AI Engineer Code 2025
No More Slop – swyx
Loading slides…
Read the talk
No More Slop
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.
Full transcript
Loading transcript…