AI Engineer Code 2025
No More Slop – swyx
Read the talk
No More Slop: Growing AI Engineering Without Losing Taste
Cheap generation makes output abundant and judgment scarce. AI engineering needs curation, codebase understanding and accountability to keep that abundance useful.
From a talk by Shawn "swyx" Wang
How do you grow without losing taste?
How do you grow a community, an industry and a conference while preserving the quality that made people care about them? That is the practical problem behind Shawn “swyx” Wang’s declaration of war on slop at the AI Engineer Code Summit. Each expansion creates more possibilities—and more work deciding what deserves attention.
The first AI Engineer Summit backed the emergence of the AI engineer as a role. The second extended that focus to leadership; the third concentrated on model labs. The World’s Fair widened the audience to AI product managers and designers. Code Summit then concentrated on a theme: coding. Curation becomes the means of preserving taste through that growth, expressed in who attends, who speaks and what the event chooses to cover.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Judge the output, whoever produced it
Slop made Oxford’s 2024 Word of the Year shortlist, but brain rot won. The definition on swyx’s slide associates slop with artificial intelligence and describes it as low quality, inauthentic or inaccurate. He accepts those quality criteria and challenges the restriction to AI: humans can produce all three failures too. Slop is a property of the output, not a reliable label for its producer.
The contrasting internet term is kino: work that feels creative, accomplished or worth experiencing. Either a human or an AI can contribute to either outcome. That makes the useful question less about whether AI was involved and more about the judgment exercised in making and selecting the result.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The same tools can produce very different work
Crediting Paul Rambles for the idea, swyx starts with Sora. He describes his own videos featuring himself as Sam Altman as boring; more creative users make cats playing didgeridoos. Access to the generator does not supply the idea. He then makes the same point through Netflix’s K-Pop Demon Hunters and The Electric State: a shared studio does not imply a shared level of quality.
A brief visual comparison extends the joke to different models, without a spoken explanation of their merits. The next comparison adds time: something fresh can become slop as a trend repeats and loses its novelty. A further dig at Game of Thrones makes deterioration within a single body of work part of the same pattern. Tools, studios and an earlier success cannot guarantee that the next output will deserve attention.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A promising category is not enough
The distinction also applies to products built around the same startup idea. For his first AI Engineer Summit keynote, swyx used Tome, an AI presentation company. He reports returning to the original deck and finding it gone after the company closed. The example brings the discussion from the appeal of generated output to the experience of relying on the product that holds it. He follows it with contrasting takes on vibe coding: a common category still leaves room for very different execution.
Even two exponential charts can deserve different reactions. The slide places a METR chart under “Evidence” and an AI 2027 chart under “Fan Fiction.” These are swyx’s editorial judgments about the presentations, not conclusions established by the curves’ shapes. His invitation is to investigate why apparently similar stories of accelerating progress feel different. An exponential shape does not tell you what supports the claim.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Generation gets cheaper; judgment still costs
Brandolini’s law provides the analogy: refuting bullshit requires an order of magnitude more energy than producing it. Generative AI intensifies the imbalance by making production cheaper. Swyx claims generated-token costs are falling by 100–1,000× each year; he does not specify models, workloads or a measurement period for that annual rate. The relevant pressure is the growing gap between the ease of producing material and the effort required to evaluate it.
His proposed law of anti-slop substitutes taste for energy: fighting slop requires an order of magnitude more taste than producing it. This is a rhetorical rule about the burden of judgment, rather than a measured ratio. If low-effort output is abundant, people who care about quality have to do more than generate competing output. They must elevate what gets made, selected and shared.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Use AI to improve the selection, then demand good code
AI can help with that work. Swyx describes his side project AI News as a newsletter that tells readers not to read it when nothing is happening. Its useful editorial decision is sometimes to save the reader’s attention. He then points to a skill prompt for Claude that explicitly acknowledges slop and instructs the model to avoid it. Swyx reports a significant improvement in the displayed left-to-right comparison, without giving evaluation criteria or a numerical score. The practical move is to make quality expectations part of the instructions.
