Can junior engineers outrun AI?
Anthropic recently published data on how jobs are changing inside the company. One number stuck with me: juniors went from 27% of its technical hires in early 2025 to 9% today. It isn't just one lab. Before the pandemic, new grads were about 15% of hires at the biggest tech companies. By 2024 it was 7%.
The usual explanation is "AI can do junior work now." I wanted a sharper question: how fast does a junior have to grow to stay ahead?
The race
Think of your scope as the biggest task you can own end to end, measured in hours of skilled work. A new grad owns something like a one-day ticket. A staff engineer owns projects that run for months. METR measures AI in the same unit, the length of task a model can finish, and that number has been doubling every four to seven months.
So I set up a race. Your scope grows along a typical career ladder. The AI bar starts at new-grad scope in December 2025, when Anthropic says its junior postings fell off, and doubles every 6.5 months, METR's long-run trend. Fall under the bar and an agent can do your job. Reach a safe level first, say owning six months of work (roughly staff), and you've escaped.
[Chart: race.png goes here]
What it says
Take someone who graduated in 2023:
- Today they're about 4× ahead of the AI bar.
- On a typical career path, the bar catches them in September 2028.
- The bar reaches staff-level scope in October 2029. A typical engineer from that class gets there in November 2032, three years too late.
- To make it, they'd have to climb the ladder about 2.1× faster than typical, around +123% growth in scope a year on average.
[Chart: escape.png goes here]
Every recent class looks like this. Under the long-run trend, only the class of 2022, on a fast track, reaches staff before the bar does.
The pace is everything
Change one number and the story flips. At METR's recent pace, doubling every 4.3 months, the bar reaches staff level by mid-2028 and nobody escapes. If progress slows to doubling once a year, the deadline moves to December 2032 and the classes of 2022 and 2023 make it on a typical path.
So the real question isn't how good juniors are. It's whether AI keeps compounding faster than careers do, and careers compound slowly: maybe 1.5–2× a year early on, less later.
[Chart: bigtech.png goes here]
If the 2019–2025 trend simply continues, new grads fall from 7% to under 4% of big tech hires by 2030.
What this leaves out
A lot. AI also speeds up your own growth. Demand for software could keep growing. AI is uneven, great at some tasks and weak at others. And jobs change shape: Anthropic says its engineers increasingly direct agents instead of writing code. The career ladder is my own estimate, and the headcount line is modeled, not measured.
But the arithmetic is hard to argue with. When one side doubles every half-year and the other every year or two, a head start buys you time, not safety.
You can change the class, the AI pace and the safe level yourself in the interactive version.
Sources: Anthropic, "How employment is changing at Anthropic"; SignalFire State of Talent reports; Indeed Hiring Lab; Stanford Digital Economy Lab; METR time-horizon results.