insights-and-analytics

Across 25,392 US job postings that state both a salary and an experience requirement: pay rises 5.3% a year for a decade, then stops. Only 0.03% of postings ask for more than 20 years.

A new line has spread through AI job ads: "this is not a research role." Across 62,659 postings it turns out to be rare, it carries no pay premium, and the ads that use it often ask for the evaluation work they claim not to want.

Remote postings advertise 6% more than onsite ones, and that premium is a composition artifact. What survives is geography, and it belongs to the employer: the same remote senior engineering role runs $57,082 apart depending on where the company is anchored, while the state itself explains almost nothing.

Not one of the 28 named technologies that peaked before 2002 still holds half its peak. Then the churn stopped for seventeen years - and LLMs just climbed faster than anything we can time.

We traced 568,663 career histories back to the 1990s. Software Engineering fell from 69% of tech hiring to 41% - and the six roles that replaced it differ by 54x in how many of their people hold a PhD.

Research Scientist ($215K) and ML Engineer ($209K) top the senior rung. But the gap to a plain software engineer is $16K, while the gap between a mid and a staff engineer doing the same job is $65K.

Em dashes appear in 36.5% of ICLR 2026 peer reviews, up 6x in two years. But a blind judge's high-confidence machine-drafted calls stayed flat, model vocabulary fell, and non-native English markers dropped a quarter. It's polish, not drafting.

Data roles are a near-constant 10.5% of the tech org, from 10-person teams to 10,000+. What changes with scale isn't the size — it's the mix. And employers are hiring a very different mix than the workforce currently has.

For 52.5% of job postings, most of their closest skill matches carry a different job title. Product Engineer postings look like another job 81% of the time; Product Designer, 9%.

arXiv AI submissions grew 18.9x in a decade. We decomposed the growth: it's ~28x more researchers, each publishing just 9% more. And only 0.51% of 84,420 papers have a near-duplicate.