The Highest Loop — swyx
Read the talk
The Highest Loop
swyx connects the work of choosing an AI system’s level of abstraction to the human work of gathering, deciding what to build next, and carrying those decisions back into teams and daily life.
From a talk by Shawn "swyx" Wang
At a glance
Ideas worth remembering
Choosing the right abstraction is part of AI engineering: move upward to pursue productivity and inference scale, and downward to diagnose reliability problems.
Use the conference’s tracks to address the level where you work or the level where you are blocked.
The highest loop is collective work that decides what loops to create next; its practical result should travel back into teams and personal workflows.
From tokens to scheduled work
A token, a chat, a tool call, a goal, an automation: swyx opens his AI Engineer World’s Fair address by moving through increasingly broad units of AI work. The progression reaches cron jobs and software factories, where the concern extends beyond an individual response to repeated activity and the organization of that activity. He calls this “loop stacking.”
The sequence deliberately skips steps. Its purpose is to introduce a question that survives changes in tools: which loop are you working in? The essay he names, Loopcraft, gives that question a home. Here, the core engineering skill is recognizing the current level of abstraction and deciding whether it is the useful place to work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Move up for scale, down for reliability
The practical decision has two directions, each serving a different need:
- Move up when ready to scale. Work at the next level of abstraction to pursue greater productivity and scale inference. The unit of attention becomes a broader loop.
- Move down when something breaks. Return to a lower level to investigate reliability and fix the bugs that the broader view leaves unresolved.
“A stack of while loops” is the programmer’s analogy. It makes movement between levels feel ordinary: engineers, leaders, researchers and founders all need to reason within a loop and step into or out of it. This is an organizing model for work; the address does not supply a software architecture or a test for when a system is ready to move upward.
That model also explains the conference’s abundance of tracks. “I’m not gonna apologize for good content,” swyx says. The attendee’s job is to route themselves toward the level where they currently work or where they are blocked. Choosing a session becomes a practical version of choosing an abstraction: find the material that addresses the problem at its useful level.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The human analogy expands from heartbeat to collective work
The human version begins with a heartbeat, then expands through speech, tool use and work: “eat, sleep, code, repeat,” with prompting offered as an update. Teams, small groups and families widen the frame again. Tiny teams matter here because they place individual activity inside a shared undertaking.
The analogy then reaches beyond personal, family, company and national contexts to collective intelligence and the human projects those groups contribute to. The question changes from how an individual repeats useful work to where that work is taking everyone. This is a proposed way to think about coordination and purpose, rather than a claim that human life literally executes like a program.
The conference program gives that expansion a concrete expression. Healthcare, finance, go-to-market and forward-deployed engineering receive tracks as AI work spreads beyond coding. swyx presents this as agents generalizing across domains; the tracks show the intended scope of that exploration, without establishing how reliably a capability transfers from one domain to another.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The organizers’ stress curves
The address briefly returns from civilization-scale coordination to the very human problem of selling conference tickets. The first AI Engineer event in 2023 had a sales curve swyx describes as roughly linear. The first World’s Fair in 2024 brought about 2,000 people. For the current event, he corrects his initial figure: the target was 6,000, and the total had crossed 7,000 that day.
What do the sales curves reveal about when people commit? That is the useful comparison behind the joke about organizers’ “stress curves.” Earlier ticket purchases had been requested, but purchases had become more concentrated instead. swyx describes the Gini coefficient as rising—“the wrong direction”—and asks attendees to join earlier next time. No coefficient values or calculation details are given, so the supported lesson is about the reported concentration of buying, rather than its numerical size.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The loop that makes loops
Why put this work into a conference? The closing definition supplies the title: the highest loop is where humans gather to figure out what the next loop should be. “The loop that makes loops” puts the choice of future work inside a collective activity. A gathering earns its place when it helps people decide what processes to create next.
AI Engineer’s summits give that idea an organizational form. swyx names New York, Paris, Singapore, Melbourne and Miami, describes organizing with partners, and invites people to organize in their home countries with help from a partnerships team. He calls the World’s Fair the “summit of summits”: a gathering that brings the wider network together.
The final request is to carry something home: find a generative process for creating loops, then propagate it into teams, personal workflows and life. The intended output of the gathering is a way to keep choosing and creating useful work after the event ends.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Read the complete timestamped transcript
- 1:05
In the beginning, there was the token.
- 1:10
Then there was the chat.
- 1:13
Then we learned to use tools.
- 1:17
Then we learned to set goals. I'm skipping a few steps.
- 1:22
And these days, we're all about automations. We're all about cron jobs and the loops.
