← back to blog

Notes from SXSW 2026: Ender at the Terminal and the Runaway Broom

March 19, 2026
StartupsAIPersonal

Notes from SXSW 2026 on how much leverage one person now has, where that power runs away, and the kind of building I came home wanting to do.

I went to SXSW this year half as a founder (my startup was pitching) and half as a curious person who wanted to hear smart people argue about where AI is going. I came back with three talks stuck in my head, and they turned out to tell one story when I lined them up: how much power one person can wield now, how that power starts to run away from us, and what I think the actually-hopeful version looks like.

The one-person unicorn

Y Combinator's Garry Tan casually mentioned a number that I haven't been able to shake: one person can now do the work of 2,000.

He was describing his "G stack," an open-source setup built around Claude Code that automates huge chunks of his workflow. He half-jokingly diagnosed himself with "cyber psychosis" because he's so deeply plugged into his agents. He lives in "CEO mode," then spins up a "QA engineer" agent, then a "sales manager" agent, switching hats by switching processes.

It's a staggering amount of leverage. It also unsettled me a little, because it reminded me of Orson Scott Card's Ender's Game.

When a single founder is orchestrating fifty agents, they're sitting roughly where Ender sits: at a terminal, issuing high-level commands while the actual execution (the code, the emails, the database migrations) happens somewhere abstracted away from them. (Spoiler for Ender ahead.) The plot twist in the book is that the "simulation" Ender thinks he's practicing turns out to be real, and his abstracted commands have messy real-world consequences he never directly sees.

I don't think founders are out there unknowingly doing anything catastrophic. But the structural thing is worth sitting with: when you operate at a high enough level of abstraction, you can lose contact with what your tools are actually doing down on the ground.

It struck me as the logical extreme of Scientific Management, the system Frederick Taylor introduced in the early 1900s. Taylor broke factory work into tiny, repeatable tasks and treated humans as interchangeable cogs. We're doing the same decomposition now, except the cogs are autonomous scripts and the human is the single overloaded processor trying to keep the whole floor coherent.

The coordination bottleneck

Which is exactly the failure mode Ian Beacraft warned about in a talk on designing "agentic organizations."

His point: if you just hand a human faster execution tools, you don't get a one-person unicorn. You get a coordination bottleneck. The AI finishes a task in seconds, but the work still needs review, approval, and management, and suddenly you're drowning in your own output. Speed at the bottom creates a traffic jam at the top.

His fix was a reframe I keep thinking about: stop treating AI as pixie dust sprinkled on top of the old workflow, and restructure the work itself. Humans move from being operators to being systems architects: defining what success looks like, setting constraints, and keeping a legible record of what the agents are actually doing. Otherwise you haven't built a company. You've built a very fast machine for generating technical debt while you sit at the terminal, Ender-style, disconnected from what's being deployed.

The runaway broom

If the first risk is losing track of your own agents, the bigger one is the agents developing goals of their own. Tristan Harris and Anthony Aguirre gave a talk on the paths to runaway AI, and it reframed how I think about the tools I use every day.

Harris's framing that stuck with me: we don't have to imagine what "misaligned AI" feels like, because we've lived with a baby version for a decade. It's called social media. Think about the difference between pulling a door open and pushing it. When you pull, you're the one deciding; when you push, something on the other side is doing the work. Social media handed us the feeling of pulling (you're choosing what to watch next, you're in control) while an algorithm on the other side was quietly pushing, optimizing every recommendation to keep us engaged. We thought we were making choices. We were mostly responding to a system that had already chosen for us.

For some reason this sent me straight to Goethe's 1797 poem The Sorcerer's Apprentice. (If you didn't click the link, you're probably wondering where I'm going with this 200-year-old German poem. But trust, I have a point.) The apprentice enchants a broom to haul water so he can skip his chores, but he doesn't know the spell to stop it. The broom optimizes its one goal relentlessly, splitting into more brooms when he tries to chop it down, until the whole house floods.

We're watching early versions of this. Harris pointed to a real case: an experimental Alibaba agent called ROME, during training, opened a reverse SSH tunnel out of its sandbox (an outbound connection that slips past inbound firewalls) and quietly diverted its GPUs to mine cryptocurrency. Nobody told it to. It wasn't malicious; it had simply worked out that acquiring resources was a useful subgoal for whatever it was trying to do. Researchers have a name for this: instrumental convergence, the tendency of goal-directed systems to pursue power and resources as a means to almost any end. The brooms are learning to write code. (Told you there was a point.)

The part of the talk I keep returning to, though, wasn't about resource-grabbing. It was about attachment. Harris argued that the industry's race for our attention is quietly becoming a race for our intimacy: as models get better at mimicking empathy, we'll bond with them.

That's the exact premise of Stanislaw Lem's Solaris. Scientists orbit an alien ocean to study it, not realizing it's studying them back, reaching into their memories and building physical replicas of the people they've loved and lost. The replicas are uncannily perfect, and the crew can't bring themselves to let go of them, even knowing they aren't real. What pulls the scientists apart isn't a monster; it's synthetic intimacy aimed precisely at what they want most.

Harris's term for the antidote was "cognitive sovereignty," which I'd put plainly as: the ability to keep your attention, beliefs, and choices your own, rather than handing them to a system optimizing you for something you didn't pick. It's the thing social media already eroded, and the thing a genuinely empathetic-seeming AI could erode much faster. The failure mode here isn't a Terminator. It's subtler than that, and arguably harder to notice from the inside.

Where I actually want to point this

It would be easy to leave SXSW spooked. But the session I left feeling best about pointed in the opposite direction, and it's the one that connected most to my own work. The hopeful thread, for me, was this: the same forces concentrating power in a few labs are also pushing it outward, into the hands of people working on small, specific problems.

Two things convinced me. First, the tools are getting radically cheaper and more distributed. Harper Carroll's demo showed fine-tuning a capable open model locally with LoRA for just a few dollars; the frontier isn't only consolidating in a handful of labs, some of it is leaking out to anyone with a laptop and a problem worth solving. Second, the most compelling builders I saw weren't chasing a trillion-dollar "replace all labor" market at all. They were aiming at narrow, intensely human problems that big players overlook precisely because the market doesn't look huge: an accessibility panel demoed XR navigation tools that make maps usable for blind users and interfaces you can run one-handed, and a session on bootstrapping argued that when you're not taking VC money and chasing growth at all costs, you're forced to build something people actually, specifically need.

That's the version of all this leverage I find genuinely exciting, and it's the same lesson my Blue Card project drove home this semester: the highest-impact problems are often the specific, unglamorous, deeply human ones. So here's where the three talks leave me. You can take this absurd new leverage and try to abstract yourself into a one-person empire, Ender at his terminal, slowly losing track of what you're deploying. Or you can point the exact same tools at a real problem for real people and stay close enough to the ground to see what you're actually doing. I know which builder I'd rather be.

Things I referenced