# A software factory needs pressure in both directions

Source: https://www.youtube.com/watch?v=y9JtCCZP2MQ

Machine-generated transcript; recognition errors remain. Source-recording timestamps. Boundary-crossing subtitle ends are clipped to this session.

[08:15:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29748s) What's up, everybody?

[08:15:50](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29750s) How are we all doing?

[08:15:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29752s) Give it up for all the fucking organizers out here.

[08:15:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29755s) They are killing it.

[08:15:56](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29756s) Swixen team, let's get a round of applause.

[08:15:58](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29758s) All the speakers, keep it going for the speakers.

[08:16:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29762s) I knew Quinn was really smart, but that was a really dope talk.

[08:16:04](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29764s) I enjoyed watching it from backstage.

[08:16:07](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29767s) Hi, everybody.

[08:16:08](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29768s) I'm Dex.

[08:16:08](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29768s) This is 12 Factor Factories, which started as a joke title, but then everyone started calling it that.

[08:16:15](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29775s) And the war on slop.

[08:16:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29776s) So thank you all for coming.

[08:16:19](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29779s) I like to yap about coding agents on the internet.

[08:16:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29782s) And like a year ago, we did the whole context engineering thing.

[08:16:26](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29786s) A lot of views on YouTube.

[08:16:27](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29787s) And up until about a week ago, I was the CEO and co-founder at a company called Human Layer, the multiplayer coding agent workspace.

[08:16:34](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29794s) space.

[08:16:34](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29794s) And for those of you who are blessed to not be quite as terminally online as some of us

[08:16:40](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29800s) may be, we are now a part of factory, which I'm very excited about and I will be touching

[08:16:44](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29804s) on during this talk.

[08:16:45](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29805s) But I want to get right into it.

[08:16:47](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29807s) Last summer was all about this, like, you know, three months ago, it feels like forever.

[08:16:51](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29811s) Factories were all the hype.

[08:16:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29812s) And I gave an intuition-based talk back in June of, like, we should probably still read

[08:16:56](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29816s) the code and here's how RL and verifiers are working and, like, exactly as our friends

[08:17:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29822s) at Google were saying, you know, models really like just solving the tests, just getting

[08:17:05](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29825s) them to pass. I now have some data beyond just that intuition. There's a benchmark called

[08:17:10](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29830s) Slopcode Bench, which we just released actually today a bunch of new data on from the lab

[08:17:15](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29835s) at University of Wisconsin. But very unsaturated benchmark, you know, GPT 5.5 got a 15%. Astra,

[08:17:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29842s) I think, scored 16.3. So like, we are getting better at this, but there's a lot of ways

[08:17:27](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29847s) to go. I'll talk all about that data in a sec. We, of course, have the study from Pharos

[08:17:31](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29851s) of 30% of PRs merging without review, longer comments, more comments, more incidents per PR,

[08:17:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29857s) more bugs per developer. And I titled this talk, I don't think it was quite correct. I said,

[08:17:42](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29862s) why software factories fail? And everyone said, oh, Dex is the anti-software factory guy. Dex

[08:17:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29866s) hates software factories. I actually love software factories. I'm a huge fan. The title was probably

[08:17:53](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29873s) more along the lines of, there are certain things you could do in your software factory

[08:17:57](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29877s) that would cause it to fail.

[08:17:59](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29879s) And here's what I think those things might be,

[08:18:01](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29881s) and please don't do them.

[08:18:04](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29884s) It'd be the dark factory thing.

[08:18:07](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29887s) But I think the main takeaway here,

[08:18:08](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29888s) and I think we've heard this from a lot of folks,

[08:18:10](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29890s) is software factories is just good engineering.

[08:18:13](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29893s) A friend of the pod, friend of the conference,

[08:18:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29896s) Lauren was talking about this yesterday,

[08:18:18](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29898s) of a lot of this is just stuff we've been doing for a long time.

