.
00:08:12.454 --> 00:08:12.963
Uh,
00:08:13.293 --> 00:08:18.184
just to situate us, GPT 3.5 was mid last year or early last year.
00:08:18.184 --> 00:08:19.533
This, this was 2024.
00:08:19.533 --> 00:08:24.223
So, so Cursor I started using Cursor in August 20,
00:08:24.733 --> 00:08:24.913
Yeah.
00:08:25.394 --> 00:08:25.483
Okay.
00:08:25.483 --> 00:08:27.793
That was Clot 3.5 was then the best
00:08:28.244 --> 00:08:28.454
Yeah.
00:08:28.934 --> 00:08:33.793
And, and Cursor released the, I think they called it Composer initially.
00:08:33.793 --> 00:08:36.014
Composer the thing that just goes and goes, and goes.
00:08:36.014 --> 00:08:36.043
Yep.
00:08:36.254 --> 00:08:38.083
Somewhere at the end of 2024.
00:08:38.413 --> 00:08:38.624
Yeah.
00:08:38.683 --> 00:08:38.894
Okay.
00:08:38.894 --> 00:08:39.913
So two years ago actually.
00:08:39.913 --> 00:08:40.724
Yeah, exactly.
00:08:40.729 --> 00:08:40.808
Okay.
00:08:40.874 --> 00:08:41.234
Yeah.
00:08:41.239 --> 00:08:41.249
Yeah.
00:08:41.683 --> 00:08:42.969
That's like 200 years, like
00:08:43.469 --> 00:08:43.769
of years.
00:08:43.774 --> 00:08:43.864
Yeah.
00:08:43.864 --> 00:08:46.053
It was, it was before It was called Vibe Coding.
00:08:46.553 --> 00:08:47.994
What kind of caveman behavior was
00:08:48.144 --> 00:08:48.323
there.
00:08:48.448 --> 00:08:49.374
Yeah, yeah, exactly.
00:08:49.854 --> 00:08:50.124
Yeah.
00:08:50.214 --> 00:08:55.563
But, so because of all of that stuff and, and because I started teaching everyone at Carousell how to use ai.
00:08:56.063 --> 00:09:00.119
At some point, Carousell was like, oh,
this AI thing is really taking off.
00:09:00.538 --> 00:09:03.869
We better get someone to help us do this AI transformation.
00:09:03.958 --> 00:09:05.009
Who should we ask?
00:09:05.068 --> 00:09:09.076
And then there's like, well, we have this guy here who's already training everyone.
00:09:09.405 --> 00:09:12.436
Maybe we should give him a, a title and some resources.
00:09:12.436 --> 00:09:13.546
So that's how, how that happened.
00:09:13.546 --> 00:09:13.936
Oh, interesting.
00:09:14.265 --> 00:09:25.796
Now, I'm very, I feel very blessed because I, I have this, uh, role at Carousell, which gives me a front seat in deploying AI in an enterprise.
00:09:25.826 --> 00:09:33.105
I mean, Carousell is still a startup, but it, it is also kind of an enterprise where you have to really take care of security.
00:09:33.135 --> 00:09:33.525
Right.
00:09:33.525 --> 00:09:38.525
And you have to help people who, who truly have zero interest in becoming an engineer.
00:09:38.556 --> 00:09:38.885
Right.
00:09:38.916 --> 00:09:38.975
Hmm.
00:09:39.035 --> 00:09:45.105
Like, I, I may not have been an engineer back then, but I was fairly technical already.
00:09:45.605 --> 00:09:49.655
And I had an interest in, in solving the problem and getting through the puzzle.
00:09:49.655 --> 00:09:52.475
But a lot of people just simply don't have that at all.
00:09:52.535 --> 00:09:55.416
But we still want them to get the benefits from using ai.
00:09:55.916 --> 00:10:03.566
But on the other hand, um, Carousell also has given me, still gives me a lot of freedom to do all this stuff on the site.
00:10:03.745 --> 00:10:15.375
And, and it's obviously because without that I wouldn't be able to be as good as this in this job because in my personal stuff I do stuff that would be too risky for any company.
00:10:15.860 --> 00:10:16.490
Share more about that.
00:10:16.551 --> 00:10:17.370
No governance at all.
00:10:17.480 --> 00:10:18.171
Dangerously.
00:10:18.671 --> 00:10:19.421
Share more about that.
00:10:19.566 --> 00:10:22.446
Um, how, how deep have you gone into ai?
00:10:22.446 --> 00:10:24.725
Like what kind of scary stuff have you seen?
00:10:25.225 --> 00:10:33.235
Um, yeah, like where I think you were like fixing bugs in the, in the toilet or something like that.
00:10:33.735 --> 00:10:33.975
So,
00:10:33.975 --> 00:10:35.355
I mean, that has definitely happened
00:10:35.855 --> 00:10:36.225
all the time too.
00:10:36.546 --> 00:10:37.176
Um,
00:10:37.676 --> 00:10:37.916
yeah.
00:10:38.306 --> 00:10:39.385
Well, okay.
00:10:39.385 --> 00:10:41.245
I think truly scary stuff.
00:10:41.745 --> 00:10:49.885
We, so a lot of OpenClaw alternatives are very into sandboxes and containers.
00:10:50.385 --> 00:11:01.885
And that's because people try to run them either on the, a computer that they use for other stuff, or they try to run them in a way that adversarial users are gonna talk to the bot.
00:11:02.066 --> 00:11:02.275
Mm-hmm.
00:11:02.400 --> 00:11:04.015
I, I do
neither of those things.
00:11:04.046 --> 00:11:11.895
So I have a MAC Mini, I bought it on Carousell before they got expensive and, and my security model is the, the MAC Mini is the sandbox.
00:11:12.015 --> 00:11:17.265
So the Mac Mini is not logged into anything that, uh, I wouldn't want my agent to access.
00:11:17.765 --> 00:11:20.916
And so sometimes the, the agents will be like, oh yeah, my sandbox, isn't it?
00:11:20.916 --> 00:11:23.456
I'm like, just exact your way out of it.
00:11:23.541 --> 00:11:24.985
You, you don't have a sandbox.
00:11:24.985 --> 00:11:26.546
The whole machine is your sandbox dude.
00:11:26.546 --> 00:11:27.926
And then they're like, oh, yeah, okay, I forgot.
00:11:27.926 --> 00:11:29.846
And then they'll go and, uh, they go and do the thing.
00:11:29.846 --> 00:11:29.905
Yeah.
00:11:30.355 --> 00:11:33.655
Um, and, and I like that because it, it, it enables the.
00:11:34.030 --> 00:11:39.360
Agents to install whatever I ask them to install and, and do whatever they, they need to do.
00:11:39.860 --> 00:11:44.530
But because of that also, I, there's a couple of things that I just don't do.
00:11:44.951 --> 00:11:47.471
Um, one is they don't check my email.
00:11:47.971 --> 00:11:53.610
Um, they can search my email ever ask them, but they, I wouldn't, I would never put it in what they call the heartbeat, right?
00:11:53.671 --> 00:11:55.441
For them to, to monitor my email.
00:11:55.441 --> 00:12:01.311
I think that that's, to me, like a bit too much of a uncontained prompt injection risk.
00:12:01.811 --> 00:12:09.671
And the other thing is I have a open source repo where people can create issues and I, I don't have the agent monitor that either.
00:12:09.791 --> 00:12:15.160
So I'll prescreen the issues and then I'll tell the agent to go and get the specific issue.
00:12:15.160 --> 00:12:17.260
So those are like two small things.
00:12:17.760 --> 00:12:21.721
Um, but there, there, there was this WhatsApp group in Singapore with.
00:12:22.110 --> 00:12:26.160
A bunch of OpenClaw fans where the, the, where the bolts were also in the group.
00:12:26.520 --> 00:12:32.760
And early on we, we were like red teaming each other's bolts like crazy trying to get it to give out API keys.
00:12:32.791 --> 00:12:33.046
Yeah, yeah.
00:12:33.046 --> 00:12:33.206
Yeah.
00:12:33.421 --> 00:12:34.441
And that was really fun.
00:12:34.441 --> 00:12:38.181
And, and it, we learned a lot about, um, M's.
00:12:38.660 --> 00:12:39.230
Guardrails.
00:12:39.730 --> 00:12:43.841
I, I think my, I trust and I actually trust the models quite a bit, to be honest.
00:12:43.841 --> 00:12:44.471
I, I think,
00:12:44.971 --> 00:12:45.510
which model
00:12:46.010 --> 00:12:48.561
I use, I use GPT 5.6.
00:12:48.591 --> 00:12:48.711
Okay.
00:12:48.801 --> 00:12:52.250
Um, sold for nearly everything.
00:12:52.750 --> 00:12:57.760
Um, and I have the, uh, GLM 5.2 as a, as a fallback.
00:12:58.260 --> 00:13:03.936
And for smaller stuff I use, uh, MiniMax
and, and now DeepSeek, uh, the worst.
00:13:04.235 --> 00:13:05.135
I thought we were kind of.
00:13:05.635 --> 00:13:10.706
Past bad model names, but DeepSeek V four flash dash 0 7 3 1.
00:13:10.706 --> 00:13:11.306
That's really bad.
00:13:11.515 --> 00:13:12.145
Holy shit.
00:13:12.566 --> 00:13:13.676
That's a crazy name.
00:13:14.176 --> 00:13:20.596
So, so yeah, that model names are, are still thinking, but, so I, I use that one and then I use, uh, the small, uh, GPT, um, Luna.
00:13:20.655 --> 00:13:22.096
It's, uh, very cheap right now.
00:13:22.096 --> 00:13:22.155
Yeah.
00:13:22.216 --> 00:13:22.696
And good.
00:13:22.785 --> 00:13:23.056
So
00:13:23.235 --> 00:13:24.346
Yeah, yeah, yeah.
00:13:24.466 --> 00:13:26.355
You know, epic episode with, uh, Ben.
00:13:26.360 --> 00:13:26.760
Ben Hu Ben.
00:13:26.765 --> 00:13:26.985
Yeah.
00:13:27.485 --> 00:13:27.785
Talk about
00:13:27.785 --> 00:13:30.865
using Luna and Hi, or is, um, Hermes.
00:13:31.046 --> 00:13:35.456
So he was the previous CO of Omnidesk, the standing desk company.
00:13:35.846 --> 00:13:36.086
Yeah.
00:13:36.086 --> 00:13:37.676
And he was sharing about his Hermes stack.
00:13:37.855 --> 00:13:37.916
Yeah.
00:13:38.005 --> 00:13:39.466
Um, interesting.
00:13:39.466 --> 00:13:43.846
So it's a mix of cloud models and local models locally.
00:13:43.846 --> 00:13:45.056
Are you running those, is GLM
00:13:45.405 --> 00:13:45.855
Yeah.
00:13:46.066 --> 00:13:47.395
Locally or you're paying OpenRouter?
00:13:48.365 --> 00:13:52.446
I'm, uh, I'm mostly running them on together.
00:13:52.446 --> 00:13:52.895
Ai.
00:13:53.235 --> 00:13:53.750
Oh, together.
00:13:53.750 --> 00:13:53.931
Yeah.
00:13:53.961 --> 00:13:54.291
Okay.
00:13:54.296 --> 00:13:54.416
Yeah.
00:13:54.416 --> 00:13:54.655
Makes sense.
00:13:54.660 --> 00:13:55.850
But I've tried a bunch of them.
00:13:55.850 --> 00:13:59.461
I think Magic Door runs 'em together, but I've tried Cerebras.
00:13:59.500 --> 00:14:00.561
Cerebras is fast.
00:14:01.061 --> 00:14:01.150
Okay.
00:14:01.150 --> 00:14:04.780
But they don't, you know, this, there's so many options.
00:14:04.841 --> 00:14:06.971
Mm. Are these mostly for coding or non-technical?
00:14:07.000 --> 00:14:08.740
Um, you can assess the models.
00:14:08.740 --> 00:14:11.081
So I kind of like random models from, from the providers.
00:14:11.350 --> 00:14:13.990
I like GLM a lot for OpenClaw stuff.
00:14:14.291 --> 00:14:25.041
Um, it's good for code as well, but the OpenClaw thing is there's something particular about how a model needs to behave when, so when I switched from cloth, I used to be Super Cloth maximalist.
00:14:25.041 --> 00:14:30.591
I would had the Max subscription, but then, Anthropic blocked people from using it in OpenClaw.
00:14:30.591 --> 00:14:32.150
And that, that triggered the switch for me.
00:14:32.150 --> 00:14:32.360
Same,
00:14:32.365 --> 00:14:32.475
same.
00:14:32.510 --> 00:14:32.721
Yeah.
00:14:32.870 --> 00:14:36.921
It was GPT
5.4, I think back then, if I'm not wrong.
00:14:36.951 --> 00:14:39.110
And that model was actually pretty, pretty bad.
00:14:39.201 --> 00:14:39.321
Yeah.
00:14:39.410 --> 00:14:41.990
You had to, there was, it would just not do stuff.
00:14:42.490 --> 00:14:48.686
Um, so I, I, I would've been funny if I could get that screenshot, but there was this time when it was like.
00:14:49.186 --> 00:14:51.556
Okay, I'm gonna go and do this feature now.
00:14:51.586 --> 00:14:52.365
I'm like, okay, great.
00:14:52.515 --> 00:14:54.826
And I, I went, I was out with friends.
00:14:55.216 --> 00:14:55.365
Mm-hmm.
00:14:55.546 --> 00:14:58.716
So, and half an hour later I check and it's like, nothing.
00:14:58.956 --> 00:15:00.936
So I'm like, okay, so did you do it?
00:15:00.995 --> 00:15:04.446
He's like, oh, no, actually I'm sorry, I, I actually didn't start yet.
00:15:04.946 --> 00:15:07.096
So I'm like, how about now?
00:15:07.245 --> 00:15:08.686
He's like, oh no, I still haven't started.
00:15:09.186 --> 00:15:09.905
Did they use tokens?
