IA-004 / 2025
On the Biology of a Large Language Model is a node-link graphDots joined by lines, showing what connects to what.Nodes joined by edges, showing what connects to or causes what.glossary showing circuitsA chain of parts inside the model that work together to do one job.Connected components and the paths between them.glossary, about planningThe model deciding where a sentence is going before it writes it.Selecting a future output before generating the text that leads to it.glossary, hallucinationThe model stating something confidently when it has nothing to base it on.Producing confident content for which the model has no basis.glossary, jailbreakGetting a model to do something it was trained to refuse.Circumventing a model's trained refusals.glossary and unfaithful-reasoningThe explanation a model gives not matching the reasoning it actually did.Stated reasoning that does not correspond to the computation actually performed.glossary.
- ID
- IA-004
- Name
- On the Biology of a Large Language Model — attribution graphs
- Artefact type
- Interactive article
- Visual form
- Node-link graphDots joined by lines, showing what connects to what.Nodes joined by edges, showing what connects to or causes what.glossary
- Makes visible
- CircuitsA chain of parts inside the model that work together to do one job.Connected components and the paths between them.glossary
- Role of the visual
- Exhibit
- Runnable by a visitor
- Pre-set examples only
- Date published
- 2025-03-27
- Year
- 2025
- Link status
- Live
- Built by (people)
- Jack Lindsey; Emmanuel Ameisen; Adam Pearce; Joshua Batson; et al.
- Organisation
- Anthropic
- Venue / published in
- Transformer Circuits Thread
- Sector
- Frontier lab
- Country / region
- US
- Open source
- Yes
- Method / technique
- Attribution graphs built on cross-layer transcoders; a 'replacement model' with interpretable nodes; causal intervention to validate edges
- Model(s) studied
- Claude 3.5 Haiku
- Access required
- Internal / proprietary
- What it visualises
- Step-by-step computational paths from prompt to response — which features caused which — across ten case studies including planning, multilingual reasoning, hallucination and jailbreaks
- Interaction affordances
- Step through each case study; expand and collapse graph nodes; trace individual paths; follow intervention results
- Pragmatic vs Basic science
- Both
- Reverse-eng vs Concept-based
- Reverse-engineering
- Observational vs Interventional
- Interventional
- Intended audience
- Researchers; Practitioners; General public
- Description (card)
- Ten case studies of Claude 3.5 Haiku's internal mechanisms rendered as explorable attribution graphs — showing it plans rhymes ahead, reasons across a shared multilingual concept space, and sometimes fabricates reasoning backwards from a hinted answer.
- Why it matters
- The clearest existing demonstration that a model's actual reasoning can diverge from its stated reasoning, made legible by graph.
- Visual / design notes
- Interactive graphs you expand case by case; the pedagogy is in the pacing. Companion methods paper carries the technical load so the biology paper can stay readable.
- Tags
- modality:languageText. Models that read and write words.Models that generate or process text.glossarymethod:attribution-graphA map of which internal pieces caused which, for one specific prompt.A causal graph of interpretable components explaining a single forward pass.glossarymethod:transcoderA variant that reconstructs what a layer passes onward, rather than what it received.A sparse dictionary trained to reproduce a layer's output rather than its input.glossarymethod:activation-patchingSwapping one internal value for another to see whether it mattered.Substituting one activation for another to measure its causal effect on the output.glossaryphenomenon:planningThe model deciding where a sentence is going before it writes it.Selecting a future output before generating the text that leads to it.glossaryphenomenon:hallucinationThe model stating something confidently when it has nothing to base it on.Producing confident content for which the model has no basis.glossaryphenomenon:jailbreakGetting a model to do something it was trained to refuse.Circumventing a model's trained refusals.glossaryphenomenon:unfaithful-reasoningThe explanation a model gives not matching the reasoning it actually did.Stated reasoning that does not correspond to the computation actually performed.glossary
- Citation
- Lindsey, J., Gurnee, W., Ameisen, E., et al., 2025. On the Biology of a Large Language Model. Transformer Circuits Thread.
- Related entries
- IA-005
- Confidence
- High
- Source of info
- Read the Anthropic research write-up in full, 2026-08-06
- Date added
- 2026-08-06
- Added by
- Claude
- Notes
- Paired with the Circuit Tracing methods paper. Tools open-sourced May 2025; Neuronpedia hosts a public graph explorer. Access classified Internal / proprietary: Claude 3.5 Haiku is not public, though the circuit-tracing method was open-sourced for use on open models.