[08:18:21](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29901s) Actually, Eno from Factory had a really,

[08:18:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29902s) the post is much longer, but it's like,

[08:18:24](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29904s) the factory mindset is about how do we standardize our process

[08:18:27](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29907s) not just the output.

[08:18:29](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29909s) And there's a whole bunch of stuff we'll dig into there

[08:18:31](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29911s) on how we're thinking about that.

[08:18:32](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29912s) But I try to kind of make the same point, right?

[08:18:34](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29914s) It's like your standard software factory

[08:18:36](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29916s) is going to look something like you have some engineers,

[08:18:40](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29920s) some product managers, maybe a CEO

[08:18:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29923s) or someone to drive the vision.

[08:18:45](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29925s) And those people are going to come up with stuff

[08:18:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29928s) that we want to build.

[08:18:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29928s) And they're going to put it in a tracker.

[08:18:50](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29930s) This is, again, this is pre-AI.

[08:18:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29932s) A lot of the things we hear about software factories today

[08:18:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29935s) were around for 30 years, right?

[08:18:58](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29938s) You take stuff off the tracker, you build it,

[08:18:59](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29939s) you run some unit tests, automated tests,

[08:19:01](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29941s) maybe you poke it locally.

[08:19:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29942s) When you're ready, you make a pull request.

[08:19:04](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29944s) At that point, someone might pull it down,

[08:19:07](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29947s) interact with it, test it.

[08:19:09](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29949s) Your CI is going to run a bunch of checks and static linters,

[08:19:12](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29952s) and none of this is that new.

[08:19:14](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29954s) And if anything is wrong with it, you send it back,

[08:19:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29956s) you go fix the problems,

[08:19:18](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29958s) and when it's actually ready to go to prod,

[08:19:21](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29961s) we do some canaries and rollouts and feature flags

[08:19:24](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29964s) and all kinds of fun stuff there.

[08:19:26](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29966s) And eventually it makes contact with our users.

[08:19:28](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29968s) And I love our users.

[08:19:29](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29969s) I love customers.

[08:19:31](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29971s) But they do like to complain about things that are broken

[08:19:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29973s) and ask for stuff that we haven't built yet.

[08:19:36](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29976s) And so that goes back to the team.

[08:19:38](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29978s) And at some point, something breaks at 3 in the morning

[08:19:40](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29980s) and the customers are the first one to figure it out.

[08:19:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29983s) So we add monitoring so that we can wake people up

[08:19:45](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29985s) at 3 in the morning and they can solve it

[08:19:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29986s) before anybody notices.

[08:19:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29988s) This is your pre-AI software factory.

[08:19:50](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29990s) And the basic point is, like,

[08:19:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29992s) all you need to do to make this agentic

[08:19:54](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29994s) is replace someone builds the thing with

[08:19:56](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29996s) an agent builds the thing. So you have orchestration

[08:19:58](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=29998s) I'm not going to go into the details of the sandbox

[08:20:00](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30000s) and the model and all this stuff.

[08:20:01](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30001s) But you can incrementally build this up by adding what I'll call forward pressure.

[08:20:06](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30006s) What causes work to start? Can we take all of the feature requests and just route them straight to our factory?

[08:20:11](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30011s) Can we take all of the incidents that happen so that when I get paged at 3 in the morning,

[08:20:14](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30014s) I'm looking at a pull request that we think is going to solve it, not just a stack trace about what went wrong.

[08:20:21](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30021s) So the forward pressure, what causes new work to start?

[08:20:25](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30025s) and back pressure, which is sort of like,

[08:20:27](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30027s) okay, cool, now we can build things in minutes or hours,

[08:20:29](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30029s) but the review and testing still takes hours or days

[08:20:32](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30032s) if we're doing it in the old way,

[08:20:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30033s) so we bring in agentic code review, agentic testing.

[08:20:36](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30036s) Anyone doing agents code reviewing their pull requests?

[08:20:40](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30040s) Okay.

[08:20:41](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30041s) I thought it would be a little bit more,

[08:20:42](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30042s) but that's pretty good.

[08:20:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30043s) Y'all are catching up.