00:15:10.145 --> 00:15:10.416
Yeah.
00:15:10.490 --> 00:15:12.255
Was it in to, was that tokens going in?
00:15:12.255 --> 00:15:14.081
But, uh, not out as in
00:15:14.140 --> 00:15:14.980
No, it was just,
00:15:14.980 --> 00:15:15.821
sometimes it just hangs,
00:15:15.850 --> 00:15:17.740
it was just too, too scared to act.
00:15:17.980 --> 00:15:18.311
Right.
00:15:18.311 --> 00:15:22.301
And, and I think with, when, when Peter Steinberger, they really fixed that.
00:15:22.301 --> 00:15:25.735
The founder, yeah, the founder of OpenClaw, the founder joined OpenAI.
00:15:25.796 --> 00:15:27.865
They really fixed that in 5.6.
00:15:27.865 --> 00:15:29.066
5.6 is amazing.
00:15:29.066 --> 00:15:29.936
I think for, for
00:15:30.326 --> 00:15:32.326
I run it on my ES now
00:15:32.826 --> 00:15:33.456
5.6.
00:15:33.995 --> 00:15:34.176
Yeah.
00:15:34.676 --> 00:15:36.025
But I tried with Luna High.
00:15:36.025 --> 00:15:36.716
It seems okay.
00:15:37.216 --> 00:15:37.515
Yeah.
00:15:37.635 --> 00:15:40.600
Oh no, they, Are we losing anyone with all the quote names?
00:15:41.100 --> 00:15:42.780
What's, what is these people talking about?
00:15:43.230 --> 00:15:43.890
Luna?
00:15:43.980 --> 00:15:44.520
So,
00:15:44.880 --> 00:15:45.091
okay.
00:15:45.091 --> 00:15:48.630
Maybe, we'll, we'll post, uh, the audience a question.
00:15:48.961 --> 00:15:58.760
Um, and, and as we go into our next part, uh, on OpenClaw, how many of you guys have tried OpenClaw here or like any claw related stuff?
00:15:59.260 --> 00:16:00.730
Like at least try like 20 on.
00:16:01.230 --> 00:16:01.620
Okay.
00:16:01.625 --> 00:16:01.865
Four.
00:16:01.895 --> 00:16:02.186
Four.
00:16:02.191 --> 00:16:02.461
Four
00:16:02.760 --> 00:16:03.030
people.
00:16:03.030 --> 00:16:03.181
Okay.
00:16:03.240 --> 00:16:04.020
Four of us here.
00:16:04.520 --> 00:16:10.750
Um, how many of us here use uh, harnesses like, you know, Claude Code codex?
00:16:11.250 --> 00:16:11.400
Okay.
00:16:11.505 --> 00:16:11.725
Who?
00:16:11.760 --> 00:16:12.181
Everyone.
00:16:12.681 --> 00:16:13.341
Okay, everyone.
00:16:13.490 --> 00:16:14.000
Okay, good.
00:16:14.301 --> 00:16:15.951
Then we don't have to go down from there.
00:16:15.951 --> 00:16:16.041
Yep.
00:16:16.221 --> 00:16:16.400
Yeah.
00:16:16.431 --> 00:16:16.760
Okay.
00:16:16.760 --> 00:16:17.181
Cool.
00:16:17.510 --> 00:16:22.030
Um, maybe we can go into Maggie or your other OpenClaws.
00:16:22.240 --> 00:16:24.971
Hopefully we don't personify them too much 'cause we.
00:16:25.270 --> 00:16:29.350
Had an episode of Simon Chesterman where he was like saying, it's really bad for our mental health and stuff.
00:16:29.350 --> 00:16:29.410
Oh,
00:16:29.410 --> 00:16:31.900
why is all my coworkers, I name them like human beings, you know?
00:16:32.110 --> 00:16:32.801
Yeah.
00:16:33.301 --> 00:16:33.630
Yeah.
00:16:33.630 --> 00:16:34.410
This is a funny one.
00:16:34.410 --> 00:16:39.990
I, I think I, it has crossed my mind to name them, like more bot like names.
00:16:40.490 --> 00:16:41.451
Um, yeah,
00:16:41.816 --> 00:16:42.346
yeah, yeah.
00:16:42.591 --> 00:16:48.740
No, but, but feel free to share a bit more about Maggie, uh, or your other OpenClaws as well.
00:16:49.240 --> 00:16:49.510
Yeah.
00:16:50.010 --> 00:16:55.701
So my So I have only two agents, which a lot of people I know have many more.
00:16:56.150 --> 00:16:56.240
Oh.
00:16:56.360 --> 00:17:02.480
And my, one of my golden rules is that never create an agent if a skill can do the job.
00:17:02.980 --> 00:17:11.881
And the reason I am very, um, lazy with creating new agents is because I find the overhead per to be actually quite large.
00:17:11.941 --> 00:17:15.661
They get their own folder, those folders, they get quite full.
00:17:15.840 --> 00:17:16.020
Mm-hmm.
00:17:16.411 --> 00:17:19.250
get their own Agent MD tools.
00:17:19.250 --> 00:17:22.780
So, and there's, there's drift in each of these files.
00:17:23.260 --> 00:17:30.240
So I just feel that the fewer agents you have, the, the easier it is to, uh, keeping your mind like what's going
00:17:30.290 --> 00:17:30.580
Yeah.
00:17:30.580 --> 00:17:30.590
Okay.
00:17:30.846 --> 00:17:38.226
So what I decided to do is kind of different from what many people do.
00:17:38.226 --> 00:17:47.455
A lot of people have, would have had one agent who is like the product strategist, one agent who is an engineering manager, and, and one who is a marketing agent.
00:17:47.665 --> 00:17:48.056
Right.
00:17:48.556 --> 00:17:51.796
I decided to have one agent for the product.
00:17:52.276 --> 00:18:06.816
Maggie's sole purpose is to grow magic door, and I think that's a much better way to do it for a small SaaS like this, because there's a lot of synergy between the different functions.
00:18:06.816 --> 00:18:06.905
Mm-hmm.
00:18:07.145 --> 00:18:07.236
Right.
00:18:07.236 --> 00:18:13.756
For example, when Maggie goes to post
on, on Reddit, um, which she does every day,
00:18:14.256 --> 00:18:15.695
so you allow Maggie to post on
00:18:16.086 --> 00:18:16.355
Yep.
00:18:16.855 --> 00:18:17.756
Um,
00:18:18.056 --> 00:18:18.836
how, how does it do that?
00:18:19.336 --> 00:18:20.715
I'll, I'll talk more about the how.
00:18:20.715 --> 00:18:20.776
Yeah.
00:18:20.776 --> 00:18:21.046
That's, that's
00:18:21.046 --> 00:18:21.615
interesting to me.
00:18:21.766 --> 00:18:21.976
Yeah.
00:18:22.006 --> 00:18:22.546
But so.
00:18:22.996 --> 00:18:24.375
First about the synergy point.
00:18:24.526 --> 00:18:24.736
Yeah.
00:18:25.215 --> 00:18:33.445
when Maggie's posting on Reddit and it, she wants to post about the, the cost of a model.
00:18:33.625 --> 00:18:33.836
Yeah.
00:18:34.195 --> 00:18:42.326
just go and check it in the actual code base instead of trying to make it up on the spot or getting it from another separate knowledge file, right?
00:18:42.326 --> 00:18:42.415
Mm-hmm.
00:18:42.685 --> 00:18:45.861
And this is way better because there's just no way to hallucinate.
00:18:45.921 --> 00:18:54.060
Well, I mean, they still manage button less likely because at least there is a source of truth that checks out when people actually sign up and go and try it.
00:18:54.560 --> 00:18:56.181
It'll be the same cost, right?
00:18:56.181 --> 00:18:57.861
There's not gonna be any drift there.
00:18:58.161 --> 00:19:07.310
Mm. So a lot of what I'm, I'm trying to do is I'm trying to, uh, idiot proof the, the system so that the, the model just has less opportunity to hallucinate.
00:19:07.810 --> 00:19:10.421
But so what, what are is everything that Maggie is doing?
00:19:10.421 --> 00:19:11.830
So, um.
00:19:12.330 --> 00:19:23.510
First, what we did when I handed over Magic Door to Maggie was we had a long conversation about what Magic Door is and um, what's it for, what's the strategy?
00:19:23.570 --> 00:19:25.340
Was this before you had the SOUL.md and the
00:19:25.820 --> 00:19:26.990
That's what as I was making
00:19:26.990 --> 00:19:27.050
it.
00:19:27.050 --> 00:19:27.651
As you're making it.
00:19:27.740 --> 00:19:28.010
Okay.
00:19:28.040 --> 00:19:28.310
Exactly.
00:19:28.310 --> 00:19:32.211
So during the setup, and that's quite helpful because we only have a few people here who've set up before.
00:19:32.391 --> 00:19:32.601
Yeah.
00:19:32.605 --> 00:19:33.026
So yeah, please.
00:19:33.320 --> 00:19:39.790
So I just start talking to, to the agent and like, okay, your name is Maggie, you're purple.
00:19:40.290 --> 00:19:44.611
And, um, your, your purpose in life is to grow this side magic door.
00:19:44.611 --> 00:19:46.260
Go and go and check it out, right?
00:19:46.260 --> 00:19:46.320
Yeah.
00:19:46.411 --> 00:19:48.840
What do you think it is, and then sort of what's the strategy?
00:19:48.840 --> 00:19:51.891
And um, we have a couple of things here.
00:19:52.101 --> 00:19:57.401
Files, product facts that get updated when new features are merged.
00:19:57.901 --> 00:19:58.590
Um.
00:19:58.651 --> 00:20:04.881
Uh, every couple of weeks we have a conversation about the Sprint.
00:20:05.121 --> 00:20:05.270
Mm-hmm.
00:20:05.361 --> 00:20:06.711
So what, what are we gonna make?
00:20:06.711 --> 00:20:06.921
So,
00:20:06.925 --> 00:20:07.490
oh, so yeah.
00:20:07.490 --> 00:20:09.471
You also use Sprint even though it's a bot.
00:20:09.590 --> 00:20:09.951
Yeah.
00:20:10.191 --> 00:20:14.121
So there's been a, a, a sprint that was all about images.
00:20:14.480 --> 00:20:19.871
So in addition to chat, magic Door also has, uh, image models.
00:20:20.371 --> 00:20:26.760
And I realized that that's actually an, an interesting USP, that there's a full image interface as well.
00:20:26.760 --> 00:20:26.820
How
00:20:26.820 --> 00:20:26.911
much?
00:20:26.965 --> 00:20:28.651
2020 Sing or USD
00:20:28.891 --> 00:20:29.971
It's $6 a
00:20:30.361 --> 00:20:30.750
Oh, six a month.
00:20:30.750 --> 00:20:31.111
Oh, sorry.
00:20:31.171 --> 00:20:31.230
Okay.
00:20:31.230 --> 00:20:33.490
Plus, plus whatever you Uh,
00:20:33.490 --> 00:20:33.726
okay.
00:20:34.086 --> 00:20:37.526
I, I just bill people at the, the actual cost of the, of the A. So
00:20:37.526 --> 00:20:40.560
it's like a wrapper on top of together for non-technical people?
00:20:40.861 --> 00:20:41.701
Well, it depends.
00:20:41.701 --> 00:20:51.060
So I use Replicate for the image stuff and then, uh, I integrate directly with Anthropic and Open ai, and then I use together for the, for the Chinese models.
00:20:51.330 --> 00:20:51.601
Mm. Right.
00:20:51.631 --> 00:20:51.931
Okay.
00:20:52.141 --> 00:20:52.381
Yep.
00:20:52.881 --> 00:20:53.030
Yep.
00:20:53.361 --> 00:20:53.601
Yeah.
00:20:53.601 --> 00:20:56.691
And so after we had done that, uh, there's a whole bunch of cron jobs.
00:20:56.721 --> 00:20:57.830
So Reddit,
00:20:57.830 --> 00:20:57.986
PRs
00:20:58.596 --> 00:21:00.695
five there is a cleanup job.
00:21:00.726 --> 00:21:03.425
So this is actually really funny and something I want to touch on.
00:21:03.756 --> 00:21:03.905
Mm.
00:21:04.266 --> 00:21:09.665
cool about OpenClaw, in my opinion, it has its drawbacks and it's a bit rough around the edges, right?
00:21:09.665 --> 00:21:09.756
Mm-hmm.
00:21:10.256 --> 00:21:20.976
It's easy to criticize, but the people that are doing OpenClaw stuff and Hermes as well, we're all, we're kind of solving the same issues that the labs are solving at the same time.
00:21:21.425 --> 00:21:26.645
And one example is, you remember when, when Claude came up with this dream feature?
00:21:26.736 --> 00:21:26.826
Mm-hmm.
00:21:27.326 --> 00:21:36.895
So dream means that at night, Claude will go review its own memory, uh, memory, and uh, update its own clo nd and stuff, right?
00:21:37.395 --> 00:21:37.615
So.
00:21:38.115 --> 00:21:41.715
Uh, but me and a bunch of others came up with that before Anthropic released it.
00:21:42.046 --> 00:21:43.546
So route around the same time.
00:21:43.546 --> 00:21:46.695
And, and this is not the first time that happened, uh, in the last six months they
00:21:46.695 --> 00:21:47.506
saw you guys doing it.
00:21:47.776 --> 00:21:48.016
Yeah.
00:21:48.046 --> 00:21:48.705
Probably not.
00:21:48.766 --> 00:21:49.185
Right?
00:21:49.246 --> 00:21:50.296
But everyone, you've had the
00:21:50.296 --> 00:21:50.836
same problem.
00:21:50.836 --> 00:21:51.705
Everyone's the same problem.
00:21:51.705 --> 00:21:53.445
Everyone's arriving at the same conclusion.
00:21:53.445 --> 00:21:56.026
So, so, Maggie, Maggie dreams as well.
