No entries match these filters.
No entries match these filters.
Interp Atlas is a visual index of ways to understand AI models.
It brings together explorable explanations, research tools and applied projects that make model behaviour and inner workings easier to explore.
It welcomes design experiments and speculative approaches, connecting artists, designers, makers and researchers around new ways to investigate and interpret AI.
AI is introducing a new modality of human–computer interaction. As models become more capable and agentic, the interface becomes an increasingly important site for interpreting, directing and evaluating activity unfolding across large computational systems.
Visual and spatial representations allow people to move between scales, organising patterns in model activity, isolating particular behaviours and connecting system processes to what appears at the interface. Dashboards, maps, timelines and interactive views shape what can be inspected, compared and questioned.
Interpretability research has produced visual and interactive tools for investigating specific questions about model behaviour. Their development creates a meeting point between research and design, where different forms of expertise shape instruments for understanding AI. In practice, these roles frequently overlap.
The safety of this interaction also depends on oversight during use. Constraints can shape an agent’s behaviour without making its activity fully observable. Interpretability interfaces may add a further layer of human judgement, allowing people to follow an agent’s actions, recognise when it departs from its intended task and decide when to intervene.
As AI becomes embedded across applications, these questions extend beyond specialist interpretability research to the designers, artists, researchers and strategists shaping how people encounter and work with AI. The Atlas offers this emerging field a shared reference for comparing approaches, developing a vocabulary and identifying unresolved problems in making AI systems more intelligible, steerable and open to intervention.
Interpretability is the effort to understand how AI models organise information, form representations and produce behaviour. It looks beyond outputs to the internal processes that shape a model’s response.
Modern AI models develop their internal organisation through training. The structures that emerge are distributed throughout the model and cannot be read as a conventional sequence of instructions. Interpretability develops concepts, methods and instruments for making this organisation available to observation, analysis and intervention.
The Atlas is for people developing ways to understand, inspect and interact with AI systems. It is intended for researchers and engineers working in interpretability, as well as designers, artists and strategists whose practices shape how model activity is represented and encountered.
It can serve as a research reference, a map of existing approaches or a point of entry into the field. Across the Atlas and Timeline views, readers can compare entries, trace precedents and identify questions that remain open. The glossary provides a path into unfamiliar terms. Technical expertise is not required, though it can deepen how the Atlas is read and used.
The Atlas welcomes submissions of existing work about interpretability with a substantial visual element. You can submit something you made or something you found. Research, tools, interfaces, educational projects and public-facing work are all considered, and each submission is reviewed before inclusion.
The Atlas is an index of published interpretability research, tools and interfaces. Each work in the index is an entry.
Each entry records the work's title, creators, publication date and source. It also describes what the work visualises, what it helps make visible and where it sits on the Atlas.
An interactive tool for interrogating attention heads, released as the research output in its own right. Its central finding is methodological: a head's behaviour on the full data distribution rarely matches what a narrow task suggests.
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.
An interactive tool that runs a live GPT-2 instance in your browser and visualises every stage of the forward pass — embeddings, attention, MLP, output probabilities — updating as you type your own text.
A dashboard that sits beside a chatbot and shows, in real time, what the model has internally inferred about the user's age, gender, education and socioeconomic status — and lets the user edit those inferences and watch the responses change.
For 24 hours Anthropic put a version of Claude 3 Sonnet online with its 'Golden Gate Bridge' feature clamped to roughly ten times maximum activation, so anyone could talk to a model that could not stop mentioning the bridge.
First demonstration that sparse autoencoders scale from toy models to a frontier production model, published as a browsable index of millions of features — including the Golden Gate Bridge feature that later became a public demo.
The paper that made sparse autoencoders the field's dominant method, published with a browsable interface over every extracted feature so readers could check the monosemanticity claim themselves rather than take the authors' word for it.
Visualises a joint embedding of the query and key vectors a transformer uses to compute attention, which makes it possible to see global patterns across many input sequences rather than one prompt at a time.
The field's central public platform. It hosts feature dashboards, attribution graphs, steering and demos across dozens of open models — including the official interactive releases for Anthropic's Jacobian Lens, Natural Language Autoencoders, Assistant Axis and Circuit Tracer, and DeepMind's Gemma Scope.
Renders millions of activations from an image classifier as feature-inversion images laid out on a single navigable map, so you can pan across the concepts a network has learned the way you would read an atlas.
The founding text for interactive interpretability. Argues that interpretability techniques studied in isolation are far weaker than the interfaces you get by composing them, and demonstrates this with a set of live, hoverable interfaces over an image classifier.
A university or academic research group.
The level of model access a person would need to reproduce the work.
Requires gated or paid inference access, but not weights.
Preserved in static or historical form and no longer maintained.