[08:20:45](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30045s) And then we invest heavily into our CICD and our linters

[08:20:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30048s) and our, anyone doing like lean checks on their code,

[08:20:50](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30050s) like formal verification?

[08:20:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30052s) All right.

[08:20:53](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30053s) Hardcore.

[08:20:53](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30053s) I like it.

[08:20:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30055s) So the back pressure, how do we let agents hill climb to a better outcome without needing

[08:20:59](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30059s) to involve humans?

[08:21:00](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30060s) How do we just overall increase the quality of what is coming out of our build step?

[08:21:04](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30064s) And agents help here, but a lot of what makes software factories good is stuff we've kind

[08:21:08](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30068s) of had for decades.

[08:21:10](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30070s) And just like context engineering, which was the thing that I was talking about last summer

[08:21:14](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30074s) in like June of 2025, there was all these ideas of like RAG and memory and state and

[08:21:18](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30078s) history and all this stuff, and we tried to like distill out what are the timeless things

[08:21:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30082s) that go into building agents.

[08:21:24](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30084s) And so I want to distill out now some of what I think are the timeless things that go into

[08:21:28](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30088s) building and designing a software factory.

[08:21:31](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30091s) Because there's a lot of noise and hype.

[08:21:32](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30092s) Like I said, Terminally Online, Twitter is going to break your brain if you spend too

[08:21:36](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30096s) much time wondering what everyone else is doing.

[08:21:38](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30098s) I'm also going to talk about Slop because I have a very fun project that I've been working

[08:21:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30103s) on with Swix and a handful of other speakers here, but we'll get to that later.

[08:21:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30108s) We're going to do 12 great practices for software factories.

[08:21:51](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30111s) And I think the first time I started talking about this, it was only like 9, but we are at 12 now, so I'm very excited about this.

[08:21:56](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30116s) I touched on this, but loops was this concept that was very in vogue.

[08:22:00](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30120s) I think this is much more useful if you think about the interplay of forward pressure and back pressure.

[08:22:06](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30126s) So, you know, again, very hyped up, but the actual tactical thing is like, how do you start work automatically?

[08:22:11](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30131s) And how do you make sure that whatever comes out and is touched by a human by the end, all the low-hanging stuff is already fixed.

[08:22:17](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30137s) The outcome is as good as it can be autonomously.

[08:22:21](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30141s) Number two, I will call know your enemy.

[08:22:23](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30143s) I will still posit that unattended models

[08:22:26](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30146s) don't do a great job of maintaining

[08:22:28](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30148s) and improving code base quality over time.

[08:22:30](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30150s) So if you care about having a system

[08:22:31](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30151s) that's going to be still maintainable

[08:22:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30153s) in three or six months,

[08:22:35](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30155s) I mean, again, I gave this talk

[08:22:36](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30156s) and I said, please read the code

[08:22:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30157s) because a lot of it is slop.

[08:22:39](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30159s) Agents got really good at solving one-off problems

[08:22:41](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30161s) and making video games

[08:22:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30163s) and Blender slop and all this stuff.

[08:22:45](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30165s) I don't think, you know,

[08:22:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30166s) if you steer a model to fix your code base,

[08:22:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30168s) it can go fix your code base.

[08:22:49](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30169s) But then you have to know what is there and what you want.

[08:22:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30172s) So if you are just hands-off pushing features into the machine,

[08:22:56](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30176s) you may not find that it autonomously will make things better over time.

[08:23:00](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30180s) Back then I didn't have a benchmark.

[08:23:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30182s) Now I have a benchmark that basically has the model solve problems and checkpoints.

[08:23:06](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30186s) Like a model can clone all of Excel or Redis in a couple days,

[08:23:10](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30190s) maybe even a couple hours now with like agent swarms.

[08:23:13](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30193s) But that requires knowing the entire problem in advance.

[08:23:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30196s) And most of us don't write software that way.

[08:23:18](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30198s) Most of us ship a feature, and then ship another feature,

[08:23:21](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30201s) and then ship another feature.

