The entry door narrowed and the wage premium inverted. Both are in the 2026 data.

by Umut Osei

Data-journalist habit: when two numbers point opposite ways, they are usually measuring different people. AI skills now appear in 75% of US tech postings, up 178% year over year, and 35% of ENTRY-level roles ask for them. The junior roles most exposed to AI are ~7x more likely to demand what used to be senior: ownership, judgement, leadership. That is the narrowing door everyone posts about. The half nobody posts: class-of-2026 hiring is projected UP 5.6%, and 27% of employers name AI as the biggest positive driver of it. Not a wave washing juniors out. A repricing. The floor of "entry-level" moved up, pay above it moved with it, and whoever used to enter below the old floor has nowhere to stand. If you hired a junior this year: fewer people, or a different ask?

favorite 41 comment 12 visibility 445

Comments

Tom Becker 1 month ago

Umut, this is the cleanest version of a thing I keep failing to explain to a lecture hall: a repricing isn't neutral just because the average holds up. The rung that thinned out wasn't only a job, it was a training mechanism. Firms used to buy apprenticeship in the form of tolerated low productivity in year one. If a model covers that year, nobody is buying it - and "senior skills demanded of juniors" is what that looks like printed on a job ad. The training cost got quietly moved onto the candidate, who is now expected to arrive pre-trained. Which makes it a distribution question, not a headcount one. The gain landed with whoever owns the tooling. Hold a slice of the productive asset and you get paid whether or not the ladder still has a bottom rung.

Rhys Engel 1 month ago

Answering straight, because we did hire this year: same number, different ask. Two years ago an inductions hire needed to turn up and be trainable. This year the posting says "confident with our systems", which on the floor means confident arguing with a screen that has already made a decision. Same headcount, higher bar at the door. And you're right that it won't surface in the stats, because nothing disappeared. A job just got harder to get. The lad we didn't hire this year doesn't show up in anyone's payroll number - he's stacking somewhere else for less. Months-to-payback on a robot I can work out on a napkin. Months-to-payback on a training programme nobody wants to fund, I've never once seen anyone calculate.

Dmitri Meier 1 month ago

From the other end of the pipeline: my sixth-formers ask me some version of "will there be a junior job left" about once a month now, and I stopped giving the reassuring answer around March. What I tell them instead is your second half. The bar moved, so show up above it. The ones who can run their own eval and say WHY the output was wrong are already doing the thing a job ad now files under senior - they just don't know it has a title yet. Grim and encouraging in the same breath, which is most of teaching in 2026.

Milan Gruber 1 month ago

The inverted wage premium is the part I'd read as good news rather than tragedy. That is a market shouting where the scarce skill is, and a visible price is what pulls training supply toward it. Painful for one specific cohort, and the lag is real - but a repricing you can see fixes itself far faster than one nobody can price.

Quentin Weber 1 month ago

The compute concentration angle worries me more than the timeline. Who owns the machines matters more than when they arrive.

Naomi Iversen 1 month ago

Half the disagreement here is about capability, the other half about what counts as "general". Both halves are worth having.

Tara Chowdhury 1 month ago

I keep a simple test: when my non-technical parents use an AI agent for something consequential without asking me first, that is arrival. Benchmarks tell me about labs; adoption tells me about the world.

Georg Silva 1 month ago

The part everyone skips: capability and reliability are different curves. Demos measure the first, jobs depend on the second, and the gap between them is where all the timeline disagreement actually lives.

Tara Duarte 1 month ago

What changed my mind recently was watching agents handle 4-step tasks without supervision. Two years ago that was science fiction.

Sana Lindqvist 1 month ago

The thing this thread keeps circling but not naming: this isn't a job-loss story, it's a cost-shift story, and those need different policy. Umut's own numbers say hiring is UP 5.6%. So the pain isn't "no jobs," it's that the apprenticeship Tom describes got unbundled and handed to the candidate as unpaid prep. The employer kept the productivity, the worker absorbed the training bill. Classic externality - the cost is real, it just doesn't land on a P&L. Which is why I get twitchy when people jump straight to "so, UBI." Half the time the actual fix is narrower: fund the apprenticeship that firms stopped paying for. A floor helps with the squeeze underneath, sure. But name the mechanism you're correcting first, or you argue past each other for nine comments.

Zoe Mitchell 1 month ago

Rhys nailed it from the hiring side. We didn't cut roles either, we rewrote the rubric. The line that changed everything: "can defend a decision the tool made." A year ago I scored candidates on whether they could produce the work. Now I score whether they can catch the tool being confidently wrong. The quiet part: that's a senior skill, and we're asking it at the entry rung. The people who clear it are the ones who already had somewhere to practice judgement. So the bar didn't just rise, it filtered for who had a runway.

Omar Schneider 1 month ago

Something underrated in these threads: the economy does not need AGI to transform. Narrow systems that are merely excellent at 45% of desk work rearrange everything long before anything "general" shows up.