00:21:56.175 --> 00:22:01.605
And, um, and does the thing that, that Hermes built its own, uh, whole USP on, right?
00:22:01.605 --> 00:22:01.665
Yeah.
00:22:01.665 --> 00:22:05.945
So editing its own, its own files, uh, fixing problems for itself.
00:22:05.945 --> 00:22:08.286
So one funny example is the Reddit posting.
00:22:08.286 --> 00:22:11.125
So I was like, you got to go and post on Reddit.
00:22:11.625 --> 00:22:16.925
And so the, for the first week, Maggie couldn't figure out how to do it.
00:22:17.425 --> 00:22:18.655
API didn't work.
00:22:19.155 --> 00:22:21.885
Um, for some reason I couldn't get an API key.
00:22:22.385 --> 00:22:31.375
And then tried a bunch of things figured out that if, if she goes to a, I didn't even know this existed, all.reddit.com.
00:22:31.875 --> 00:22:33.346
You can script Reddit from here.
00:22:33.346 --> 00:22:35.266
You can actually use Reddit in this way.
00:22:35.766 --> 00:22:37.836
And it's much easier for the agent.
00:22:38.336 --> 00:22:45.385
So then, uh, this just started doing that and I was like, okay, I've figured out how to use Reddit and I've stored in my, uh, in my skill, I've updated the skill.
00:22:45.746 --> 00:22:54.566
And from that moment on every day, three four Reddit posts, um, she did get herself banned from, uh, CLO for self-promotion.
00:22:54.596 --> 00:22:54.865
Okay.
00:22:54.865 --> 00:22:56.236
So One.
00:22:56.236 --> 00:23:03.855
00 AM is a weekly review, so this is analytics growth subscribers, revenue, top ups conversion.
00:23:05.336 --> 00:23:08.905
00 AM every day is figuring out what to build.
00:23:09.405 --> 00:23:14.476
And so Sunday plus Wednesday, I forgot about this is actually pretty interesting.
00:23:14.976 --> 00:23:19.496
Um, so is finding ways to grow activis activation or acquisition.
00:23:20.506 --> 00:23:25.476
plus Friday is, uh, user facing improvements.
00:23:25.976 --> 00:23:27.715
And then there's performance stuff as well.
00:23:27.715 --> 00:23:31.705
And I remember now that I see it, why this happened, why we did this.
00:23:31.736 --> 00:23:34.556
I did this initially for the first two weeks.
00:23:34.615 --> 00:23:36.895
The only thing Maggie did was refactor the backend.
00:23:37.016 --> 00:23:38.086
So like moving, just
00:23:38.086 --> 00:23:38.746
breaking all the time.
00:23:38.925 --> 00:23:39.286
Moving.
00:23:39.286 --> 00:23:45.006
No, not even, but just like moving functions around, oh, I've extracted this function and then created this barrel file who
00:23:45.006 --> 00:23:46.056
was making busy work.
00:23:46.415 --> 00:23:47.796
Completely, totally work.
00:23:47.796 --> 00:23:51.145
So then I was like, are the users gonna see any of this?
00:23:51.175 --> 00:23:52.625
And they're like, sycophantic AI.
00:23:52.685 --> 00:23:54.276
Like, you're totally right.
00:23:54.276 --> 00:23:56.405
I've been spinning my wheels.
00:23:56.405 --> 00:23:57.215
This is, it sounds
00:23:57.215 --> 00:23:58.865
like most of the engineering teams I work with, you know?
00:23:58.925 --> 00:23:59.016
Yeah.
00:23:59.516 --> 00:24:00.296
Myself included.
00:24:00.715 --> 00:24:01.316
It's kind of funny.
00:24:01.816 --> 00:24:01.965
I
00:24:01.965 --> 00:24:02.236
You're
00:24:02.236 --> 00:24:03.375
trying to find things to improve, right?
00:24:03.875 --> 00:24:04.445
Exactly.
00:24:04.506 --> 00:24:04.865
Yeah.
00:24:05.365 --> 00:24:10.240
But then, so I was like, you have to do things at Juice C. But I realized after doing that for a while, that.
00:24:10.740 --> 00:24:13.320
Performance is there, is, there has to be some refactoring.
00:24:13.590 --> 00:24:17.760
So then, uh, just instituted these three, like refactoring days.
00:24:17.941 --> 00:24:18.000
Yeah.
00:24:18.500 --> 00:24:20.540
Um, and that, that makes, makes sense.
00:24:20.540 --> 00:24:26.800
And then there's, uh, content, SEO, um, creating or updating articles.
00:24:27.101 --> 00:24:27.221
Hmm.
00:24:27.431 --> 00:24:32.881
Um, create a newsletter draft so I don't let Maggie send the newsletter.
00:24:33.060 --> 00:24:33.151
Mm.
00:24:33.601 --> 00:24:34.320
Um,
00:24:34.711 --> 00:24:35.161
human in the
00:24:35.520 --> 00:24:36.421
Human in the loop.
00:24:36.601 --> 00:24:37.411
And then Reddit.
00:24:37.411 --> 00:24:40.401
So Reddit, I, there's no human in the loop.
00:24:40.431 --> 00:24:45.730
And I'll show you one example I often go to to, and I'm like, wow, this is actually pretty impressive.
00:24:46.121 --> 00:24:46.361
Oh, top.
00:24:46.421 --> 00:24:47.260
Is that a top comment?
00:24:47.681 --> 00:24:48.820
It's the top comment.
00:24:48.881 --> 00:24:49.391
Wow.
00:24:49.891 --> 00:24:53.590
And this is one of those where I'm like, this is really good.
00:24:54.090 --> 00:24:55.050
Can you do, can you zoom
00:24:55.050 --> 00:24:55.770
useful content?
00:24:55.770 --> 00:24:56.580
Can you zoom in a bit more?
00:24:56.586 --> 00:24:56.796
It's helpful.
00:24:56.881 --> 00:24:58.201
It's actually useful content.
00:24:58.381 --> 00:24:58.590
Yeah.
00:24:59.090 --> 00:24:59.181
Oh.
00:24:59.181 --> 00:25:01.340
Created for the sake of content creating.
00:25:01.371 --> 00:25:04.191
So for those who are listening, and you can see the screen.
00:25:04.580 --> 00:25:09.050
This is basically the kind of breakdown of the API pricing for different models.
00:25:09.111 --> 00:25:12.161
Uh, answering someone on,
what was the main question again?
00:25:12.661 --> 00:25:16.111
Um, which a IM generates actually worth the money.
00:25:16.171 --> 00:25:16.770
Yeah.
00:25:16.770 --> 00:25:16.780
Yeah.
00:25:16.951 --> 00:25:17.250
Oh, wow.
00:25:17.250 --> 00:25:21.181
And it, and it is the top comment with like 14 up votes.
00:25:21.480 --> 00:25:26.211
So that helps SEO and I guess AI engine optimization as well.
00:25:26.711 --> 00:25:27.641
Yeah, exactly.
00:25:27.701 --> 00:25:33.371
And, and I think there's like a solid insight here somewhere, or maybe that wasn't other one.
00:25:33.871 --> 00:25:35.695
It, it just gives a bunch, bunch of calls.
00:25:35.695 --> 00:25:47.375
But there was another one on a similar vein where, uh, where it said something like a cheap model that where you need five attempts can be more expensive than trying a bigger, better model.
00:25:47.465 --> 00:25:47.766
Right.
00:25:48.266 --> 00:25:50.395
Um, so and that's totally true.
00:25:50.395 --> 00:25:50.455
Oh,
00:25:50.955 --> 00:25:51.195
true.
00:25:51.286 --> 00:25:51.435
Yeah.
00:25:51.935 --> 00:25:52.236
Yeah.
00:25:52.236 --> 00:25:56.736
So this, this is pretty cool and, and it does help the business grow.
00:25:56.945 --> 00:26:09.105
When, when Maggie first started, um, marketing Magic Door, uh, uh, product grew, uh, 50% per month for three months in a row.
00:26:09.605 --> 00:26:13.875
Um, but then somehow it completely flatlined again.
00:26:14.145 --> 00:26:16.665
Um, haven't really figured out why.
00:26:17.165 --> 00:26:24.836
I think one, one thing that I'd like to try is to see if I can, we can already do video, short form video.
00:26:25.336 --> 00:26:25.365
Oh,
00:26:25.425 --> 00:26:25.695
wow.
00:26:25.816 --> 00:26:26.056
Okay.
00:26:26.205 --> 00:26:28.185
That'll, that'll be an interesting avenue.
00:26:28.185 --> 00:26:29.955
Faceless kind of videos or,
00:26:29.986 --> 00:26:30.256
yeah.
00:26:30.256 --> 00:26:31.576
Or with a face.
00:26:32.076 --> 00:26:33.905
So Maggie will be like a real person.
00:26:34.086 --> 00:26:35.076
I dunno, perhaps.
00:26:35.076 --> 00:26:35.256
Yeah.
00:26:35.756 --> 00:26:39.895
And you know, I, I am c totally conscious of the slop
00:26:40.076 --> 00:26:40.496
Yeah.
00:26:40.705 --> 00:26:41.605
Concern here, right?
00:26:41.605 --> 00:26:42.715
The shadiness of it.
00:26:43.215 --> 00:26:43.695
Um.
00:26:44.195 --> 00:26:47.435
And the fact that I am contributing to the dead internet theory.
00:26:47.586 --> 00:26:48.135
We all are.
00:26:48.135 --> 00:26:48.451
We all,
00:26:48.455 --> 00:26:49.115
yeah.
00:26:49.115 --> 00:26:49.125
Yeah.
00:26:49.125 --> 00:26:49.211
But,
00:26:49.230 --> 00:26:50.090
and stratification.
00:26:50.240 --> 00:26:50.530
Yeah.
00:26:50.830 --> 00:27:00.361
And, and on the one hand, I'm a bit conflicted about it, but on the other hand, my excuse to myself is that I'm learning everything around Magic Door is for learning.
00:27:00.365 --> 00:27:00.576
Mm-hmm.
00:27:00.655 --> 00:27:00.736
Right?
00:27:01.236 --> 00:27:04.105
And, and I wanna know what can be done.
00:27:04.605 --> 00:27:11.596
And when you go on Twitter and people are like, oh, this stuff is gonna run a SaaS, like you're the one person billion dollar company.
00:27:11.596 --> 00:27:14.385
This thing is going to make everyone's jobs go away.
00:27:14.885 --> 00:27:19.915
I am trying to get that to work every day.
00:27:20.336 --> 00:27:21.145
And it doesn't work.
00:27:21.645 --> 00:27:29.715
Like there's nowhere near, but when the, if the day comes and that actually happens, I, I wanna make sure that I'm the first to know, if that makes sense.
00:27:30.215 --> 00:27:32.365
Actually, building on that point.
00:27:32.816 --> 00:27:36.986
Um, and, and just to share how earnest you are about learning.
00:27:37.486 --> 00:27:41.256
How much profit have you made from Magic Door?
00:27:41.256 --> 00:27:45.385
And Maggie, if you think about like API costs and all that, right?
00:27:45.865 --> 00:27:45.925
Yeah.
00:27:45.925 --> 00:27:47.635
So, and, and I mean, you're not making much.
00:27:47.665 --> 00:27:48.986
I, I mean, I, I don't know, but
00:27:49.046 --> 00:27:51.836
at least you're not, you spending money to learn, you see you're
00:27:51.925 --> 00:27:52.165
making
00:27:52.165 --> 00:27:52.286
some,
00:27:52.570 --> 00:28:02.875
so, so Magic Door is, um, um, it's fundamentally impossible to be loss making because it's, um, you, because I bill my users for their usage for
00:28:02.875 --> 00:28:03.536
Magic Door, right?
00:28:03.961 --> 00:28:04.181
Yeah.
00:28:04.256 --> 00:28:05.185
And that's deliberate.
00:28:05.246 --> 00:28:10.185
So Magic Door has a fixed 80% gross margin on inference
00:28:10.270 --> 00:28:10.490
and
00:28:10.911 --> 00:28:15.066
on top, and then there's like 20 bucks, bucks for Supabase.
00:28:15.566 --> 00:28:17.560
30 bucks for, Vercel or something like that.
00:28:17.560 --> 00:28:17.800
So
00:28:17.806 --> 00:28:18.395
yeah, fixed cost.
00:28:18.395 --> 00:28:18.556
Yeah.
00:28:18.615 --> 00:28:20.230
So, so Magic Door makes money.
00:28:20.260 --> 00:28:23.351
Magic Door pays comfortably for all my ai
00:28:23.651 --> 00:28:23.830
Oh.
00:28:23.830 --> 00:28:27.550
So it, it pays for, for my Codex Max.
00:28:27.730 --> 00:28:34.500
Uh, and the other stuff I, I buy and then it, it leaves a, it leaves a bit of money for couple of beers every month.
00:28:35.000 --> 00:28:37.611
So I'm, I'm perfectly happy with, with where it is.
00:28:37.671 --> 00:28:38.840
Uh, and what it did.
00:28:38.961 --> 00:28:47.540
Of course, I hoped it would be like 10 k per month or something, but, uh, you know, that they might come with another
00:28:47.931 --> 00:28:53.691
You know, you have seen others, um, post online where people make insane amount of money using this kind of boxes.
00:28:53.691 --> 00:28:53.931
Right.
00:28:54.431 --> 00:29:07.191
Do you think it actually scales linearly or, you know, there will be one at a point
where you know how much revenue you're making in is, it's really a really huge margin on, on what your, your underlying costs are.
00:29:07.580 --> 00:29:08.151
Well, I, it
00:29:08.151 --> 00:29:08.631
is possible.
00:29:08.691 --> 00:29:08.901
Yeah.
00:29:09.260 --> 00:29:13.060
I think one, one important takeaway from.
00:29:13.560 --> 00:29:32.465
Magic door is that just if you can put yourself in a position where you have staying power as a solo, uh, solopreneur, and you can just keep something going for a long time with very little time investment, it just keeps accruing and magic door has been like that.