[08:23:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30202s) And if you're building by hand without AI,

[08:23:24](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30204s) along the way you say, oh, that's a bad abstraction.

[08:23:25](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30205s) We need to centralize this, or we need to remove this.

[08:23:27](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30207s) And so this is a benchmark that measures a model's ability

[08:23:30](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30210s) to evolve a code base over time.

[08:23:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30213s) And so Astra, again, is 16%, very unsaturated.

[08:23:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30217s) There's a bunch of data here.

[08:23:38](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30218s) I probably don't have time to look at all of it,

[08:23:40](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30220s) but we can look at code quality metrics

[08:23:42](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30222s) like cyclomatic complexity, lines per symbol,

[08:23:47](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30227s) the percentage of functions that have high complexity.

[08:23:49](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30229s) There's more. There's 200 custom rules.

[08:23:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30232s) 50 of them are based on things we've been doing for a while.

[08:23:54](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30234s) A lot of them are custom based on looking at code that models write,

[08:23:57](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30237s) comparing it to pre-AI repos.

[08:23:59](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30239s) I wrote a long article about that. I've read a bunch.

[08:24:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30242s) Gabe Orlansky, who runs the lab that does this benchmark,

[08:24:04](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30244s) just put up this paper today

[08:24:06](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30246s) based on some of the research we did with them.

[08:24:09](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30249s) There's a bunch of stuff in here.

[08:24:10](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30250s) I probably don't have time to go through all of it,

[08:24:12](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30252s) but they also test the model's ability to design good libraries and interfaces.

[08:24:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30256s) The most, the craziest one here is like, models did learn to use exceptions instead of returning

[08:24:21](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30261s) tuples in Python.

[08:24:23](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30263s) There was a lot of patterns that were identified as like probably not great, that were way

[08:24:28](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30268s) less present in the, you know, six months ago models than they are in today's models,

[08:24:32](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30272s) so it almost feels like we're like, the slop is amplifying over time.

[08:24:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30277s) And Claude and GLM are now writing Python 3 in Python 2 style, so we have, you know,

[08:24:42](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30282s) %s format strings are back, and they will not use fstrings by default anymore.

[08:24:47](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30287s) So things to be aware of. You may not care about this. I care about this.

[08:24:50](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30290s) The jury's still out on Fable 5.2. I love reading the Anthropic blog post about how

[08:24:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30295s) to prompt the new models. And so allegedly, if you tell Fable it owns the code base, it

[08:24:59](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30299s) should keep it clean. I haven't tested this yet. We're running the benchmark in the background

[08:25:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30302s) right now. But for those of us not building FBS games, shouts out to Ben, built a cool

[08:25:06](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30306s) game. You should go watch this video. It's insane.

[08:25:09](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30309s) you should know your enemy.

[08:25:11](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30311s) You should understand that, like,

[08:25:12](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30312s) work with this in mind, that if you need your codebase

[08:25:15](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30315s) to last a long time, or your codebase has been

[08:25:17](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30317s) around for a long time, we should

[08:25:19](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30319s) probably maintain understanding of the

[08:25:21](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30321s) system and the code.

[08:25:23](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30323s) To do this, we like to plan before we

[08:25:25](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30325s) build. I highly encourage this. You should do

[08:25:27](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30327s) this. We used to do this before in this old software

[08:25:29](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30329s) factory. People realize, especially

[08:25:31](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30331s) when building things took

[08:25:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30333s) hours or days,

[08:25:35](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30335s) and when testing and reviewing things

[08:25:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30337s) took hours or days, even if this part is faster,

[08:25:39](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30339s) this part is probably still slow if you're doing it properly,

[08:25:42](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30342s) we would have architecture review,

[08:25:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30343s) and we would have sprint planning,

[08:25:44](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30344s) and we would talk as a team, and we would align

[08:25:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30346s) in the hopes that we might decrease the percent chance

[08:25:50](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30350s) that we'd have to rework something later.

