Sean Wang, known as swyx, runs the Latent Space podcast and the AI Engineer conferences. He has a rule for the people he pays: "I can pay for my own psychosis. That's fine." He said he doesn't want to pay for someone else to go through "LLM psychosis." "When you work for me and I'm paying your tokens, you better be like actually producing thoughtful stuff," he said. Tokens are the units AI providers bill by, so an employer pays for every prompt an employee sends.
Wang made the remarks to Nathan Labenz, host of The Cognitive Revolution, in a conversation included in the show's weekly highlights edition, published October 8, 2026. Labenz asked how software engineers are feeling right now: empowered or threatened?
More demand, but of a specific kind
Wang said college kids are "a bit worried." For everyone else he was optimistic. If you are "relatively plugged in and very capable of AI engineering tooling," he said, "you're in more demand than you've ever been because your expected value is higher than it's ever been."
He called this a Jevons paradox situation. The term comes from economics: when something gets cheaper to use, people can end up using much more of it. Wang argued that because creating software now costs less, demand for it has grown a lot. But the demand is narrow. It is for people who can manage coding agents productively, he said, "instead of producing a whole bunch of slop." Coding agents are AI systems that write and change code largely on their own.
Paying for slop
Wang described himself as both an engineer and an employer of engineers, and of people who are not engineers but write software by vibe coding, meaning they describe what they want and let the AI generate the code. He said two or three of his employees are currently "basically under performance review" because they keep handing him low-quality output from Anthropic's Claude.
According to Wang, they don't see the problem. "They're like, what do you mean? I think this is perfectly fine," he said. His reply is that they are not producing value: he could prompt Claude himself. "I don't need you."
Where years of experience still count
Labenz asked whether years of coding experience still matter when Wang hires, or whether product sense and high standards matter more. Wang said he mostly doesn't care about experience, with two exceptions: security roles and anything involving back-end scalability, meaning the systems that must keep working as usage grows.
For the second exception he pointed to an interview he had just recorded with founders who are scaling Postgres, a widely used database, "to a scale that we've never seen before." He described a self-managed sharding solution that works at YouTube scale. Sharding means splitting one database across many machines. Wang said only one person in the world can build that, and the company hired him. In that kind of job, he said, you have to avoid losing data, avoid downtime and scale economically. The pressure is growing, he said, because agents consume databases at something like 500 times the rate of human developers.
The transcript renders the founders' company as "super based," and Wang did not name a specific project. The likely context is Supabase's Multigres project, whose first alpha release Sugu Sougoumarane announced on June 4, 2026. It describes infrastructure for running Postgres with high availability, connection pooling and backups. That release covered a single shard: sharding was not yet included, and Supabase labeled it alpha rather than production-ready.
Taste, juggling and reading the data
Outside those specialties, Wang said, "you can mostly buy code," and what matters is taste. Long experience can even work against people, he argued, because they have "a set way of doing things" and don't know how to run more than one agent at a time. He said engineers should now be comfortable juggling five to 10 ongoing things. That makes them "much more of a manager than a individual contributor," meaning someone who does the work personally.
He also said people who look at data rather than code are more valuable now. In practice that means capturing logs and traces (step-by-step records of what a program or agent did), the database schema (how its data is structured) and the actual inputs and outputs, then turning them into an eval. An eval is a repeatable test of whether an AI system does its job. Wang described this as the merging of AI engineering and machine-learning engineering, and said people need to build up those skills.
Check the modules, not every line
Wang contrasted two ways of reviewing code. An engineer with 10 or 20 years of experience, he said, "really reviews every line and tries to make every line make sense." Now, he argued, "we just need to make the modules make sense." A module is a self-contained part of a program with a defined job. He said he allows some slop inside modules because it helps him go faster, as long as he contains it in places where he understands the whole system. Things go wrong, he said, when there are too many modules and too many black boxes he doesn't understand. The code itself also gets confused, he added.
His example came from a Slack competitor he is building. Messages weren't loading. He refreshed and they loaded. He refreshed again and they didn't. "It's the same exact code. What the hell is going on?" he recalled thinking.
The cause was two code paths and a race condition, where the result depends on which of two processes happens to finish first. Two different coding agents had worked on the feature at different times, Wang said, and each made its own thing. A human would never do that, he said, but an agent sometimes will, because things fall out of the context window, the limited amount of text a model can keep in view at once.
His answer is to keep oversight at the module level. A person should understand and own each module, he said, and "whatever inside is inside of the module can be a black box."