Opinion

Two confident stories about AI and jobs, and one dataset at a time. The payroll records show a 19% entry-level gap. The national series shows nothing. Both readings can be true.

Randomised trials measure real benefits of AI, from 34% for novice support agents to a 44% cut in radiologist workload. Nine in ten executives report nothing at all. Where the gap comes from, and what would close it.

A company paid $63 million because nobody could tell an agent what the job actually was. That is not a model problem.

Give away the tier below your best, sell the product built on your best. That is not generosity. It is a funnel.

A three-week release cycle means the model you validated against is a generation old before your evaluation finishes.

A technology that needs four industries to make it safe to use has not finished being built. It has finished being sold.

Products assume someone who rephrases, verifies and explores. Most people ask once, believe the answer, and leave if it is poor.

Hallucination implies a malfunction. Correct and incorrect output come from the same process, and the vocabulary decides what people trust.

Latency, reliability, interruption and the cost of checking. Most AI features trade a function that always works for one that usually does.

Error compounding, absent memory and no failure recovery. The best evidence agents work arrived as a security incident, not a launch.

A 27x revenue multiple, a $500B guarantee against hardware with a four to six year life, and 5x annual compute growth. What breaks first.

Leakage runs from 1% to 45% on popular benchmarks, and even Meta found 16% of MMLU contaminated. The scores measure memory as much as ability.