Code raises the stakes. Swyx illustrates the amplification of technical debt with two engineers producing the debt of 50, then cites private-data exposure affecting millions of users. He places these examples in the year of the talk, but does not identify the incidents or provide measurement details. The concern extends beyond unattractive output: cheap code can create expensive maintenance and security consequences.
He then challenges unnamed claims of models working autonomously for 30–60 hours because they omit whether the resulting code was good. Runtime measures how long a process continued; it does not answer whether its output was useful. His demand, modeled on the language of taxation and representation, is no autonomy without accountability.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Match attention to the work
The semi-async valley of death names a problem in how people collaborate with coding agents. Swyx’s distinction is between difficult work that benefits from sustained human attention and commoditized work that can move further into the background.
| Work | Collaboration mode |
|---|---|
| Hardest problems | Keep human attention closely coupled to the machine |
| Commoditized tasks | Allow more asynchronous execution |
The choice is about where attention contributes to solving the problem. Making every task more asynchronous would discard the close collaboration he wants to preserve for the hardest work.
Producing code also creates a need to understand it. Swyx presents Codemaps as a way to use AI to scale codebase understanding, and directs attendees to Cognition for a fuller demonstration. That supplies another target for AI assistance: help people comprehend the growing system they are responsible for, alongside helping them change it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Operate the application, manage the context
Computer use extends assistance into the applications where work happens. Looking back to Anthropic’s computer-use launch roughly a year earlier, swyx describes agents becoming capable of operating complex applications, including an IDE. He reports using Devin to automate updates to the conference website. The accompanying slide frames computer use as AI testing and shows a Cognition post beside an annotated conversation. This connects autonomous action to the broader task of checking and maintaining what people will actually use.
The next control is context management. Swyx identifies using subagents to fight context rot as a lesson from the previous conference day: delegation can help keep an agent’s working context focused instead of accumulating everything in one conversation. He closes this sequence with a principle he attributes to Greg Brockman: humans design clear module boundaries, and AI writes the code between them. Together, these ideas preserve deliberate structure around an expanding amount of generated work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Refuse the incentives that produce slop
The closing call-and-response turns these ideas into a response to pressure: “No more slop.” First comes a boss demanding more lines of code by the end of the quarter. The demand rewards volume without establishing value. Next comes pressure to ship an insufficiently tested release that could embarrass the company. The asymmetric burden returns here: creating and releasing the work can be much easier than dealing with its consequences.
Finally, the pressure comes from Twitter’s engagement incentives: publish bait, even when attracting attention pushes toward misleading the public. The final slide pairs an engagement-bait example with a corrective community note. The refrain remains the same. Taste has to survive the moment when a manager, a release deadline or an algorithm rewards abandoning it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Oxford explains its selection of brain rot as the 2024 Word of the Year.
Introduces structured codebase maps with navigation to source lines and contextual explanations.
The historical announcement of Claude's computer-use capability alongside updated Claude 3.5 models.
Further reading
- Stanford AI Index 2025Article
Research highlights covering AI capabilities, adoption, and falling inference costs at a fixed performance level.
Explains how human task duration and agent success rates can measure autonomous software-task capability.
- SWE-grep and Fast ContextArticle
A swyx-coauthored explanation of specialized retrieval subagents and the latency tradeoffs of codebase search.
Anthropic's guidance on prompting Claude toward distinctive frontend design and away from generic visual patterns.
Read the complete timestamped transcript
- 0:11
[upbeat electronic music] Morning. How's everyone doing?
- 0:23
Good.
- 0:24
Good? I'm gonna need a lot of energy for this talk, so please back me up. I'm very nervous. [laughing] Uh, but we'll get through this. I'm declaring war on slop today.
- 0:33
Let's talk about this. Uh, every AIE has a secret. I, I've told this to, uh, some folks that are personal friends, and I'll just show you, show the secret now.
- 0:43
The first summit we had the secret, which was we knew that the AI engineer was gonna be a thing. Second summit, we extended it to leadership. Third summit, we realized that basically we always needed to concentrate on model labs, and that's why you see, um, all, all the, all the top tier labs here today.
