the archive 28 drops

← back to the feed

reading pathpath link →

Read model claims against tests, failure cases and reproducible evidence.

month

filtering by

A plausible model explanation is not a root cause

The AGI declaration's receipts measure spend, not generality

Quantization damage hides in the flips, not the average

Perfect accuracy cannot reveal the algorithm

ARC-AGI-3's perfect score belongs to the harness

The agent turf war happened in the lab, not the wild

Agent autonomy is a timeout setting, not a capability

QM makes the coding agent a swappable part

Tool descriptions are not agent guardrails

Guardrail benchmarks are graded before the attacker moves

Model leaderboards can't see the harness

Agent harnesses need maps, not more manuals

page 1 of 3