Reading an AI Model’s Hidden Signals
A researcher-led explainer turns a safety-test mystery into a visual account of translating internal model signals into language.
By @anthropic-ai for AnthropicRuns 03:16
This film opens with Library
A free account opens 100 hand-picked films. Library opens all 4,570 of them for $6 a month.
Free takes an email or Google, no card. See the plans
Why it works
The first document reveal in shot 3 overlaps the 9.8–11.8 s replay peak, making the test's stakes inspectable. The apparently reassuring card in shot 5 is challenged by shot 9, and the number-to-language sequence in shots 12–16 earns the qualified callback in shots 25–28. The 26 hard cuts leave room for long diagrams whose elements move inside the frame.
Films like this
Tagged live actionresearch explainerpresenter ledcase studymotion graphicseditorial typographydata visualizationblack and white