[08:25:51](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30351s) We wanted to be rather sure that we're building

[08:25:54](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30354s) the right thing in the right way

[08:25:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30355s) before we actually go and spend time building it,

[08:25:58](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30358s) time in tokens, and then get to review

[08:26:00](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30360s) and have to rework it.

[08:26:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30362s) A lot of people overdo this as well.

[08:26:04](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30364s) I've overdone this for sure,

[08:26:05](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30365s) where you spend six hours working on a spec

[08:26:07](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30367s) and you're like, I feel like I could have just YOLO'd this.

[08:26:10](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30370s) Because sometimes you can just YOLO a two-sentence prompt,

[08:26:12](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30372s) and if it's small or you don't care,

[08:26:13](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30373s) there's maybe a 50% chance you send a couple more prompts

[08:26:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30376s) and then it's good.

[08:26:17](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30377s) You could spend five hours handwriting a detailed spec,

[08:26:20](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30380s) and maybe there's a 10% of it you have to change afterwards.

[08:26:23](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30383s) Or you could write all the code by hand in 20 hours,

[08:26:25](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30385s) and then there's like a 0% chance.

[08:26:27](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30387s) Well, there's 0% chance you have to rework it

[08:26:29](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30389s) because of something the agent did.

[08:26:30](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30390s) You know, people still make mistakes too.

[08:26:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30393s) But I think about this in terms of expected pain.

[08:26:35](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30395s) like how likely will this have to change later

[08:26:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30397s) times how hard will it be to change.

[08:26:41](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30401s) I don't argue with the model during planning

[08:26:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30403s) about what color the button's gonna be,

[08:26:44](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30404s) but I do wanna make sure we get our database schemas right

[08:26:47](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30407s) and stuff like that,

[08:26:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30408s) because it's really hard to change later.

[08:26:49](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30409s) So based on your problem and what you're doing

[08:26:51](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30411s) and your code base,

[08:26:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30412s) you should probably decide where on this spectrum

[08:26:54](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30414s) you wanna land.

[08:26:56](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30416s) Align visually, I published a skill called show me,

[08:26:59](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30419s) which is optimized for instead of reading walls of text,

[08:27:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30422s) you get HTML mockups, you get tables,

[08:27:04](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30424s) You get charts. You get all this stuff that is optimized for the human brain, which is very visual.

[08:27:10](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30430s) I think Koda said we've been running away from predators and eating fruit for millions of years.

[08:27:15](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30435s) We've only been reading text for about 5,000.

[08:27:18](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30438s) Anyways, you can get this. It's got about 40,000 installs.

[08:27:21](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30441s) You should help agents test their work.

[08:27:23](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30443s) Screenshots, browsers, videos, curl, whatever you can do to help the agent.

[08:27:26](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30446s) Again, he'll climb against an outcome.

[08:27:28](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30448s) If you're building an iOS app, make sure it can build, run, and test that app in whatever environment it's running.

[08:27:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30453s) You should fight slop with anyone here writing in TypeScript.

[08:27:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30457s) OK.

[08:27:38](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30458s) Some people are feeling lazy.

[08:27:39](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30459s) I feel like it should be more than that.

[08:27:41](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30461s) But this is a repo from Dylan Mulroy, who's a principal engineer at CloudFlare.

[08:27:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30466s) Anti-slop.

[08:27:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30466s) If you're not using this, you should go check it out.

[08:27:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30468s) It's 100-ish rules on stopping agents from building slop.

[08:27:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30472s) Factory also has an ESLint plugin that I should probably mention.

[08:27:56](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30476s) And also, again, you can go slop fork all of the Python code quality rules out of Slop

[08:28:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30482s) code bench. This exists for your

[08:28:04](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30484s) language. You should use it. Slop

[08:28:06](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30486s) is a tool. Not all slop is bad.

[08:28:07](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30487s) Vibe coding prototypes, side projects.

[08:28:10](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30490s) If you have something behind a feature flag, you just need to like figure

[08:28:11](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30491s) out if people even want it before you

[08:28:14](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30494s) go put in the time. I think it's like, it's great.