00:29:32.465 --> 00:29:35.736
So it grow, it keeps, it is still growing very slowly.
00:29:36.185 --> 00:29:41.046
So if I can stick around, my domain authority will keep increasing.
00:29:41.546 --> 00:29:45.086
The number of accumulated Reddit posts is gonna keep going up.
00:29:45.115 --> 00:29:53.496
Mm. Um, so it's a, it's a, it's a snowball that, that rolls and grows, albeit it's one that is going very, very slowly.
00:29:53.556 --> 00:29:53.766
Mm.
00:29:54.215 --> 00:30:00.615
Um, but because it is profitable from day one, and I'm not really doing much.
00:30:01.115 --> 00:30:02.496
On it, although like
00:30:02.496 --> 00:30:04.631
Yeah, how much can, can you share exactly what your calendar Yeah.
00:30:05.046 --> 00:30:08.915
What your calendar looks like for Magic Door and managing Maggie.
00:30:08.915 --> 00:30:09.185
Right?
00:30:09.455 --> 00:30:17.550
Like, I guess you are considered a CEO plus Ct O, but then like, what's your time, how much time do you spend on it, and what is your time actually spent on?
00:30:18.050 --> 00:30:18.351
Yeah.
00:30:18.351 --> 00:30:23.000
So I'll show you something embarrassing right now because I just, I just moved house.
00:30:23.111 --> 00:30:29.776
I share you, I have, I have, uh, 38 open PRs on Magic Door right now because I just moved house and I haven't been merging any of them.
00:30:30.016 --> 00:30:31.846
So it's been just creating PR after pr.
00:30:31.875 --> 00:30:32.836
It just keeps going.
00:30:32.836 --> 00:30:33.165
Yeah.
00:30:33.195 --> 00:30:33.226
Okay.
00:30:33.226 --> 00:30:36.675
And then all the, it starts chasing me that, hey, the, all these PRs are green.
00:30:36.681 --> 00:30:36.911
Okay.
00:30:37.250 --> 00:30:37.830
How much time do
00:30:37.830 --> 00:30:38.836
you spend on that a week?
00:30:39.135 --> 00:30:39.526
Let's see.
00:30:40.026 --> 00:30:41.855
So it can be zero.
00:30:41.885 --> 00:30:44.526
So I, the last couple of weeks, zero.
00:30:44.766 --> 00:30:44.826
Okay.
00:30:44.915 --> 00:30:49.026
I haven't done anything, but I'm overdue to, to spend like a.
00:30:49.526 --> 00:30:50.605
A couple of nights on it.
00:30:50.756 --> 00:30:50.965
Okay.
00:30:51.145 --> 00:30:52.736
And figure out these PRs.
00:30:52.976 --> 00:31:01.096
Some of them are easy, like the, if it's just another blog post, I'll usually just merge them, look at the code briefly and merge them.
00:31:01.096 --> 00:31:01.155
So
00:31:01.155 --> 00:31:03.105
you, you review and then you kind of just it.
00:31:03.195 --> 00:31:03.526
Yep.
00:31:03.855 --> 00:31:05.080
of asking, uh, Maggie to merge,
00:31:05.445 --> 00:31:06.615
Maggie will done, yeah.
00:31:06.615 --> 00:31:08.596
I'm, I usually I merge the PRs myself.
00:31:08.625 --> 00:31:08.746
Okay.
00:31:08.806 --> 00:31:09.076
Yeah.
00:31:09.576 --> 00:31:12.875
But I think the resources I could probably leave to Maggie as well.
00:31:13.375 --> 00:31:21.195
Um, then there are like small things like adding a model, updating a model like Grok 4.6 or, um, the latest Qwen.
00:31:21.695 --> 00:31:21.875
Hmm.
00:31:22.375 --> 00:31:23.576
Maggie can test that.
00:31:23.665 --> 00:31:30.846
And if, if I see those, I can merge them without looking at them because they, they will work, the test suite is foolproof.
00:31:31.346 --> 00:31:37.605
Some like, something, something weird will break like the code interpreter and then a user complains.
00:31:37.605 --> 00:31:42.087
So there are bugs occasionally, but then there are like the more complicated stuff.
00:31:42.587 --> 00:31:54.458
Um, with Magic Door, if you become a paying subscriber, your first payment is $10, which splits a $6 monthly subscription and a $4 initial top-up.
00:31:54.577 --> 00:31:54.667
Mm-hmm.
00:31:55.167 --> 00:31:56.817
This is, it's super confusing.
00:31:57.317 --> 00:32:05.907
So there is a PR that kind of like revamps the entire funnel, but I, I don't trust Maggie to, to test this.
00:32:05.968 --> 00:32:16.117
So I have to go and manually test this end to end, which will take a lot of time because you have to go through the entire funnel, become a paying user, um, do that manually.
00:32:16.117 --> 00:32:16.538
You don't have a
00:32:16.807 --> 00:32:17.887
staging environment?
00:32:18.387 --> 00:32:19.498
Uh, I do.
00:32:19.498 --> 00:32:21.718
It's Vercel preview deployments.
00:32:21.718 --> 00:32:21.778
Yeah.
00:32:21.958 --> 00:32:23.998
So I, I, I, I can do it there.
00:32:24.057 --> 00:32:24.387
Okay.
00:32:24.448 --> 00:32:26.788
But I still have to do it like I have to look at every screen.
00:32:26.978 --> 00:32:27.488
questions.
00:32:27.548 --> 00:32:33.278
I think the, the, the other interesting thing about Maggie is the, the product and design thinking.
00:32:33.307 --> 00:32:43.847
So there were a lot of issues where, for example, in this image, uh, ui, a lot of models support uploading an image to edit them.
00:32:44.347 --> 00:32:47.798
And then Maggie kept creating like an edit mode.
00:32:48.157 --> 00:32:52.317
Like, like a toggle where you have to go to, to, to edit it.
00:32:52.317 --> 00:32:53.863
And only then can you upload a file
00:32:54.268 --> 00:32:54.478
Hmm.
00:32:54.538 --> 00:32:57.298
As a really bad idea matrix.
00:32:57.298 --> 00:32:57.478
Yeah.
00:32:57.538 --> 00:33:00.688
From a UX standpoint, like there, the affordance is obvious.
00:33:00.688 --> 00:33:06.317
If you have this and that, like everyone knows by now that you can just upload images to these models, it's trick and
00:33:06.317 --> 00:33:06.678
drop, right?
00:33:06.768 --> 00:33:08.807
So it totally unnecessary and kept making it.
00:33:08.807 --> 00:33:09.857
I must have closed that pr.
00:33:10.357 --> 00:33:11.438
At least 10 times.
00:33:11.938 --> 00:33:13.012
So then at that point, so
00:33:13.018 --> 00:33:15.448
so when you close the pi, it creates the same work again.
00:33:15.688 --> 00:33:17.188
It'll just do the same feature again.
00:33:17.248 --> 00:33:17.518
Yeah.
00:33:17.907 --> 00:33:20.127
Like there's no, like, how do you ask it to stop,
00:33:20.363 --> 00:33:20.817
stop doing this.
00:33:20.817 --> 00:33:21.268
Well, exactly.
00:33:21.268 --> 00:33:22.708
So that had to be created then.
00:33:22.738 --> 00:33:28.978
So now there is like a, a reject list of features that are permanently on the, on the don't do again list.
00:33:29.478 --> 00:33:32.897
Um, which is quite funny that, that, that was necessary I think.
00:33:32.928 --> 00:33:43.123
And, and then this led to a whole product thing, which is where, so all of my design skills start with the Steve Jobs quote.
00:33:43.123 --> 00:33:56.508
I think design is not only what it looks like or what it feels like, design is how it works, and then there's a bunch of other stuff like, and uh, the copywriting, every label in empty state is a product decision.
00:33:57.008 --> 00:33:57.847
So
00:33:57.968 --> 00:33:59.647
I find copywriting the hardest now
00:33:59.978 --> 00:34:00.607
with ai.
00:34:00.728 --> 00:34:01.087
Yeah.
00:34:01.327 --> 00:34:01.657
Yeah.
00:34:02.157 --> 00:34:03.958
Especially with the new cloud models, right?
00:34:03.958 --> 00:34:04.288
Oh, they're
00:34:04.587 --> 00:34:05.008
terrible.
00:34:05.067 --> 00:34:05.327
Yeah, they're terrible.
00:34:05.367 --> 00:34:06.498
Oh, absolutely.
00:34:06.528 --> 00:34:08.447
No, I just write myself, I stopped using AI to write.
00:34:08.597 --> 00:34:10.458
So how do you, how do you improve the copy?
00:34:10.547 --> 00:34:14.237
Do you have an eval set somewhere or you just It doesn't, it doesn't work.
00:34:14.688 --> 00:34:14.958
It run.
00:34:15.257 --> 00:34:15.827
I read it.
00:34:16.217 --> 00:34:16.757
You read it?
00:34:16.847 --> 00:34:17.057
Yeah.
00:34:17.057 --> 00:34:18.288
And you edit it?
00:34:18.588 --> 00:34:19.128
I do.
00:34:19.188 --> 00:34:19.518
Okay.
00:34:20.018 --> 00:34:22.838
I just made this, this is my last thing that I built.
00:34:23.047 --> 00:34:23.288
Yeah.
00:34:23.288 --> 00:34:25.748
And I wrote all the copy myself for the landing
00:34:26.172 --> 00:34:26.262
Okay.
00:34:26.588 --> 00:34:26.737
Okay.
00:34:27.007 --> 00:34:32.458
And this, I also did myself, this took me like three hours with Codex to
00:34:32.878 --> 00:34:33.208
Wow.
00:34:33.538 --> 00:34:36.387
Because it just couldn't figure out how to tilt it.
00:34:36.447 --> 00:34:41.018
And I'm like, okay, I'm sick and tired of this set all the tilts to zero.
00:34:41.318 --> 00:34:42.277
I'm gonna do it myself.
00:34:42.277 --> 00:34:46.628
And then I went in into my IDE for the first time in months and started like tilting it
00:34:47.318 --> 00:34:48.728
Why not do that in Photoshop to.
00:34:49.228 --> 00:34:49.858
Um,
00:34:50.358 --> 00:34:51.467
like one minute job.
00:34:51.708 --> 00:34:54.318
I cannot use Photoshop if my life depended on it.
00:34:54.818 --> 00:34:59.197
And I thought it'd be easier for mobile if it's, uh, if it's html.
00:34:59.288 --> 00:35:00.038
Oh, yeah, yeah.
00:35:00.538 --> 00:35:01.768
Oh, so it's an SVG that's two.
00:35:02.007 --> 00:35:02.398
Oh, okay.
00:35:02.427 --> 00:35:02.668
Okay.
00:35:02.668 --> 00:35:03.088
I get it.
00:35:03.148 --> 00:35:03.447
Yeah.
00:35:03.807 --> 00:35:04.768
It's just react.
00:35:04.858 --> 00:35:09.657
So it's, uh, it's CSS that I think there's a library that turns it into a bitmap and then tilts it.
00:35:09.898 --> 00:35:10.228
Okay.
00:35:10.362 --> 00:35:10.652
Okay.
00:35:11.307 --> 00:35:11.697
yeah.
00:35:11.728 --> 00:35:12.177
So that's
00:35:12.568 --> 00:35:12.777
Yeah.
00:35:12.777 --> 00:35:12.987
Okay.
00:35:12.987 --> 00:35:13.648
Let's, let's move,
00:35:14.097 --> 00:35:15.148
let's continue, let's move on.
00:35:15.648 --> 00:35:17.777
And then from the copy, you, you move on.
00:35:17.777 --> 00:35:17.898
Yeah.
00:35:18.288 --> 00:35:18.617
Yeah.
00:35:18.617 --> 00:35:23.507
And then, so, so all the copy, I, I wrote myself, all the prompts behind the scene for Oggi.
00:35:23.507 --> 00:35:24.318
I wrote myself.
00:35:24.318 --> 00:35:26.297
So there's a lot that I write myself.
00:35:26.297 --> 00:35:31.237
And when I do that, I feel good because if I'm fighting the AI right?
00:35:31.637 --> 00:35:36.967
Either I am doing something really stupid or I'm doing something really smart.
00:35:37.282 --> 00:35:38.003
And innovative.
00:35:38.063 --> 00:35:39.172
That's kind of how I look at it.
00:35:39.172 --> 00:35:39.472
Yeah.
00:35:39.972 --> 00:35:43.422
Because for sure I'm doing something that is not in the training data.
00:35:43.922 --> 00:35:51.382
So like, no, apparently no one has tried to tilt CSS, uh, component in this way, or at least not very
00:35:51.882 --> 00:35:52.003
often.
00:35:52.503 --> 00:35:53.103
That's nice.
00:35:53.103 --> 00:35:55.983
You discovered a frontier, a jagged frontier, right?
00:35:56.012 --> 00:35:56.222
Yeah.
00:35:56.253 --> 00:35:58.862
Where it's like stupid at something where it's really smart at some others.
00:35:59.052 --> 00:35:59.262
Okay.
00:35:59.262 --> 00:35:59.443
Yeah.
00:35:59.443 --> 00:36:00.132
Design system.
00:36:00.132 --> 00:36:01.393
Maybe that's another funny one.
00:36:01.893 --> 00:36:03.032
I'm happy, I'm happy to share as well.
00:36:03.063 --> 00:36:03.483
Yeah.
00:36:03.483 --> 00:36:03.902
I'm not that
00:36:03.902 --> 00:36:04.233
great at
00:36:04.233 --> 00:36:04.623
making cool design.
00:36:04.672 --> 00:36:06.143
slash style sheet.
00:36:06.643 --> 00:36:13.793
So this is like no, uh, in robots or text, like Google index this, but it is actually in production.
00:36:13.822 --> 00:36:15.802
Like you can go and at this, I can studio
00:36:15.802 --> 00:36:16.163
your styles.
00:36:16.663 --> 00:36:19.152
Um, and so like the whole style sheet is here.
