Frontier Radar¶

This page records developments that change a judgment, not every piece of news. Each entry carries an evidence level and a next validation step.

Current focus¶

World models¶

Video generation is moving from “looks plausible” toward predicting future states and constructing interactive environments. The key tests are not single-frame quality, but object permanence, action-conditioned prediction, long-horizon consistency, and state representations useful for planning.

VLA¶

VLAs place visual observations, language tasks, and action sequences in one loop. The main bottlenecks remain data coverage, cross-embodiment generalization, failure recovery, and safe online learning.

Verifiers¶

Test-time compute and self-improving systems share the same foundation: reliable evaluators. Without independent holdouts, rollback mechanisms, and anti-reward-hacking tests, “self-improvement” may only be self-confirmation.

Evidence levels¶

L0 launch claim · L1 paper experiment · L2 public-code reproduction · L3 independent reproduction · L4 long-term deployment feedback

Do not treat L0 as a capability conclusion.

Update cadence¶

Write one short memo each month: which judgment changed, what the evidence was, and what to validate next month. Start by using GitHub Watch; add an RSS/Atom feed once the publishing cadence is stable.