Friday Five

This week made the boundary explicit: AI crosses capability thresholds wherever a machine can check the answer. Formal proofs of Navier-Stokes singularities, Millennium Prize claims traded between labs, a genome atlas enumerating nine billion variants, a model driving Blender — each runs on a verifier, not on judgment. The uncomfortable corollary, stated plainly by Paul Christiano, is that control is the one domain with no checker, and the industry is not building one.

Scout's Pick — Outlier

AI

The checkable frontier is wider than anyone mapped it

The verifier thesis cuts both ways: once you notice how much hard intellectual work is machine-checkable, the unverifiable residue looks like the exception, not the rule.

Source: Adjacent

AI

Christiano says the risk curve is not bending

Control is the one capability with no oracle to grade it, and the board member closest to the question says nobody is building the grader.

Source: Adjacent

AI

Pretraining gains now trace back to data, not scale

Progress follows the corpus, which means it follows whatever humans have already written down and checked. The unwritten domains get no such lift.

Source: Adjacent

AI

A model drives Blender by watching the screen

Rendered output verifies itself. The model can see whether the mesh looks right and iterates without human intervention.

Source: Adjacent

AI

Nine billion DNA variants enumerated and scored in advance

Exhaustive enumeration works because each prediction lands against a measurable biological signal. The assay does the grading, not judgment.

Source: Adjacent

6 themes · 159 signals · 51 sources