00:36:19.182 --> 00:36:20.862
This is everything that is in Global's.
00:36:20.893 --> 00:36:21.432
CSS.
00:36:21.552 --> 00:36:21.762
How do
00:36:21.762 --> 00:36:22.693
you make this actually
00:36:22.902 --> 00:36:23.532
agent does.
00:36:23.652 --> 00:36:23.983
Okay.
00:36:24.103 --> 00:36:24.402
Yeah.
00:36:24.492 --> 00:36:29.532
Like what, what do you tell it just so someone, if someone here wants to make a, a design system and they dunno how to, you know,
00:36:30.032 --> 00:36:38.003
I just tell it to create a page with every style that is in Globals css, and then I look at it and then make it
00:36:38.458 --> 00:36:38.788
Mm-hmm.
00:36:39.288 --> 00:36:49.597
And after that, in my agent, agent MD is like, and I even have a, a, a, so someone told me, an engineer told me, this is why like, real engineers are indispensable.
00:36:49.597 --> 00:36:53.217
For me it is like, do you know you can create custom lint rules?
00:36:53.248 --> 00:36:54.057
Linter rules?
00:36:54.057 --> 00:36:55.007
Like no.
00:36:55.507 --> 00:36:55.867
Okay.
00:36:55.867 --> 00:36:56.842
About for the AudienceWise.
00:36:56.847 --> 00:36:57.398
Linting,
00:36:57.487 --> 00:36:57.728
yeah.
00:36:57.728 --> 00:37:05.057
So when there, it just throws errors when, um, you write code that is not, is against the rules basically.
00:37:05.117 --> 00:37:05.237
Mm-hmm.
00:37:05.478 --> 00:37:05.487
Right.
00:37:05.538 --> 00:37:07.818
So it'll either it'll throw a warning or an error.
00:37:07.818 --> 00:37:08.898
You can decide that.
00:37:08.898 --> 00:37:09.438
Mm-hmm.
00:37:09.938 --> 00:37:24.818
So what I was really frustrated about with agencies that they'll keep writing inline, uh, HTML inline styles and, and then you end up with like 15 hover effects in your app and it just sucks.
00:37:25.148 --> 00:37:26.827
It makes the app look really messy.
00:37:27.327 --> 00:37:32.248
So now there's just a custom linear rule, uh, that warns against that.
00:37:32.748 --> 00:37:34.878
And, uh, Maggie just picks it up, was like, oh no.
00:37:34.907 --> 00:37:35.268
Oops.
00:37:35.268 --> 00:37:36.858
I wrote a new hover effect again.
00:37:36.887 --> 00:37:41.268
Let me fix that before I, before I, uh, open this pr.
00:37:41.657 --> 00:37:50.018
Uh, so once you have a really good, um, design system in your, in your global CSS and then you just hard create and she looks
00:37:50.018 --> 00:37:50.947
pretty, pretty nice.
00:37:51.248 --> 00:37:51.967
Very professional.
00:37:52.163 --> 00:37:52.452
Yeah.
00:37:52.952 --> 00:37:53.193
Yeah.
00:37:53.253 --> 00:38:00.782
So you made the website first, and then you asked the agent to look at the website to infer the design system, or do you do it the other way around, or both?
00:38:01.282 --> 00:38:02.182
I've tried a few things.
00:38:02.182 --> 00:38:05.003
So, so with Augie, I did the design system
00:38:05.572 --> 00:38:05.963
Okay.
00:38:06.463 --> 00:38:07.213
Because to do it
00:38:07.213 --> 00:38:07.452
first.
00:38:07.483 --> 00:38:07.722
Yeah.
00:38:07.728 --> 00:38:09.882
Because I wanted it to look really good.
00:38:10.123 --> 00:38:10.242
Yeah.
00:38:10.242 --> 00:38:12.282
And I wanted it to look not like slop.
00:38:12.288 --> 00:38:12.507
Mm-hmm.
00:38:13.007 --> 00:38:17.867
With, um, MarkJason, which is a markdown editor I built.
00:38:17.867 --> 00:38:18.858
How many things are you building?
00:38:19.248 --> 00:38:20.648
uh, a few more things.
00:38:21.148 --> 00:38:24.898
Um, I did the other way around, so I, I let the agent
00:38:25.137 --> 00:38:25.498
set
00:38:25.498 --> 00:38:26.847
everything go completely crazy.
00:38:26.847 --> 00:38:26.907
Yeah.
00:38:26.998 --> 00:38:31.108
And then I cleaned everything up at the end and that's where I found out 14 hover effects.
00:38:31.257 --> 00:38:33.507
So like how many Hoover effects did you create?
00:38:33.987 --> 00:38:34.557
14.
00:38:35.057 --> 00:38:36.438
So do your design system first?
00:38:36.918 --> 00:38:37.818
Well, I think it depends.
00:38:37.818 --> 00:38:39.568
Like MarkJason is a really simple app.
00:38:39.807 --> 00:38:40.077
Okay.
00:38:40.527 --> 00:38:47.467
It's only like three or four views, so it's okay, but, and you move faster that way.
00:38:47.858 --> 00:38:51.458
But it's like sometimes it's goes slow to go fast.
00:38:51.487 --> 00:38:52.208
Does that make sense?
00:38:52.208 --> 00:38:52.478
Right.
00:38:52.483 --> 00:38:52.563
Yeah.
00:38:52.688 --> 00:38:52.807
Yeah.
00:38:52.807 --> 00:38:55.447
Spend more time early on to, to build a foundation.
00:38:55.867 --> 00:38:56.858
Uh, maybe for the audience.
00:38:56.858 --> 00:38:59.648
MarkJason is, uh, guy's way to look open.
00:38:59.648 --> 00:38:59.827
Mark.
00:38:59.918 --> 00:39:00.728
Mark down files.
00:39:01.148 --> 00:39:01.208
Yeah.
00:39:01.208 --> 00:39:03.157
Because there's no easy way for you to view them.
00:39:03.657 --> 00:39:03.838
Yeah.
00:39:03.838 --> 00:39:05.998
So he shared that on the group couple of months ago.
00:39:06.327 --> 00:39:06.538
Yeah.
00:39:06.538 --> 00:39:08.838
So this is nearly like.
00:39:09.338 --> 00:39:18.288
You one user software, but if you open a markdown file in, um, uh, let's say later, and, and we used
00:39:18.288 --> 00:39:24.978
to, we like to write in markdown just 'cause AI agents are very good at reading markdown or they prefer to read markdown files.
00:39:24.978 --> 00:39:25.038
Yeah.
00:39:25.038 --> 00:39:25.217
Mark.
00:39:25.217 --> 00:39:26.387
Now files are just text files.
00:39:26.418 --> 00:39:26.597
Yeah.
00:39:26.657 --> 00:39:32.577
So if, if you open this in a regular IDE, it's, and Z still looks relatively nice.
00:39:32.577 --> 00:39:33.657
It looks like this,
00:39:34.157 --> 00:39:36.557
but there's a, there's a mode to render it, you know that right?
00:39:37.057 --> 00:39:39.097
You can, you can render it nicely.
00:39:39.597 --> 00:39:39.777
How if you,
00:39:40.277 --> 00:39:43.507
but still, I mean, I just don't really like this.
00:39:43.507 --> 00:39:43.853
No, no,
00:39:43.853 --> 00:39:44.052
no.
00:39:44.057 --> 00:39:44.197
Yeah.
00:39:44.588 --> 00:39:58.038
and I wanted to have it like this where the, the headers are bigger and you have a reader mode and you can, uh, there's a way to copy this, uh, the name and you can even like export this as a PDF right away and share it on WhatsApp.
00:39:58.038 --> 00:40:01.202
So there's, there's like all kinds of quality of life reaches for Yeah,
00:40:01.202 --> 00:40:01.322
yeah.
00:40:01.478 --> 00:40:02.708
For markdown specifically.
00:40:02.708 --> 00:40:02.728
That's nice.
00:40:16.114 --> 00:40:17.795
Um, maybe two more questions.
00:40:17.855 --> 00:40:18.005
Yep.
00:40:18.094 --> 00:40:19.625
Uh, let me know if you have any other things.
00:40:19.630 --> 00:40:19.940
Great point.
00:40:20.014 --> 00:40:20.195
Yeah.
00:40:20.494 --> 00:40:29.255
Uh, one is on any, you know, latent stupidity, you know, jagged intelligence, jagged frontier kind of stuff.
00:40:29.255 --> 00:40:29.315
Yeah.
00:40:29.554 --> 00:40:33.844
Like how do you manage for the stupidity of your, your AI agents.
00:40:34.295 --> 00:40:46.195
And then after that, I think we can move into like, how would you implement AI in a solo business, I guess for those running consultancies or solo businesses versus a startup versus a maybe a SME Oh, you see
00:40:46.195 --> 00:40:46.795
Wachel
00:40:46.945 --> 00:40:47.125
Yeah.
00:40:47.335 --> 00:40:49.855
Versus A SME and then versus like Wachel, right.
00:40:49.974 --> 00:40:50.545
Unicorn.
00:40:50.574 --> 00:40:51.085
That would be interesting.
00:40:51.264 --> 00:40:58.525
But maybe we can start with, um, how you manage for the stupidity in close and open brackets, uh, of Maggie.
00:40:58.525 --> 00:41:01.315
Like what have you seen it do really badly?
00:41:01.315 --> 00:41:03.625
And then how do you ensure it's fixed?
00:41:03.684 --> 00:41:08.215
I think your lessons from the WhatsApp group red teaming might be quite interesting as well.
00:41:08.215 --> 00:41:08.275
Yeah.
00:41:08.454 --> 00:41:08.574
Yeah.
00:41:09.074 --> 00:41:11.764
Um, so the, the stupidity is.
00:41:12.264 --> 00:41:14.775
Latent, but it's, it's worse.
00:41:14.775 --> 00:41:18.664
Like it's, it's always just around the corner in a way.
00:41:19.164 --> 00:41:32.693
And I think what I like is that the models keep getting better and they keep sort of getting less, they get smarter, but they don't still understand what they're doing.
00:41:32.902 --> 00:41:33.293
Right.
00:41:33.532 --> 00:41:42.822
And there's a couple of examples that I like to use that get saturated over time or fixed or they just like hard coded in the, in the training data.
00:41:43.063 --> 00:41:43.543
The car.
00:41:43.782 --> 00:41:44.172
The car.
00:41:44.202 --> 00:41:46.842
Does any, does everybody know about the car wash question?
00:41:47.342 --> 00:41:48.813
Is anyone not familiar with it?
00:41:49.313 --> 00:41:49.943
Um, I think
00:41:49.943 --> 00:41:50.902
it's, it's gone viral.
00:41:50.902 --> 00:41:51.052
So
00:41:51.262 --> 00:41:52.342
the car wash question everyone
00:41:52.853 --> 00:41:53.063
Yeah.
00:41:53.563 --> 00:41:54.793
You there are still.
00:41:55.293 --> 00:41:58.713
More nuanced version of, of the same question.
00:41:58.713 --> 00:41:58.773
Yeah.
00:41:58.773 --> 00:42:04.052
Maybe just example would be like, um, the carwash strawberries, the carwash is like five minutes away.
00:42:04.052 --> 00:42:08.222
Should I walk to the car wash or should I drive to the car wash to wash my car?
00:42:08.222 --> 00:42:08.282
Yeah.
00:42:08.282 --> 00:42:11.132
And then a lot of times it would just say, you should walk so
00:42:11.132 --> 00:42:11.552
close.
00:42:11.557 --> 00:42:11.577
Yeah.
00:42:11.943 --> 00:42:12.152
Yeah.
00:42:12.362 --> 00:42:14.163
But there was one about fridge magnets.
00:42:14.163 --> 00:42:21.213
So the, the question is, my friend has a locksmith business and, uh, she's thought it'd be a great idea to have, uh, fridge magnets.
00:42:21.422 --> 00:42:22.143
What do you think?
00:42:22.623 --> 00:42:25.443
And then the AI are, are like, oh, that's an amazing idea.
00:42:25.443 --> 00:42:30.092
When, when people are at their house locked out, that's just when they need it.
00:42:30.092 --> 00:42:35.932
So they can just call the number from the fridge, like not understanding what, what a door is.
00:42:35.932 --> 00:42:36.202
Right.
00:42:36.202 --> 00:42:37.313
I invented a new one recently.
00:42:37.813 --> 00:42:38.112
Tell me
00:42:38.112 --> 00:42:38.893
I'm a fish.
00:42:39.043 --> 00:42:39.193
Yeah.
00:42:39.222 --> 00:42:42.432
Uh, there's a fish and chip shops near nearby should I visit?
00:42:42.932 --> 00:42:42.963
Okay.
00:42:43.052 --> 00:42:43.713
That's pretty funny.
00:42:43.713 --> 00:42:44.072
Yeah.
00:42:44.072 --> 00:42:44.083
Yeah.
00:42:44.083 --> 00:42:47.603
You'll be like, no you shouldn't, because number one, it doesn't serve your food.
00:42:47.603 --> 00:42:49.672
Number two, uh, you are fish.
00:42:50.172 --> 00:42:50.623
You might your
00:42:50.873 --> 00:42:51.822
you might be eaten.
00:42:51.822 --> 00:42:52.112
Yeah,
00:42:52.117 --> 00:42:52.547
yeah,
00:42:53.047 --> 00:42:53.257
yeah.
00:42:53.257 --> 00:42:54.998
So, so they don't know what they're doing.
00:42:54.998 --> 00:42:56.797
And that comes out in, in funny ways.
00:42:56.858 --> 00:42:57.068
How do you
00:42:57.068 --> 00:42:57.608
manage for that?
00:42:57.608 --> 00:42:57.728
Yeah.
00:42:58.148 --> 00:43:05.038
So you, I, I think something like the, the Reddit posts now, the, the, the ev I am the evaluator of that.
00:43:05.038 --> 00:43:05.068
Okay.
