Three separate people in today's feed are making the same argument from different angles: AI was supposed to shrink demand for knowledge work, and instead it keeps expanding it. The debate has a name now, and it's worth knowing.
01
Software job postings are up. AI was supposed to kill software jobs.
Box CEO Aaron Levie points to data that's inconvenient for the "AI replaces coders" narrative: software job postings are outpacing other fields right now, not declining. His explanation is Jevons paradox, the 19th-century observation that making something cheaper tends to increase total consumption, not reduce it. Lower cost per unit of software means people want far more software built, for use cases that weren't worth the cost before. ---
Why it matters: If you've been waiting for a wave of developer layoffs to confirm AI's productivity story, this is your counter-signal. The companies quietly winning right now aren't replacing their engineering teams. They're using cheaper software production to expand into adjacencies they couldn't afford to touch two years ago.
The Jevons argument, but for all of knowledge work
Swyx, the AI developer and researcher behind Latent Space, takes Levie's point and extends it further. His argument: most people encountered Jevons paradox through the lens of software specifically, but the same logic applies as coding agents "break containment" into all other knowledge work. As the unit cost of any knowledge task falls, total demand for that work goes up. Coding was the first visible case. It won't be the last. ---
Why it matters: Your legal team, your finance team, your marketing team are all about to face the same dynamic developers did. The people who thrive won't be the ones who worried about AI replacing them. They'll be the ones who used AI to take on work their department previously couldn't justify doing.
Thibault Sottiaux posted a five-minute setup guide for pointing Claude Code (the "orange crab") at OpenAI's GPT-5.6 Sol model instead of Anthropic's own. The trick uses a proxy called CLIProxyAPI and a shell alias. It's a workaround, not a supported integration, but the 3,800+ likes suggest a lot of developers were looking for exactly this. --- *Dropped: Garry Tan on California housing reform (no AI content) and Matt Turck on Argentina's World Cup match (sports, not AI).*
Why it matters: When developers are writing shell aliases to swap models inside a competitor's tool, it tells you the interface has won and the underlying model is becoming a commodity slot. Claude Code is sticky. The model powering it is increasingly negotiable.