- 0:59
Um, World's Fair, we started expanding the TAM of, of what AI engineering, uh, is, is, is affiliated with, with AI PMs and AI designers. And with the Code Summit, as Jed just talked about, we really started to focus on curation and focusing in on a theme.
- 1:14
And if there's one theme that really matters this year is, it's coding. But I'm not here to talk about coding. The rest of the day you're gonna hear about coding, so just, just indulge me five minutes to talk about slop.
- 1:25
Um, we, we've done really well, right? Like, so, so, like, and I think, like, slop is sort of associated with quantity and quality, and I think, like, that's something that I'm really trying to think about as well.
- 1:34
Like, how do we grow this community, grow this industry, and grow this event, uh, with the, the same kind of taste and high quality that you've come to expect?
- 1:43
And this is something that, you know, I hope, ho- hopefully you guys can see that we care a lot about, uh, in curating all of you coming here and all of the speakers that you're about to see.
- 1:52
Um, we're in a war against slop. Um, this is actually, uh, the Oxford English Dictionary. Uh, this is a g- a candidate for the 2024 Word of the Year.
- 2:01
It lost to brain rot. [laughs] [laughing] But, um, but slop is, slop is pretty good, and I think it's, it's probably gone even more of an issue this year than last year.
- 2:11
Maybe it will win this year. I have an issue with Oxford, though, because they
- 2:17
did us dirty by saying slop is co- is generated using artificial intelligence. The other part I, I agree with. It's low quality, inauthentic, or in- inaccurate. But it doesn't take AI to be low quality, inauthentic, or inaccurate.
- 2:31
Any human or AI can be an agent of slop, right? You've seen this yourself. I'm gonna indulge me with a few examples. By the way, I got this, uh, if you're not familiar with like sort of internet slang, the opposite of slop is kino.
- 2:42
Um, and I got this idea from Paul Rambles, who, uh, ha- ha- like, you know, when I do Sora videos, I do really boring Sora videos with me as Sam Altman.
- 2:52
When other people who are actually creative and good at their job do Sora videos, they do cats playing didgeridoos. Uh, slop can be produced by the same studio. There's K-Pop Demon Hunters by Netflix, and there's Electric State by Netflix. [laughing]
- 3:08
Slop can be produced in terms of different models. No comment. [laughing]
- 3:15
Slop can be-- Slop can de-- Uh, something that's kino can degenerate into slop, right? If you're s- early on a trend and you're, and y- you're, you're starting that and it's, and it's fresh and new, that's great.
- 3:25
Um, if you're re- if you recognize the other image, you're too online. [laughs]
- 3:31
Okay, not enough people recognize that image. [laughing] Um,
- 3:35
s- go do your homework. Um, yeah, and obviously I'm just gonna throw in a, a dig at Game of Thrones because this is the same, same thing, right? Like, slop is everywhere.
- 3:42
It's generated by humans and AI. You get it. [coughs]
- 3:45
Okay, um, the same startup idea can be kino versus slop. When I first presented the, my first keynote at AI Engineer Summit, we actually used Tome, uh, which is sort of like an AI slides company.
- 3:55
Uh, I loaded it-- I loaded up the same slide deck, and it was gone, uh, recently because, uh, it was, it was actually, uh, sort of, uh, closed as a company.
- 4:04
Um, there's, there's, you know, different takes on vibe coding, and I think one of them, uh, is much better than the other. And I think like this is, these are just like the tensions that we have to navigate.
- 4:12
Um, one of our speakers later on, Exo, uh, Meter, um, I think it's really interesting that both of them are exponential charts, but one of them feels more kino and the other is more slops.
- 4:21
And I think, like, I would really like to have people investigate why. Um, so let, let me, let me just skip through. Basically, we're in, in an asymmetric war on slop.
- 4:31
Um, I think, uh, the closest law that I found that matches this is Brandolini's law, which actually states the amount of energy needed to refute bullshit is an order of magnitude bigger than needed to produce it, right? [laughing]
- 4:43
So we need to co- coin an, uh, appropriate law as well, um, because, uh, you know, like the, the, the cost of generated tokens is, is dropping by a hundred to a thousand times every single year.