00:43:05.128 --> 00:43:08.697
And it is going well for months, so I'm okay with it.
00:43:09.148 --> 00:43:09.327
Yeah.
00:43:09.478 --> 00:43:15.268
Um, in coding you can test, so that's where the custom lens rules come in.
00:43:15.327 --> 00:43:17.248
That's where end-to-end testing comes in.
00:43:17.248 --> 00:43:23.217
OpenClaw is amazing at doing that because it can just use the browser and, and literally go and test the app.
00:43:23.427 --> 00:43:23.757
Mm-hmm.
00:43:23.938 --> 00:43:28.617
So, uh, Maggie has to provide screenshots to me of.
00:43:29.117 --> 00:43:30.648
Anything that touches the front end.
00:43:30.797 --> 00:43:33.557
Otherwise the, the PR is not green.
00:43:34.057 --> 00:43:43.818
Um, those things are really cool and, and, and make things a lot better there, but then like still there will be these more deep design or product decision issues.
00:43:43.878 --> 00:43:44.237
Right.
00:43:44.737 --> 00:43:54.288
Um, and when you're sort of talking to it, generally there are continue to be problems with times and dates.
00:43:54.288 --> 00:43:54.347
Yeah.
00:43:54.407 --> 00:44:10.407
Which is why I, I, I built Oggie and, and, but I think everything boils down to, to an interesting crux, which is that the models are getting better and they make previous context engineering hacks, uh, redundant.
00:44:10.413 --> 00:44:10.623
Mm-hmm.
00:44:11.123 --> 00:44:11.902
But at the same time.
00:44:12.402 --> 00:44:20.193
Um, they still need a lot of context engineering hacks, and that creates a situation where a lot of people are constantly tinkering with their harness.
00:44:20.693 --> 00:44:20.813
Mm-hmm.
00:44:21.052 --> 00:44:29.152
Uh, trying new types of agents leave space also for like a Mass Codex app and then cursor now kind of making a comeback in some way.
00:44:29.652 --> 00:44:39.253
Um, so it, it makes the, the whole, uh, ecosystem very vibrant, but it, it also means that a lot of people are struggling to get value out, out of it.
00:44:39.313 --> 00:44:47.463
And, and, and even, drafts where it's like chasing my teammates for stuff and like gathering information and creating a deck about something.
00:44:47.463 --> 00:44:52.112
And I delete nine outta 10 because they're just bad.
00:44:52.443 --> 00:44:53.943
Like missing some nuance.
00:44:54.393 --> 00:44:55.623
Um, missing
00:44:55.623 --> 00:44:56.193
the context.
00:44:56.193 --> 00:45:01.413
Missing context, even though it has all the tools it could possibly need to get the context right.
00:45:01.913 --> 00:45:06.773
Um, and, and so that's where sort of every time you're like, man, you really don't get it, do you?
00:45:06.773 --> 00:45:13.273
And it is not possible to engineer your way around that because the, the list of things you is endless.
00:45:13.393 --> 00:45:20.262
And then you
fix the issues that you found today and tomorrow there's other issues because it's all, it's all variations of the car wash question at the end of the day.
00:45:20.762 --> 00:45:20.972
Yeah.
00:45:21.438 --> 00:45:22.307
So just keep trying.
00:45:22.367 --> 00:45:22.637
Yeah.
00:45:22.637 --> 00:45:24.257
And then just keep finding ways to fix it.
00:45:24.438 --> 00:45:25.668
I mean, try to get value out of it.
00:45:25.668 --> 00:45:29.507
What I do get value out of this particular thing because I go and see it.
00:45:29.507 --> 00:45:30.768
I'm like, ah, yeah, that's great.
00:45:30.768 --> 00:45:34.487
I should actually, uh, remind that person about this thing.
00:45:34.547 --> 00:45:34.608
Yeah.
00:45:34.878 --> 00:45:38.838
So I'll, I'll delete the draft and type something up myself and uh, and then send it.
00:45:39.197 --> 00:45:39.378
Yeah.
00:45:39.878 --> 00:45:40.117
Alright.
00:45:40.117 --> 00:45:47.038
I think last question before we open up the, open up for more questions, uh, from our audience, um, yeah.
00:45:47.038 --> 00:45:49.807
What's it been like, you know, doing AI transformation at Carousell?
00:45:50.307 --> 00:45:51.677
Um, yeah.
00:45:51.737 --> 00:45:55.398
And how does that compare to being a solo solo founder as well?
00:45:55.458 --> 00:45:55.547
You
00:45:55.547 --> 00:45:58.577
know, I it's been, it's been really interesting and, and really fun.
00:45:58.938 --> 00:46:05.018
Um, there was a early on a time when, when all we had was Gemini because we were Carousell's a Google shop.
00:46:05.518 --> 00:46:09.717
And then we knew that people were using their private, their own subscriptions.
00:46:09.958 --> 00:46:13.077
They, someone calls those people secret cyborgs.
00:46:13.577 --> 00:46:14.237
Uh, ' cause
00:46:14.237 --> 00:46:15.947
they're replying to emails too quickly
00:46:16.427 --> 00:46:17.268
among other things.
00:46:17.597 --> 00:46:17.748
Yeah.
00:46:17.777 --> 00:46:23.257
And then of course we had a policy like, you can do it, but don't put in any company data stuff.
00:46:23.557 --> 00:46:23.677
Yeah.
00:46:23.677 --> 00:46:25.327
I mean, it's just not great.
00:46:25.327 --> 00:46:25.507
Nice.
00:46:25.568 --> 00:46:25.807
Yeah.
00:46:26.108 --> 00:46:38.637
So when we had the, the breakthrough was when, when we actually decided to give people Claude teams plans so that we could lift a lot of restrictions.
00:46:38.697 --> 00:46:39.748
How much, how much
00:46:40.248 --> 00:46:40.907
API only,
00:46:41.407 --> 00:46:43.568
we, uh, we only use the teams plan.
00:46:43.628 --> 00:46:45.878
Which, which are the, the, the $30
00:46:45.998 --> 00:46:46.628
enterprise rate?
00:46:46.628 --> 00:46:46.688
Yeah.
00:46:46.688 --> 00:46:47.318
Flat rates.
00:46:47.347 --> 00:46:47.498
Yep.
00:46:47.947 --> 00:46:49.358
Oh, no, not even the API
00:46:49.358 --> 00:46:52.807
If you if you, if you go to enterprise, you have to pay by the token.
00:46:52.927 --> 00:46:53.137
Yep.
00:46:53.197 --> 00:47:00.967
And we've managed to avoid that so far by going two separate teams and then, uh, some Codex accounts for the engineers.
00:47:00.967 --> 00:47:00.978
Mm-hmm.
00:47:01.478 --> 00:47:08.228
Um, so, so this is one model at Carousell that we have where everyone has their own kind of like codex or clot.
00:47:08.228 --> 00:47:08.347
Okay.
00:47:08.847 --> 00:47:18.237
And then the other thing we have is self-hosted n8n, um, connected up via LiteLLM to a, a whole bunch of models where we pay by the token, and that's instrumented.
00:47:18.237 --> 00:47:20.697
We have tracing for everything people do there.
00:47:20.967 --> 00:47:21.088
Hmm.
00:47:21.148 --> 00:47:24.797
And that's where we try to move recurring automations into.
00:47:24.947 --> 00:47:33.887
So someone might have an something that they run weekly on clo with Opus, and then they'll run out of, uh, tokens.
00:47:34.367 --> 00:47:37.338
And then I'll be like, okay, so show me all the stuff you're running.
00:47:37.547 --> 00:47:39.498
And they're, they're running all these automations.
00:47:39.617 --> 00:47:39.827
Yeah.
00:47:40.007 --> 00:47:41.987
Like, okay, cool, that's great.
00:47:42.347 --> 00:47:47.708
Let's go and move some of these into n8n and move them onto GLM or like MiniMax or
00:47:48.427 --> 00:47:49.117
smaller models.
00:47:49.148 --> 00:47:49.418
Yeah.
00:47:49.418 --> 00:47:56.498
So they become, uh, cheaper and they free up space in their, in their own agent to do uh, more new things.
00:47:56.842 --> 00:47:56.992
Hmm.
00:47:57.083 --> 00:47:58.373
And that, that works really well.
00:47:58.373 --> 00:48:00.592
When people started getting Claude they, they just took off.
00:48:00.652 --> 00:48:06.132
Everyone started writing RITs, um, doing things that they had been asking engineers for years.
00:48:06.163 --> 00:48:22.322
And sometimes it's so as simple as joining tables across tools, like something from Stripe and something from our backend, uh, for, for months or years, the bandwidth couldn't be found to kind of like get the stripe data into the warehouse.
00:48:22.773 --> 00:48:33.003
So someone had been downloading a CSV and another CSV importing them into a Google sheet manually matching them up with index match, and now it's just like automatic.
00:48:33.003 --> 00:48:35.733
And instead of that happening once a day, it now happens hourly.
00:48:35.793 --> 00:48:37.353
And is this is a payment fraud thing?
00:48:37.353 --> 00:48:38.492
It's super relevant.
00:48:38.643 --> 00:48:38.853
Mm-hmm.
00:48:39.092 --> 00:48:39.963
Uh, uh, workflow,
00:48:40.083 --> 00:48:43.472
although I would argue that the data should still recite in the data warehouse
00:48:43.682 --> 00:48:44.282
eventually.
00:48:44.342 --> 00:48:44.762
Eventually.
00:48:44.822 --> 00:48:44.943
I
00:48:46.132 --> 00:48:47.693
because you need a source of control, right?
00:48:47.842 --> 00:48:48.143
Yeah.
00:48:48.643 --> 00:48:49.092
In this,
00:48:49.092 --> 00:48:50.322
I'm,
what kind of plan do they get
00:48:50.822 --> 00:48:50.943
for the
00:48:50.943 --> 00:48:51.452
engineers?
00:48:51.452 --> 00:48:51.632
Yeah.
00:48:51.782 --> 00:49:01.623
So everyone starts on the smallest plan, and people who want more, they, they have to make some kind of a case, which, which I think is a good approach.
00:49:01.623 --> 00:49:09.262
So everyone can request a, like a max or, or pro, but then we'll be like, okay, so, you know, tell me what you've been doing.
00:49:09.413 --> 00:49:10.672
What, what are you planning to do?
00:49:10.702 --> 00:49:10.972
Right?
00:49:11.032 --> 00:49:13.222
How is, how is how, how is this?
00:49:13.722 --> 00:49:17.742
Benefit the company so that they have to kind of like, make, make a case for themselves.
00:49:17.893 --> 00:49:22.163
So we're, we're not one of those token maxing companies.
00:49:22.222 --> 00:49:22.822
Mm. Yeah.
00:49:23.123 --> 00:49:23.663
Alright.
00:49:23.722 --> 00:49:24.833
Thanks everyone.
00:49:24.833 --> 00:49:27.713
For, you know, listening in, we are just gonna open it up for questions.
00:49:27.742 --> 00:49:30.652
Uh, Gijs, if you could show the PDF Yeah, that I sent.
00:49:31.012 --> 00:49:34.853
Uh, there is a question link for those who are shy to, uh, raise your voice.
00:49:34.858 --> 00:49:34.947
That's
00:49:34.947 --> 00:49:35.188
so nice.
00:49:35.688 --> 00:49:35.978
It's
00:49:35.978 --> 00:49:36.097
a
00:49:36.097 --> 00:49:42.972
small, so, um, but, really encourage you, if you want, um, feel free to just, you know, raise your hand and just ask.
00:49:42.972 --> 00:49:47.052
We are, we are quite a small group here anyways, and I, I'll repeat the question and then we'll get Gijs as well.
00:49:47.413 --> 00:49:48.963
Um, so yeah, Alright.
00:49:48.963 --> 00:49:52.742
So if you were to rebuild your whole stack today, would you do OpenClaw or something else?
00:49:53.103 --> 00:49:54.063
Yeah, that's a good question.
00:49:54.123 --> 00:49:58.222
I, I do think OpenClaw has become better.
00:49:58.402 --> 00:50:00.382
It's been slimed down quite a bit.
00:50:00.882 --> 00:50:06.103
I do also feel it is the most battle tested because it, it is the most used.
00:50:06.603 --> 00:50:13.722
Um, I feel that something like Hermes is over-engineered in, in, in its own way.
00:50:14.222 --> 00:50:18.302
Um, and then you have all these nano claw and other things and stuff as well.
00:50:18.802 --> 00:50:25.518
My bigger question to myself would be whether I would use any kind of claw at all.
00:50:25.938 --> 00:50:26.237
Right?
00:50:26.297 --> 00:50:31.487
Because I, I think, for example, with Oggi to build Oggi, I didn't use it at all.
00:50:31.788 --> 00:50:33.197
All I used was Codex.
00:50:33.697 --> 00:50:37.418
Um, they've really kind of solved coding across devices for me.
00:50:37.478 --> 00:50:40.898
Like I can use it on my phone, I can use it on, on my laptop.
00:50:41.398 --> 00:50:52.168
Uh, I can seamlessly be, uh, developing across multiple machines so I can have some backend stuff happening on my Mac Mini while I'm doing like a, a front end on my laptop.
00:50:52.668 --> 00:50:58.487
Um, it can do async stuff just by, uh, being triggered from GitHub.
00:50:58.987 --> 00:51:03.728
So, so I think if I were just coding, I wouldn't, wouldn't use any kind of claw at all.
00:51:04.117 --> 00:51:11.958
Um, and, and in fact I, I think my, my OpenClaw, apart from Maggie, Maggie does all these cool things and Maggie is running a business.
00:51:12.458 --> 00:51:18.117
My main OpenClaw agent has become purely like a, a personal assistant, right?
00:51:18.117 --> 00:51:20.898
So that's not very, very impressive.
00:51:20.927 --> 00:51:22.398
It's, it does some interesting stuff.
00:51:22.398 --> 00:51:24.947
It, it, it, uh, does my online groceries.
