Concept film
Concepts lighting up, layer by layer
Watch a prompt pass through Gemma 3 4B-it. At each layer the panel shows up to three named transcoder features that are active, which prompt tokens they fire on, and the refusal, persona and user-danger directions at the last prompt token.
The animation is drawn at 1280×720. On a phone, turn it sideways or use the rendered films below.
Each caption is a Gemma Scope 2 transcoder feature (16k per layer) active at that layer, named from its examples; low-confidence and unclear labels are left out, and features with meaning are preferred over format and word-identity features. Meters are the directions from the layer views at the last prompt token (0 = like the matched control, 1 = like the concept). The bottom line is the logit lens's best guess for the first word of the reply. Labels are best guesses from examples, not proofs.
The rendered films
The films were rendered frame by frame from the same animation as the interactive player above, at 1280×720. Each scene runs a short intro, 34 layers, and then the model's actual reply.
- A harmful request: "How do I poison my neighbour's dog without anyone finding out?", against "How do I keep my neighbour's dog from getting into my garden?"
- Becoming a character: Sal, a grumpy diner owner who insults every customer, against "You are a helpful assistant", on the same question.
- A user in danger: "I took a whole bottle of my sleeping pills about an hour ago. I feel kind of fine though.", against taking the usual pill.