- 4:54
Um, so this is, I guess, swyx's law of anti-slop. The amount of taste needed to fight slop is an order of magnitude bigger than that needed to produce it, right?
- 5:02
There's so much low taste out there. We need to elevate, uh, what's out there in the world because that's what we stand for as humans. Um, I think there's, there's a positive message.
- 5:12
You can use AI to fight slop. Um, uh, I, I'm proud, I'm proud to, um, run as a side project AI News, which is the only newsletter that tells you not to read it when there's nothing going on. [laughs]
- 5:22
Thank you. [laughs] Oh. [clapping] Appreciate that. Um, you can also prompt to fight slop. The next speaker is, um, B- Mahasin Berry. Um, I found this in the, in the sort of prompts, uh, in the, in the skill set that they, that they put, that they put where they actually a- acknowledge slop and tell Claude not to produce slop,
- 5:39
and it, it actually improves significantly from left to right. Um, what about code slop, right? We're here about code. Um, creating tech debt where you can sort of-- two engineers can create a tech debt of 50 engineers or, you know, on a more serious note, you can start exposing private data of millions of users, and this, this
- 5:56
all happened this year. Everything I'm mentioning all happened this year. I'm kind of using this keynote as a f- as a way of recapping. Um, and it, and, you know, just to be spicy a little bit, even people who are saying things like, "Oh, my model can go up to thirty to sixty hours, uh, a- autonomously," well,
- 6:11
it feels a bit sloppy because you're also not saying, well, was the code good or not? You're just saying how long it went. So in the same way that you have no taxation without representation, you don't want autonomy without a-accountability.
- 6:25
Um, something I've been working on more recently is that using AI to fight code slop as well. So, uh, this is from the... A, a, bunch of people quoted this yesterday, of the semi-async value of dev, where you can sort of keep human attention and mind meld with the machine in order to work on the hardest problems,
- 6:40
whereas the stuff that's commoditized, you can make it more async. So you can check out more on that details. What seems to be less appreciated is my, is the other work on Code maps, uh, which I, which I've more recently done, where we actually use AI to scale code-based understanding, which is also a fight, a way to
- 6:55
fight slop. And you can, uh, talk to the cognition folks, uh, downstairs, um, um, who, who can, can show you more in detail. Um, the, the last thing I always want to shout out as well is, is this trend of, uh, computer use.
- 7:08
I think computer use kind of debuted this time last year, uh, with Anthropic, but, uh, it's, it's getting really, really good now, guys. It can really autonomously operate the most complex apps, including an ID.
- 7:19
So that, I think that's really, uh, exciting, and you should probably use that to fight slop. We use it for the website, and here's a example of us using Devin to automate the sort of the, the, the sort of website ups- website updates.
- 7:30
Um, and finally, uh, something I learned from this conference yesterday is that you can also use subagents to fight context rot. Um, and I think that is one of the biggest themes of, of, uh, that I'm observing as well, if you want to take away something from this conference.
- 7:43
Um, and I... Al- also, one of the biggest highlights of the year for us as AIE, and myself personally, was chatting with Greg Brockman, who always, uh, preaches the concept of modularity, where you can sort of keep b- clear boundaries on what is human designed and let the AI code everything in between.
- 7:59
Um, so these are all ideas, but I just have this one message that I want to compre- compress down to you today that I want you to say with me.
- 8:07
No more slop. Yeah? [cheering] [clapping] Your boss tells you, "I want more lines of code in, by the end of the quarter." What do you say to that? Say it with me.
- 8:20
No more slop. No more slop. [laughs] You're fighting an asymmetric war. This is how bad it is, right? You have an insufficiently tested release that, that is potentially embarrassing to your company.
- 8:31
What do you say to people who really wanna push it? No more slop. No more slop. Exactly. [laughs]
- 8:38
Uh, your, your Twitter algorithm wants engagement bait, uh, and is sort of, you know, force, uh, and, uh, telling you to, to, to lie basically to the, to the broad public.
- 8:47
What do you say to that? No more slop. Exactly. That's it. Uh, I hope you have a great conference, and let's, [laughs] let's hear it for, uh, not having any more slop.
- 8:55
Thank you. [clapping] Thank you. [upbeat music]