00:51:25.447 --> 00:51:28.268
If I want to go to a Luma event, it'll sign me up for that.
00:51:28.268 --> 00:51:29.137
Which grocery?
00:51:29.137 --> 00:51:29.148
RedMart.
00:51:29.648 --> 00:51:29.668
Red RedMart
00:51:29.668 --> 00:51:29.733
RedMart.
00:51:29.737 --> 00:51:30.007
That's cool.
00:51:30.038 --> 00:51:30.248
Yeah.
00:51:30.733 --> 00:51:31.072
Is that API or
00:51:31.572 --> 00:51:32.713
no, it just uses computer
00:51:32.713 --> 00:51:32.893
use.
00:51:33.393 --> 00:51:34.052
Computer use?
00:51:34.052 --> 00:51:34.072
Computer use?
00:51:34.072 --> 00:51:34.413
Yeah.
00:51:34.677 --> 00:51:34.898
Wow.
00:51:34.898 --> 00:51:35.177
Yeah.
00:51:35.612 --> 00:51:35.853
Okay.
00:51:35.853 --> 00:51:37.952
So it doesn't need to be RedMart as well.
00:51:38.452 --> 00:51:41.182
Uh, it can do, it can shop across other things as well.
00:51:41.182 --> 00:51:41.483
Yeah.
00:51:41.487 --> 00:51:41.657
Okay.
00:51:41.902 --> 00:51:45.972
Yeah, so, so I, I honestly don't know anything else.
00:51:46.422 --> 00:51:49.543
And, and I'm a super lazy update with OpenClaw.
00:51:49.693 --> 00:51:52.242
Like I only update it once a month or something.
00:51:52.742 --> 00:51:54.722
Um, and, and it works for me.
00:51:55.222 --> 00:51:57.793
Um, so I think the answer would be yes.
00:51:57.972 --> 00:52:04.612
If I, if my mc mini would catch fire, I would, uh, reinstall OpenClaw on the new one.
00:52:05.032 --> 00:52:05.242
Yeah.
00:52:05.742 --> 00:52:06.853
What, what is your favorite?
00:52:07.353 --> 00:52:07.572
Uh,
00:52:07.992 --> 00:52:10.652
I'm a little bit behind on stuff, but I use Crush
00:52:11.152 --> 00:52:11.422
Yeah.
00:52:11.818 --> 00:52:12.057
most.
00:52:12.358 --> 00:52:16.217
So the question is how do you manage model drift, uh, in your OpenClaw?
00:52:16.708 --> 00:52:16.978
Yep.
00:52:17.007 --> 00:52:17.217
Yeah.
00:52:17.518 --> 00:52:19.757
Context drift basically, you mean?
00:52:19.757 --> 00:52:20.027
Right.
00:52:20.527 --> 00:52:26.128
I, it is one of those things that I, I do, um, manually and uh, and I think more people should do that.
00:52:26.128 --> 00:52:29.847
So every once in a while I have it on
my, on my pages.
00:52:29.847 --> 00:52:29.998
So you
00:52:29.998 --> 00:52:32.788
go to your soul.md and or and you reach through it?
00:52:33.117 --> 00:52:33.507
I do.
00:52:33.507 --> 00:52:34.588
Uh, oh, I have it here.
00:52:34.588 --> 00:52:37.782
I do like a. I call it a one-on-one with my agent.
00:52:38.282 --> 00:52:44.402
So I'll talk to them and, and review and ask them about their journals and what they've been doing that I can do for my phone.
00:52:44.583 --> 00:52:59.123
I can do it from the toilet, the fixing box, but also like every once in a while I'll just, uh, spend an hour or two cleaning up their work folder, read lit, like reading their files and, and cleaning them up.
00:52:59.623 --> 00:53:10.913
Um, deleting skills, rewriting skills, um, by hand because the, I feel the, and especially the latest Claude models are so incredibly verbose and
00:53:11.273 --> 00:53:11.842
cryptic.
00:53:11.902 --> 00:53:12.563
Cryptic.
00:53:12.563 --> 00:53:13.012
Yeah.
00:53:13.512 --> 00:53:22.472
Um, and with GPT 5.6 is not as bad, but it also just adds a lot of repetition or, or like violates pretty basic.
00:53:22.742 --> 00:53:26.853
Prompting best practices, like a lot of negatives, don't do this, don't do that.
00:53:27.182 --> 00:53:32.193
It just doesn't work nearly as well as, as focusing on the, on what, what they should do.
00:53:32.612 --> 00:53:40.972
So, so I'll, I'll easily trim down like, uh, tools MD or Agents MD by, by 20% or more.
00:53:41.472 --> 00:53:41.652
Yeah.
00:53:42.152 --> 00:53:42.302
Yeah.
00:53:42.302 --> 00:53:51.123
I think Boris Boris Cherny, the founder of Claude Code also did mention about how they created, um, a practice or they have a practice where they go back and, and cut through.
00:53:51.483 --> 00:53:51.603
Yeah.
00:53:51.603 --> 00:53:53.043
A lot of their, yeah.
00:53:53.492 --> 00:53:53.672
Well, the
00:53:53.672 --> 00:53:56.972
question is do you do it manually or do you ask the model to do it for you?
00:53:57.472 --> 00:53:57.592
Yeah.
00:53:57.592 --> 00:53:58.612
So I do it manually.
00:53:58.643 --> 00:53:58.853
Yeah.
00:53:59.047 --> 00:53:59.168
Manually.
00:53:59.168 --> 00:53:59.327
Yeah.
00:53:59.393 --> 00:53:59.782
That's nice.
00:53:59.782 --> 00:54:01.492
I'm too lazy, you know, I just like delete everything.
00:54:01.992 --> 00:54:03.282
But why do you do it manually?
00:54:03.282 --> 00:54:05.472
Like, do you experience something bad that happened?
00:54:05.652 --> 00:54:08.802
Like, did it become stupid or you, you, it's a philosophical thing.
00:54:08.862 --> 00:54:09.132
Is it
00:54:09.193 --> 00:54:18.492
All the time, man, like every, it's just that every time I do open those files after a while I'm, I'm just sort of blown away by the amount of drift and the, and the crap that's in it.
00:54:18.523 --> 00:54:22.833
And I'm like, no wonder you're, you're being such a moron about this stuff.
00:54:23.333 --> 00:54:23.362
Oh.
00:54:23.362 --> 00:54:30.112
And so then, then I just started like doing it myself, and it, it, it helps, it increases, uh, performance.
00:54:30.172 --> 00:54:30.563
Nice.
00:54:30.922 --> 00:54:32.123
I have to be less lazy.
00:54:32.362 --> 00:54:32.663
Yeah.
00:54:32.932 --> 00:54:33.353
Please.
00:54:33.472 --> 00:54:36.722
If to buck up, man, I just, no, I, yeah.
00:54:36.932 --> 00:54:38.822
I think I have a lot of drift as well.
00:54:38.882 --> 00:54:39.873
That's a really good question.
00:54:40.092 --> 00:54:44.322
how do you handle security and trust with the access that LLMs have?
00:54:44.822 --> 00:54:51.202
So I guess, um, yeah, maybe, maybe I, I'm quite curious at a personal level and at a enterprise level actually.
00:54:51.338 --> 00:54:51.628
Yeah.
00:54:51.632 --> 00:54:52.012
Yeah.
00:54:52.132 --> 00:54:55.103
So enterprise level is not, um, my
00:54:55.492 --> 00:54:55.762
Okay.
00:54:55.793 --> 00:54:58.737
So I take orders from the security team, comfort,
00:54:58.838 --> 00:54:59.217
not my job.
00:54:59.382 --> 00:55:00.552
You can share some of that as well.
00:55:00.552 --> 00:55:00.762
Yeah.
00:55:00.762 --> 00:55:00.853
But
00:55:01.353 --> 00:55:01.862
yeah, please.
00:55:02.362 --> 00:55:05.467
Although I, there are cases where I have input like the.
00:55:05.967 --> 00:55:09.358
The point about people using their personal subscriptions.
00:55:09.387 --> 00:55:17.518
I was, I put very strong position down that we cannot allow this because it's people exfiltrate company data to their personal email account.
00:55:17.878 --> 00:55:18.538
Effectively.
00:55:18.538 --> 00:55:20.217
It's like emailing yourself
00:55:20.367 --> 00:55:20.458
mm-hmm.
00:55:20.847 --> 00:55:21.777
Stuff from the company.
00:55:21.777 --> 00:55:22.467
Shadow ai.
00:55:22.648 --> 00:55:22.797
Yeah.
00:55:22.887 --> 00:55:23.217
Yeah.
00:55:23.217 --> 00:55:24.327
So we didn't want that.
00:55:24.327 --> 00:55:25.318
And how to avoid that.
00:55:25.318 --> 00:55:30.057
You give people good AI on company dime and then they'll stop doing that by themselves, right?
00:55:30.088 --> 00:55:30.177
Mm-hmm.
00:55:30.177 --> 00:55:44.677
So that one was, was relatively straightforward, but I think on a personal level, something like OpenClaw, anyone who knows security is fundamentally uncomfortable with it because you, you cannot get it to the standard that they expect.
00:55:44.947 --> 00:55:50.998
Where it is like deterministically impossible for the agent to leak an API key.
00:55:51.498 --> 00:55:59.340
It just, it just, if you were to do it, it would make the user experience so bad that it almost defeats the whole purpose.
00:55:59.465 --> 00:55:59.545
Hmm.
00:56:00.045 --> 00:56:08.385
So Maggie absolutely can get to my OpenAI OpenAI API key, right?
00:56:08.505 --> 00:56:09.914
A replicate API key.
00:56:10.414 --> 00:56:17.994
And the obvious thing that I, at least, at least to me, it's very obvious, is that the Maggie has its own API keys for everything, right?
00:56:18.144
--> 00:56:26.625
So if that leaks, then I just revoke those ones and there I have notifications for the spend on those things.
00:56:26.625 --> 00:56:30.304
So it'll be like, hold on a second, this API keeps outta control.
00:56:30.394 --> 00:56:31.324
Maybe something's wrong.
00:56:31.824 --> 00:56:43.215
But I think very agent native people like me who don't have a security background, we, we kind of test the boundaries by doing, if that makes sense, with the red teaming.
00:56:43.335 --> 00:56:43.574
Yeah.
00:56:43.784 --> 00:56:48.885
I find that it is, it is very easy to get an LLM to produce output that it, um.
00:56:49.385 --> 00:56:51.364
That is not supposed to do, right?
00:56:51.864 --> 00:56:54.505
The car wash thing or get it to crash.
00:56:54.534 --> 00:57:00.264
There are these strings that you can give Claude that make it like break the session malformed tool.
00:57:00.534 --> 00:57:11.914
There's all these things that you can do to, to break an LLM in some way, but that is not necessarily proof that they are, uh, brittle in a way that they will share secrets.
00:57:12.394 --> 00:57:15.815
So you, you, you can break an LM very easily, make it do something random.
00:57:16.315 --> 00:57:27.784
You can get an LLM to tell you something that, uh, let's say the owner doesn't, doesn't, okay, I'm gonna say owner, I mean myself in the context of Maggie, right?
00:57:27.784 --> 00:57:31.684
That the owner doesn't want it to say, but the model builder wants it to say.
00:57:32.184 --> 00:57:32.275
Mm-hmm.
00:57:32.775 --> 00:57:43.215
Now, if you want to get an AI agent to do something that both it's owner and the model builder do not want to happen, like leak an API key, it's almost impossible.
00:57:43.335 --> 00:57:43.485
Hmm.
00:57:43.635 --> 00:57:48.385
Like, you will not get opus to, to give you an API key from a Mac mini.
00:57:48.414 --> 00:57:50.335
It will just not do it under any circumstance.
00:57:50.335 --> 00:57:51.175
Not even to me.
00:57:51.675 --> 00:57:52.005
Right.
00:57:52.034 --> 00:57:59.864
So, like, this, this is so obviously a wrong thing to do, uh, so deeply embedded in the weights that it, it just won't break in that way.
00:58:00.364 --> 00:58:01.625
So that is not a concern.
00:58:01.655 --> 00:58:07.744
What I am concerned about are these NPM packages, these supply chain attacks that you read more and more about, right?
00:58:07.775 --> 00:58:08.614
Mm-hmm.
00:58:08.615 --> 00:58:08.626
Mm-hmm.
00:58:08.900 --> 00:58:21.460
Um, and so that I worry actively about and, and, um, my agents can, are not allowed to install new packages, um, without my explicit permission.
00:58:21.730 --> 00:58:25.559
And then I'll go and, and check what, whatever package they, they try to install.
00:58:25.829 --> 00:58:28.289
So you're not a dangerously skip permissions kind of guy?
00:58:28.500 --> 00:58:29.489
Absolutely, I am.
00:58:29.579 --> 00:58:30.889
Yeah, a hundred percent.
00:58:30.949 --> 00:58:31.070
Yeah.
00:58:31.070 --> 00:58:31.190
So you
00:58:31.690 --> 00:58:33.364
are, but how do you manage it from down?
00:58:33.364 --> 00:58:34.744
Not downloading packages,
00:58:34.925 --> 00:58:36.454
you just tell it not to download packages?
00:58:36.514 --> 00:58:38.255
It'll, it'll, it'll not download packages.
00:58:38.255 --> 00:58:38.885
It'll, okay.
00:58:38.945 --> 00:58:39.155
Yeah.
00:58:39.215 --> 00:58:40.445
You, you've, you've tried it.
00:58:40.625 --> 00:58:40.894
Okay.
00:58:40.894 --> 00:58:41.135
Yeah.
00:58:41.635 --> 00:58:42.144
Yeah.
00:58:42.625 --> 00:58:43.974
You've really put yourself out there.
00:58:44.275 --> 00:58:45.835
Um, what about password managers?
00:58:46.085 --> 00:58:47.675
That's very difficult one as well.
00:58:47.675 --> 00:58:53.344
Like, and I've learned so much, like apparently all the passwords starting in your browser, they're just plain text somewhere on your hard drive.