- 1:29
And the rest of today on this livestream, uh, the, uh, those of you watching at home, uh, we're gonna be talking about software factories as well. So there's a lot of s- loop stacking. Uh, I wrote this essay called Loopcraft. Uh, I think Joel in the audience did a workshop on it yesterday. Can you still hear me? Are we good? Sorry, I, I, um, had a mic slip. Um, and I think this is something that is basically a core skill of AI engineer, uh, is
- 1:58
defining what loops you're working in and figuring out whether that's the right level of abstraction. And when you're ready, going up a s- loop in order to figure out the next level in productivity and scaling your inference. And when you're finding issues, going down a loop and going lower level in order to figure out reliability and fixing whatever bugs you, you find. If you want to simplify it, uh, I have a, I have a better analogy that will be less complicated. Uh, it's
- 2:28
literally just a stack of while loops, right? I think all of you in, in the audience and the programmers will understand this. Uh, and effectively, this is what we're here to do today. We're, um, I expect our engineers, our leaders, our researchers, our founders, um, to all be understanding and recursing in as a, uh, as a, as a sort of human general intelligence, all these loops that you guys are capable of, of, um, going up and down. And that is why there's so much
- 2:58
content. Um, I often tell people, a few who, who complain about, uh, the tracks, "I'm not gonna apologize for good content." Uh, it's up to you to decide and route yourself, uh, as to which level you, you're currently working at and which level you're currently blocked on. Um, so the thing that's occupying most of my time, um, recently, and, uh, what I, what I make the human analogy for, is what is the human loop? Um, and in the beginning, there's the heartbeat.
- 3:28
When you're, when you're sort of in the womb, you're, you, you sort of get that heartbeat. That is the proof of life. Um, then you learn to talk. Then we learn to use tools. Then we start to work. We eat, sleep, code, repeat or prompt. Um, and then we start working in teams. We start working in small groups and families. We- we've, uh, promoted the concept of tiny teams before, and I think that is increasingly important with, uh, everybody. Um, and I think
- 3:59
the interesting thing to think about is, like, the broader human projects as a whole. Like, where are we all going? You know, uh, extend out from your personal context, your family, your country, your company, whatever, um, to the broader project of the human civilization and what we're building, uh, with collective intelligence. Uh, I think it's very interesting to, uh, consider that. So, um, I, I think that's a little bit what I want to leave you with today. Like, the things that you're doing with coding are definitely crossing over to human life.
- 4:29
Um, the message of my last keynote at AI Engineer was that the, you know, it's very much like agents that you see are working in one domain are definitely generalizing this year. Um, this is the year that we're adding vertical tracks like, you know, AI in healthcare, uh, AI in finance, AI in GTM. Uh, we have a forward-deployed engineering track this, uh, today, which I'm very excited about. Um, all of this is meant to see how AI is just diffusing finally beyond coding. Okay. Um,
- 4:59
last World's Fair, I gave you all an update on the stress curves, uh, of the, of the, uh, the organizers. Uh, just to give you an idea, this is, um, the first World's F- uh, first AI Engineer in 2023. Uh, this is sort of it's, uh, the sales curves basically of, like, how people bought. Basically, it was kind of linear. Then we had the first World's Fair, 2024, uh, about two thousand people. 2025, uh, was, uh, I think I maybe skipped, uh, skipped a few lines. But we're, we're basically at six thousand,
- 5:30
um, and if you sort of calculate-- Uh, sorry. We, we, we, we aimed for six thousand. We actually crossed and go- went to seven thousand today. Um, and, uh, basically, I asked you all to buy your tickets earlier. You did not. Um, uh, if you measure the Gini coefficient, uh, it's gone up and not down. That's the wrong direction. Uh, so if y'all, if y'all can just gather, uh, you know, your- yourselves and, um, you know, join us earlier, I think that would be great. Anyway, um, why we want to work at a conference,
- 6:00
why we, why I work on a conference, why you all come to a conference is basically the concept of The Highest Loop. Um, I, I will submit to you that The Highest Loop of all is to, uh, if I can make this slide work. Oh my God. I, I'm gonna refresh. Okay. Um, uh, that The Highest Loop is where humans come together to figure out what the next loop is, right? The loop that makes loops. Um, so we organize AI Engineer summits all around the world. Uh, this year we have, uh, summits in New York,
- 6:30
summits in Paris, summits in Singapore, um, and, uh, you know, again, uh, uh, Melbourne and I'm forgetting one already. M- Miami. Um, and, uh, we or- organize this with partners. We want you to organize it at h- in your home countries. We're, we're having, uh, a partnerships team, uh, to do this. And this is The Highest Loop of all, which is the World's Fair, which is the summit of summits. So you're-- I'm very excited to welcome all of you today. Um, I hope that you figure out your generative process for generating loops and propagate it down to your teams,
- 7:00
to your personal workflows, to your lives. Thank you very much, and enjoy the show.