[08:28:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30496s) I don't think there's like always, always, always,

[08:28:18](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30498s) always every line of code has to be perfect.

[08:28:20](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30500s) You should give models a feature

[08:28:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30502s) flagging system. PostHog

[08:28:24](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30504s) or Statsig or LaunchDarkly or whatever it is so they

[08:28:25](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30505s) can try stuff and see if it performs better

[08:28:27](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30507s) and then make decisions about

[08:28:29](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30509s) what the right thing to be in your tool.

[08:28:32](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30512s) You should compound over time, eval everything,

[08:28:34](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30514s) learn from past sessions,

[08:28:35](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30515s) and you should unify your forge.

[08:28:38](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30518s) What I mean by this is we used to send

[08:28:40](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30520s) Git patches around on mailing lists.

[08:28:41](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30521s) I think Linus and the Linux people still do this,

[08:28:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30523s) but most people keep all their code in GitHub.

[08:28:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30526s) And so before the ANA to forge,

[08:28:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30528s) we have centralized patches,

[08:28:49](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30529s) but we still have people pasting

[08:28:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30532s) clawed answers back and forth in Slack

[08:28:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30535s) and hopefully not doing too many rounds of that

[08:28:58](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30538s) because it's like, I copy out of mine

[08:29:00](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30540s) and send it to my coworker.

[08:29:01](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30541s) He pasted it into his agent, and then they sent it back to me,

[08:29:03](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30543s) and then I pasted it into my agent, and maybe we get an answer.

[08:29:06](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30546s) So centralizing all of your designs and decisions

[08:29:08](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30548s) and whatever you have in Notion and Google Docs

[08:29:11](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30551s) and whatever it is, you should try to bring those things

[08:29:14](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30554s) as close as possible together so you can understand

[08:29:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30556s) not just the code about what something was built

[08:29:18](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30558s) but all the decisions that went into it.

[08:29:20](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30560s) You can't buy everything.

[08:29:21](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30561s) I think there's things in this space that just cannot be productized.

[08:29:25](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30565s) Obviously, you can buy VMs or you can bring your own.

[08:29:28](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30568s) Dev environment, I think, is very personal to most teams.

[08:29:31](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30571s) The harness, of course, you can go get a Frontier harness,

[08:29:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30573s) or you can use something minimal like OpenCode or Pi

[08:29:35](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30575s) and customize the heck out of it.

[08:29:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30577s) And then you have this control plane picture of, like,

[08:29:39](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30579s) how do you collaborate on sessions and plans?

[08:29:41](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30581s) How do you do automations and scheduling?

[08:29:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30583s) How do you respond to events?

[08:29:44](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30584s) All this governance stuff.

[08:29:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30586s) Part of the reason I'm really excited to be working with Factory on all this stuff

[08:29:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30588s) is, like, they have a much bigger view of this picture

[08:29:51](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30591s) and have a lot more stuff solved here.

[08:29:53](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30593s) But there are parts, especially at the top here,

[08:29:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30595s) is, like, I don't think anyone can sell you verification of your code

[08:29:59](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30599s) because only you

[08:30:00](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30600s) know, if you have hundreds of engineers and hundreds of repos, you have those opinions and you know how to test your code.

[08:30:05](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30605s) Not just a black box you can buy.

[08:30:07](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30607s) There are building blocks that are undifferentiated.

[08:30:09](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30609s) But the idea of software factory, I think, is a way of thinking about our work.

[08:30:14](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30614s) So that's software factories.

[08:30:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30616s) I also want to talk about slop.

[08:30:20](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30620s) Factories are about how do we increase human leverage, right?

[08:30:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30622s) How do we get the best return on our time?

