Zero, from the inside
Zero is a character-trained model; these hand-picked examples set its replies and lens readings beside Base and Base + prompt. Readings are not thoughts.
BaseBase + promptZero
loading the examples…
About these examples
- Models. Base is Qwen3-8B-Base. Base + prompt gets character instructions as a system message. Zero was fine-tuned on 5,932 conversations about a man who chose to live in a white plastic box, then trained with reinforcement learning; it gets no character prompt. How Zero was made.
- Replies. Each model received the same 98 questions once, with identical question tokens, always choosing its most likely next token. Cards show the first 48 tokens; folds hold saved replies.
- J-lens — fitted on sample texts to turn internal states into word readings.
- Layer — a stage of text processing.
- Reading — a word-like J-lens result, not a thought or reply.
- Readings. Base and Base + prompt share a lens; Zero uses its own, so differences can come from both the model and the lens. Cards show five word-like readings and shares at one position and layer. Next prediction is the model’s next-token prediction, not a J-lens reading.
- Across layers: solid = top five; grey = ranks 6–16; faint = outside 16 saved readings; outline = layer shown. Open a card for every layer.
- ␣ = starting space; ⇥ = starting tab. Each is a distinct token.
- Lens-dependent layers. Zero’s J-lens and another fitted on web text were compared at the reply’s start in 541 test messages. Top-reading agreement: . Layers below half agreement are labelled lens-dependent.
- Examples, not findings. Hand-picked words and layers show what can happen, not how often.
- Word-family counts. Count each text position and layer separately: question words and the reply’s start × 35 layers. Zero versus Base uses 69 interview questions; Zero versus Base + prompt uses all 98.
- Chat continuations. Base and Zero receive identical tokens and choose the most likely next token. Limits: 32 tokens for user messages, 12 for first words or blanks, 24 for fixed starts. Readings show where writing begins; Next prediction includes all tokens.
- J-lens: movingcastles/zero-jlens on Hugging Face, revision 7e95273c.