Everyone’s worried about a future where a handful of AI labs hoard the good stuff and the rest of us get a watered-down version.
Everyone’s worried about a future where a handful of AI labs hoard the good stuff and the rest of us get a watered-down version. That’s a real risk, worth taking seriously. But it’s not the gap actually deciding outcomes right now. The gap that’s already visible has nothing to do with access and everything to do with whether you’ve practiced directing an agent instead of just asking it questions.
Jarred Sumner, Bun’s creator, spent 11 days porting Bun’s 535,000-line codebase from Zig to Rust — running up to 64 Claude instances in parallel across separate worktrees, at a cost of about $165,000 in API usage. He estimates the same work would’ve taken three engineers a year by hand. Same tools everyone else has access to. The difference wasn’t the model — it was that Sumner knew the codebase well enough to direct dozens of concurrent workstreams and verify what came back.
Compare that to Klarna, which cut 700 support staff in 2023 and handed the queue to AI. Two-thirds of the volume got handled fine. The rest — the edge cases, the emotionally loaded conversations, the messy multi-step problems — didn’t. By 2025 the company was quietly rehiring. Its CEO put it plainly: “We focused too much on efficiency and cost. The result was lower quality, and that’s not sustainable.”
Its CEO put it plainly: “We focused too much on efficiency and cost. The result was lower quality, and that’s not sustainable.”
Same underlying technology, opposite outcomes. Sumner used AI to multiply a skill he already had. Klarna used it to remove the skill entirely. That’s not a story about AI capability — it’s a story about whether a human stayed in the loop directing and checking the work.
It’s not an isolated case either. Orgvue’s 2025 survey found 55% of leaders who made AI-driven layoffs now say the decision was wrong — a figure Forrester’s research landed on independently. The pattern holds everywhere I’ve looked: replace the judgment and quality drops; multiply it and output jumps by orders of magnitude.
I spent an afternoon this week running five things in the background at once — a model benchmark, a skill rewrite, a couple of code reviews — checking in on each rather than babysitting any one of them. It’s a different kind of tired than doing the work myself. Smaller scale than Sumner’s 11 days, but the same shape of thing.
Here’s the actually good news buried in that data: the skill that made Sumner’s 11 days possible isn’t gatekept. It’s not behind an enterprise contract or a research-lab badge. It’s the same Claude Code anyone reading this can open right now. What’s scarce isn’t access — it’s the practice of treating an agent as something you direct and verify at scale, instead of something you ask one question at a time.
What’s scarce isn’t access — it’s the practice of treating an agent as something you direct and verify at scale, instead of something you ask one question at a time.
Pick one task this week that you’d normally do by hand start to finish, and spend it directing and checking the work instead. That’s the whole skill, and it compounds fast once you start. The people who’ll be fine in five years aren’t the ones with the best model access — they’re the ones who put in the reps before it felt urgent.
— –
Sources
– “Don’t Go Quietly Into the AI Night” — Off-Policy, by Gavriel Cohen
— https://www.off-policy.com/dont-go-quietly-into-the-ai-night
– “Rewriting Bun in Rust” — Bun Blog — https://bun.com/blog/bun-in-rust
– “Klarna Reverses AI Push, Says Customers Prefer Human Support” —
Forbes — https://www.forbes.com/sites/quickerbettertech/2025/05/18/business-tech-news-klarna-reverses-on-ai-says-customers-like-talking-to-people/
– “The AI Layoff Trap: Why Half Will Be Quietly Rehired” — HR
Executive — https://hrexecutive.com/the-ai-layoff-trap-why-half-will-be-quietly-rehired/