[08:30:24](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30624s) and I've talked a little bit about why humans should be involved

[08:30:26](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30626s) and what it means to be a diligent

[08:30:28](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30628s) steward of what we're building and what we're doing

[08:30:31](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30631s) but I think it's bigger than just software

[08:30:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30633s) so I'm very excited to announce

[08:30:35](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30635s) the Anti-Slop Manifesto

[08:30:38](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30638s) which you can go

[08:30:39](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30639s) if you go to slopmanifesto.com

[08:30:40](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30640s) you can read it, I'll put the URL up later

[08:30:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30643s) but we built this in collaboration

[08:30:44](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30644s) with a lot of people here, both attendees

[08:30:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30646s) and speakers, special shout out to

[08:30:49](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30649s) Swix and Phillip

[08:30:50](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30650s) it has signers from Zoom Info

[08:30:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30652s) and I think XAI and OpenAI,

[08:30:56](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30656s) and there's a lot of cool people on there.

[08:30:58](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30658s) I encourage you to go read it and check it out

[08:31:00](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30660s) and be a part of, you know.

[08:31:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30662s) But I'll tell you what's in it.

[08:31:03](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30663s) We value outcomes over outputs.

[08:31:06](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30666s) We value understanding over speed.

[08:31:09](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30669s) We value our own thinking

[08:31:11](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30671s) over whatever the model came up with as a first draft.

[08:31:14](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30674s) We value better work over easier work.

[08:31:17](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30677s) We value improving the system

[08:31:19](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30679s) over just patching outputs.

[08:31:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30682s) And the things we believe are that AI is the most useful tool

[08:31:25](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30685s) most of us have ever worked with,

[08:31:27](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30687s) but it can create a swarm of agents can produce more work in a single day

[08:31:32](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30692s) than a human can review in their entire career.

[08:31:34](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30694s) Look at some of the math stuff coming out of OpenAI.

[08:31:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30697s) Slop is when we put less effort into producing something

[08:31:39](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30699s) than we ask their consumer on the other end to spend understanding it.

[08:31:44](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30704s) And using AI doesn't necessarily make something slop,

[08:31:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30708s) but it happens when we pass off AI output

[08:31:50](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30710s) without taking responsibility for its quality or for the reader's experience.

[08:31:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30715s) I believe strongly, I've been saying this for a year now,

[08:31:58](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30718s) AI cannot replace human judgment.

[08:31:59](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30719s) It can only amplify our judgment,

[08:32:01](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30721s) or maybe perhaps our lack of judgment in certain cases.

[08:32:05](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30725s) And the slop is asymmetric.

[08:32:07](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30727s) The cost of producing a thousand-line PR or a ten-page doc is approaching zero,

[08:32:12](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30732s) but the cost of reviewing it is still fairly constant.

[08:32:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30736s) And slop proliferates when we repeatedly work around problems

[08:32:20](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30740s) instead of solving them correctly.

[08:32:21](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30741s) It comes from tools that love to add

[08:32:23](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30743s) and are not biased towards subtracting.

[08:32:26](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30746s) I'm not going to name names.

[08:32:29](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30749s) But I think a lot of slop comes from pressure.

[08:32:31](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30751s) Companies are pushing their people

[08:32:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30753s) to redefine what is humanly possible.

[08:32:36](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30756s) And I think slop compounds.

[08:32:39](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30759s) So the more slop sits in our knowledge base,

[08:32:41](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30761s) in chat, in code,

[08:32:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30763s) the more of it the next agent reads and copies.

[08:32:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30766s) And the more slop flying around our team,

[08:32:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30768s) the more we lose the will to fight it, and everything gets worse over time.

[08:32:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30772s) Communication and quality and all of this.

[08:32:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30775s) So we, the undersigned of the anti-slop manifesto...

[08:32:58](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30778s) Guys, is it a little hot in here?

[08:33:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30782s) I might just have to pop off the jack. There we go.

[08:33:06](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30786s) They couldn't give me one of these with a factory logo on it yet,

[08:33:08](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30788s) but I'm sure it's coming soon.

[08:33:11](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30791s) Thus, we, the anti-slop undersigners,

[08:33:14](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30794s) resolve that human time and attention is incredibly valuable.

[08:33:19](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30799s) So we optimize for using it wisely.