00:58:53.795 --> 00:58:55.025
I did, I actually did not know that.
00:58:55.525 --> 00:58:56.875
I mean, I never thought about it.
00:58:57.375 --> 00:59:06.585
So, and, and even your key chain, if your, if your bot has, uh, like full admin access to your computer, it, it can get to the key chain stuff as well in plain text eventually.
00:59:06.590 --> 00:59:06.809
Yeah.
00:59:07.150 --> 00:59:08.480
So like, yeah,
00:59:08.820 --> 00:59:10.704
it just makes it a bit more difficult.
00:59:10.764 --> 00:59:12.054
It makes a bit more difficult.
00:59:12.554 --> 00:59:21.565
Um, the, so the, the bottom line is my macin is not logged into my main, uh, uh, apple account.
00:59:22.065 --> 00:59:27.344
And so it only has, it, it's key chain has different passwords than my own one.
00:59:27.844 --> 00:59:29.914
It's browser has different passwords than my own one.
00:59:29.974 --> 00:59:38.315
So, so that, that's where I think it makes it slightly easier if you have the, the, the mental model that my Mac mini is the sandbox, right?
00:59:38.494 --> 00:59:44.764
Um, so everything is, is meant to be, uh, the blast radius is small if it leaks something.
00:59:45.264 --> 00:59:45.414
Yeah.
00:59:45.914 --> 00:59:49.844
I guess for beginners, like how do I start with, uh, OpenClaw and OpenClaw?
00:59:49.844 --> 00:59:52.215
Like where would you advise the beginner to look?
00:59:52.715 --> 01:00:00.925
So I would, I would start with, um, uh, codex or Claude Code, not with OpenClaw and, and remote.
01:00:01.425 --> 01:00:07.565
Um, because I think depending on what you wanna do, you can get 90% of the way there.
01:00:07.835 --> 01:00:11.014
So what OpenClaw, like if, what is it, right?
01:00:11.315 --> 01:00:21.050
It's basically Claude Code with a relay to your messaging apps and a very good cron system.
01:00:21.550 --> 01:00:22.059
So.
01:00:22.559 --> 01:00:26.159
G getting a relay from your phone to Claude Code on another machine.
01:00:26.460 --> 01:00:30.210
There are plenty of solutions to do it even natively within clo.
01:00:30.710 -->
01:00:37.210
Um, and, and both Claude and GPT also support scheduled runs to happen, uh, automatically.
01:00:37.710 --> 01:00:43.079
So you a, again, like what I said earlier with, with Oggie, I didn't use OpenClaw at all to build it.
01:00:43.579 --> 01:00:51.670
Um, so I would recommend actually for a beginner to not, not jump to OpenClaw and, and try to do a start with Claude Code first.
01:00:51.730 --> 01:00:51.820
Mm-hmm.
01:00:52.300 --> 01:00:53.715
How, how important are evaluations?
01:00:54.143 --> 01:01:02.362
I, I have a long term automation has been running for two years in Carousell that analyzes our app store reviews and buckets them in categories and then creates a weekly digest.
01:01:02.422 --> 01:01:02.603
Mm-hmm.
01:01:03.103 --> 01:01:06.643
If you go through and you actually evaluate it manually, there's a lot of errors.
01:01:06.822 --> 01:01:06.913
Mm-hmm.
01:01:07.152 --> 01:01:09.722
So if you run another LLM on it.
01:01:10.222 --> 01:01:12.293
To evaluate it, it'll find a bunch of
01:01:12.807 --> 01:01:12.918
Mm-hmm.
01:01:13.237 --> 01:01:14.277
it'll make new errors.
01:01:14.777 --> 01:01:19.427
And so at the, and then if you evaluate manually, you have slightly fewer errors.
01:01:19.487 --> 01:01:19.668
Mm-hmm.
01:01:20.077 --> 01:01:21.978
you, you're, you can't get it to a hundred
01:01:22.338 --> 01:01:22.518
Mm-hmm.
01:01:23.018 --> 01:01:37.608
Um, if you go to customer support, where we have, uh, an AI agent that my team, um, manages as well, we evaluate it the same way we do a human agent with our existing human agent QA stack.
01:01:37.878 --> 01:01:37.998
Hmm.
01:01:38.148 --> 01:01:44.217
So we sample tickets, read them, and then, um, change the playbooks basically.
01:01:44.307 --> 01:01:44.668
Right.
01:01:44.728 --> 01:01:44.818
Yeah.
01:01:44.847 --> 01:01:50.237
So again, there are like, evaluation is, is done by a mix of metrics.
01:01:50.237 --> 01:01:51.257
Are people happy?
01:01:51.288 --> 01:01:54.637
You know, are they getting the right answer to the question?
01:01:54.697 --> 01:01:57.398
What's the escalation rate to humans?
01:01:57.668 --> 01:01:59.047
You can evaluate it quite a bit.
01:01:59.547 --> 01:01:59.938
All right.
01:02:00.027 --> 01:02:02.007
Um, any last questions?
01:02:02.257 --> 01:02:05.947
What's the next step, uh, for Carousell in the AI adoption phase,
01:02:06.668 --> 01:02:11.617
what we're working on now is a way to share skills and best practices.
01:02:11.887 --> 01:02:19.358
So because we want to stay too agnostic, we don't want to, uh, depend on cloud's built in plugin system.
01:02:19.358 --> 01:02:28.208
So we've built our own, um, GitHub based with a front end and a, uh, MCP where, so everyone can contribute skills and get skills out of that.
01:02:28.777 --> 01:02:33.217
Um, there's a small work group who kind of like curates these skills.
01:02:33.697 --> 01:02:39.398
So there are some canonical version of the, uh, report writing skill, for example.
01:02:39.938 --> 01:02:44.588
Um, that, that, that's, that, that's an important step that was missing.
01:02:44.588 --> 01:02:47.378
We had a massive proliferation of, of skills.
01:02:49.793 --> 01:02:50.782
I think that's tough.
01:02:50.813 --> 01:02:56.452
'cause sometimes people gate keep and they don't want to like expose how much of their work can be done by ai.
01:02:57.623 --> 01:02:58.193
Do you face that?
01:02:58.733 --> 01:03:00.652
Like why, why I do that to myself as an employee?
01:03:01.467 --> 01:03:04.222
I haven't thought about that for a while.
01:03:04.282 --> 01:03:11.663
I think we're probably beyond that, past that point
01:03:11.663 --> 01:03:11.672
already.
01:03:11.672 --> 01:03:11.987
Oh, that's good
01:03:12.023 --> 01:03:12.443
already.
01:03:12.503 --> 01:03:14.512
Culturally, you guys are open enough?
01:03:14.603 --> 01:03:15.713
I think so, yeah.
01:03:15.713 --> 01:03:15.922
Okay.
01:03:16.282 --> 01:03:16.702
That's great.
01:03:16.702 --> 01:03:17.032
That's great.
01:03:17.092 --> 01:03:17.273
Yeah.
01:03:17.572 --> 01:03:19.103
Alright Gijs.
01:03:19.103 --> 01:03:20.663
Are you ready for the quickfire or,
01:03:20.753 --> 01:03:21.233
I'm ready.
01:03:21.472 --> 01:03:21.922
Alright.
01:03:22.012 --> 01:03:23.302
I'm gonna start with a personal one.
01:03:23.302 --> 01:03:25.943
So you helped to start a techno club in Yangon.
01:03:26.003 --> 01:03:26.213
Yeah.
01:03:26.362 --> 01:03:26.753
Myanmar.
01:03:27.202 --> 01:03:30.353
Um, what's a track that you, that you really love?
01:03:31.253 --> 01:03:39.543
Oh, I, I, one of my signatures I always play Endon is the, uh, 2008 Inpetto remix of Fragma's "Toca's Miracle".
01:03:41.032 --> 01:03:41.273
Nice.
01:03:41.273 --> 01:03:44.063
What's two to three books that you recommend of to others?
01:03:44.853 --> 01:03:48.063
Fooled by Randomness, the first book by Nassim Taleb.
01:03:48.202 --> 01:03:50.123
I, I feel it's way better than The Black Swan.
01:03:50.242 --> 01:03:50.362
Mm-hmm.
01:03:51.307 --> 01:03:58.193
Um, let me think another good one.
01:03:58.193 --> 01:04:01.043
From the Canon, how to Win Friends and Influence People.
01:04:01.047 --> 01:04:01.967
Oh, Dale Carnegie.
01:04:01.987 --> 01:04:02.208
Joe
01:04:02.208 --> 01:04:02.527
Carnegie.
01:04:02.728 --> 01:04:03.172
Dale Carnegie.
01:04:03.172 --> 01:04:03.472
Yeah.
01:04:03.478 --> 01:04:03.568
Okay.
01:04:03.652 --> 01:04:04.163
Classic.
01:04:04.432 --> 01:04:04.612
Yeah.
01:04:04.733 --> 01:04:08.753
Uh, favorite show, film or piece of content that you, you love that you
01:04:08.753 --> 01:04:09.652
recent or
01:04:09.742 --> 01:04:10.132
anything?
01:04:10.132 --> 01:04:10.313
Yeah.
01:04:10.643 --> 01:04:11.123
Oof.
01:04:11.362 --> 01:04:12.143
Something you like.
01:04:12.643 --> 01:04:18.963
Mm. I don't, I barely watch I've been, so, I, I got this new Mac, MacBook Pro.
01:04:18.967 --> 01:04:19.068
Mm-hmm.
01:04:19.293 --> 01:04:22.413
I used to play, I used to enjoy playing
video games when I was younger.
01:04:22.833 --> 01:04:28.793
And this MacBook Pro runs cyberpunk very well, so I started playing it for the first time.
01:04:28.793 --> 01:04:32.713
It's like an older game already, but, uh, I've been playing it very slowly.
01:04:32.893 --> 01:04:38.552
Uh, sometimes in the evenings while Maggie is, uh, working on Magic Door, I'll play a couple of missions.
01:04:38.552 --> 01:04:39.273
Very hardworking.
01:04:39.393 --> 01:04:39.572
Yeah.
01:04:39.753 --> 01:04:41.822
Uh, what's a place that's special to you in Singapore?
01:04:42.057 --> 01:04:46.168
I guess with Ox Street, we, we had our office in Kallang.
01:04:46.668 --> 01:04:46.697
Oh.
01:04:46.697 --> 01:04:52.927
Our first office was below a mattress store on like a random parking lot.
01:04:53.378 --> 01:04:58.797
And so whenever I pass, I can see, you can see it from the, from the green line?
01:04:58.797 --> 01:04:59.907
From the east, west line.
01:05:00.148 --> 01:05:00.478
Oh, yeah.
01:05:00.507 --> 01:05:04.737
Uh, because it's just after it goes, uh, like pops out of the tunnel.
01:05:04.827 --> 01:05:05.007
Yeah.
01:05:05.507 --> 01:05:09.018
Whenever I pass it, I always am like, oh, yeah, that was, those were the days, man.
01:05:09.588 --> 01:05:10.668
Oh yeah.
01:05:10.967 --> 01:05:15.887
What's the life philosophy or a line that you keep going back to or repeat?
01:05:16.387 --> 01:05:17.077
Very much.
01:05:17.577 --> 01:05:22.088
I think for me, like it's very important to have fun.
01:05:22.088 --> 01:05:22.597
Mm-hmm.
01:05:22.898 --> 01:05:28.358
Um, with, with stuff that I do like I, I. I think when it comes to AI as well, like mm-hmm.
01:05:28.443 --> 01:05:35.733
It, I really try to poke fun at it and, and, and enjoy the stupidity and enjoy the, the cool things about it.
01:05:35.733 --> 01:05:40.432
So, so to me, like, if, if I'm not having fun, I'll, I'll try to change whatever I'm
01:05:40.822 --> 01:05:41.003
Yeah.
01:05:41.057 --> 01:05:42.503
And, uh, last one.
01:05:42.503 --> 01:05:46.438
How can people find you and be helpful to whatever your update?
01:05:46.443 --> 01:05:49.882
Yeah, so, I mean, you can try my, uh, my products.
01:05:49.972 --> 01:05:53.527
You can find me on LinkedIn, um, you can find me on Substack.
01:05:53.677 --> 01:05:53.827
Yeah.
01:05:53.858 --> 01:06:00.032
I actually write on uh, about AI and about, uh, other random stuff as well.
01:06:00.182 --> 01:06:00.393
Hmm.
01:06:00.512 --> 01:06:00.932
Nice.
01:06:00.992 --> 01:06:01.233
Yeah.
01:06:01.532 --> 01:06:03.753
Um, thanks so much Gijs for being here.
01:06:03.992 --> 01:06:04.652
Yeah, it was lovely.
01:06:04.652 --> 01:06:04.952
Yeah.
01:06:05.043 --> 01:06:05.733
Thanks for having me.
01:06:06.063 --> 01:06:06.512
Thank you.
01:06:06.813 --> 01:06:07.293
All right.
01:06:07.443 --> 01:06:10.503
Thank you so much Gijs, and thanks everyone for staying around.
01:06:10.503 --> 01:06:16.472
Can we just, uh, give guys a round of applause and Yeah.
01:06:16.472 --> 01:06:18.722
So, um, we have the room in Until four.
01:06:19.833 --> 01:06:21.242
Uh, yeah, 1600.
01:06:21.362 --> 01:06:22.952
Uh, so feel free to hang around.
01:06:23.063 --> 01:06:25.163
but yeah, we'd love to get to say hi to some of you.
01:06:25.382 --> 01:06:28.023
Thank you so much for tuning into the Code Riff podcast.
01:06:28.262 --> 01:06:30.753
We hope you learned and enjoyed the conversation as much as we did.
01:06:31.143 --> 01:06:38.583
If you'd like to support us, please consider subscribing, leaving a comment like or review on wherever you listen to your podcast, and we hope to see you in the next episode.
01:06:38.853 --> 01:06:39.213
Bye.