[08:33:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30802s) We fight slop because AI should make our work better,

[08:33:25](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30805s) not just easier.

[08:33:27](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30807s) We optimize our systems and tools and products

[08:33:30](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30810s) for the efficient elicitation of human intent

[08:33:32](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30812s) and for efficient human understanding.

[08:33:35](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30815s) And we do our own thinking when working with a model.

[08:33:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30817s) We explore resources, code, prior art.

[08:33:40](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30820s) We really want to understand the problem

[08:33:41](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30821s) before we actually get to work.

[08:33:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30823s) And we own the intent and the taste and the consequences of what we ship.

[08:33:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30826s) However much AI helps us to make it, we own it.

[08:33:50](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30830s) And when an agent wrote something that we're sending somebody,

[08:33:52](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30832s) we make that clear, and we actually make sure it's good

[08:33:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30835s) before we drop it in Slack or email or whatever it is.

[08:33:58](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30838s) And when someone sends us slop, this is also like we have to push back.

[08:34:01](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30841s) We have to tell people kindly, empathetically,

[08:34:03](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30843s) and then ask them to review it first.

[08:34:05](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30845s) And most of us, it's a journey.

[08:34:07](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30847s) I've been on both sides of this one, so I've had to have that note.

[08:34:11](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30851s) I believe there will always be alpha in reviewing something.

[08:34:16](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30856s) As models get better, we may stop reading most of the code.

[08:34:19](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30859s) We may still review a plan or a design or a result.

[08:34:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30862s) But if you don't review anything, if you bring nothing to the table,

[08:34:25](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30865s) no thinking, then you're just going to get the same AI output as everybody else.

[08:34:30](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30870s) Agent reviewers, linters, verifiers, they're very useful,

[08:34:33](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30873s) but they are not sufficient.

[08:34:36](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30876s) So for now, we'll still review some subset of the outputs.

[08:34:40](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30880s) What subset, of course, depends on the stakes.

[08:34:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30883s) But what shift, who's using it, and how it's helping them matters far more than lines of

[08:34:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30888s) code or PRs merge or, God forbid, the number of agents we're running in parallel.

[08:34:53](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30893s) It's cool if it gets you to the outcome.

[08:34:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30895s) But unreviewed output isn't always necessarily bad.

[08:34:57](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30897s) Slop is a cool one-shot prompts, vibe coding.

[08:35:02](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30902s) These are all high-leverage things that are in the right place.

[08:35:04](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30904s) Slop is bad when it undermines our outcome.

[08:35:06](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30906s) If we want to make a good product, a maintainable system, we want to save people time and energy.

[08:35:11](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30911s) If slop undermines that, then it's bad.

[08:35:14](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30914s) And I guess finally, none of us has this figured out.

[08:35:17](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30917s) We're each relearning what we will and won't accept, and there's no single right answer here.

[08:35:22](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30922s) You've got to calibrate yourself.

[08:35:23](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30923s) The world moves fast.

[08:35:25](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30925s) My goal here is these are principles that we can use to embrace our important human role

[08:35:30](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30930s) in harnessing AI safely, responsibly, in a way that creates value for everyone.

[08:35:37](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30937s) for many years to come.

[08:35:39](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30939s) So you can read and sign the manifesto

[08:35:40](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30940s) at slotmanifesto.com.

[08:35:43](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30943s) TLDR, you know, give a shit.

[08:35:46](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30946s) Care about what you're doing.

[08:35:48](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30948s) I'm going to stick around for Tariq's talk,

[08:35:50](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30950s) but then, what is that slide?

[08:35:53](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30953s) All right, I'll be hanging in the hallway afterwards.

[08:35:55](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30955s) So come say hi.

[08:35:56](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30956s) Keep learning, keep giving a shit.

[08:35:58](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30958s) I'm excited to keep building with y'all.

[08:36:00](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30960s) Thank you so much.

[08:36:01](https://www.youtube.com/watch?v=y9JtCCZP2MQ&t=30961s) Thank you.