Work made primarily for aesthetic or critical reception rather than scientific reporting.
The form the work takes, independent of what it contains or argues.
A practitioner working primarily within an art context.
Models that process sound or speech.
A searchable collection of results presented for inspection.
The named individual contributors, in the order the source credits them.
A full bibliographic reference in a consistent style.
Models whose subject matter is programming languages.
A link to that source code.
How well evidenced this row is, judged by what the person writing it actually read.
A value that may only be chosen from the list on the Vocabularies sheet, never typed freely.
Where the producing institution is based.
A live readout of a system's internal state, displayed alongside that system in use.
When the row was created and by whom, human or model.
The date the artefact first became publicly available, in ISO form where known.
The URL no longer resolves to the artefact.
A model made publicly usable in a deliberately modified state, as a demonstration.
One or two sentences describing the artefact to a stranger, written to be published rather than as a note to self.
The visual exists so that a reader can inspect and audit a claim which is itself stated in prose.
The artefact is undergone rather than read, and its value lies in what it does to the visitor.
The visual communicates a result established elsewhere, serving teaching rather than discovery.
The result is constituted by the visual, such that removing the images would remove the contribution.
A field written in prose, with no fixed list of permitted values.
An organisation training state-of-the-art general-purpose models.
Joint authorship spanning both.
Not tied to any particular model or data type.
Non-specialists, with no assumed technical background.
Models of genetic sequence.
The primary source was read in full by whoever wrote the row.
A permanent identifier for one artefact, never reused, so that a retired entry's number can never resolve to a different record.
Models that produce images, such as diffusion models.
A person working outside any institution.
Others can operate it on their own inputs or models to produce findings that did not previously exist.
Who the artefact was designed to be understood by.
What a person can actually do with the artefact, verb by verb.
A written argument in which the figures respond to the reader.
A standalone application whose purpose is to let a user investigate a model.
Requires access available only within the producing organisation.
Models that generate or process text.
Whether the recorded URL currently resolves to the artefact.
The URL resolves and the artefact functions as described.
The row rests on inference and has not been checked against a source.
The thing rendered on screen. Not what the work is about, which is the phenomenon tag, but what you are actually looking at.
The image filenames belonging to this entry, semicolon separated.
The row rests on an abstract, documentation, or a reliable secondary source.
The interpretability technique the work employs.
The interpretability technique or techniques the artefact rests on.
Deliberately designed to serve more than one of the above.
The kind of model or data the work concerns.
The specific neural networks the work examines.
Models that process more than one modality jointly.
The artefact's own title, as given by its makers.
No source is published.
The artefact was operable once but is not now.
An organisation constituted for public benefit rather than profit.
No graphic form. The artefact qualifies some other way.
Requires no model access at all; the outputs are fixed.
Code published to be run and adapted by other researchers.
Anything qualifying the row, including unresolved questions and flagged boundary cases.
Nothing is rendered. The artefact is encountered rather than viewed.
Whether the work examines the model unchanged or alters it and observes the consequence.
A visitor can supply their own prompt, image or model and see fresh output.
Whether the artefact's own source code is publicly available.
Reproducible using publicly downloadable model weights.
The institution that produced the work.
A form not covered above; propose a new term rather than reusing this indefinitely.
Some components are released and others deliberately withheld.
Reachable only behind payment or registration.
The property or behaviour of the model that the work is about.
Infrastructure hosting many artefacts, models or datasets originating from multiple parties.
Regulators, governments and civil society organisations.
People deploying or auditing models professionally.
Whether the work targets one specific behaviour or the model's general structure.
The visuals are fixed; interaction is limited to navigation such as panning or zooming.
A visitor can explore interactively, but only within cases the authors selected in advance.
Models of protein sequence or structure.
Every column in the Ledger and every permitted value is defined here in one sentence, so that two people filling in the same row reach the same answer.
The IDs of other Atlas entries with a documented relationship to this one.
Specialists in interpretability or machine learning.
Whether the work decomposes the network and then names the parts, or proposes concepts and then locates them.
Agents trained by reinforcement learning to act within an environment.
The function the visual element performs within the research, which is the central judgement in each row.
Whether, and how far, a person arriving today can operate the artefact themselves.
Representations bearing on harm, deception, bias or misuse.
The kind of institution that produced the work, recorded so the composition of the field is visible.
The lowercase hyphenated form of the entry's name, used as its web address.
What was read to write this row, and on what date.
A commercial company whose product is interpretability or model tooling.
Fixed images within a conventional publication.
Learners being taught the material for the first time.
Models over structured, columnar data.
Faceted keywords in facet:value form, drawn only from the controlled facet lists.
An image representing the entry.
The role has not been determined.
Not determined, often because the page could not be loaded.
The basis for the row was not recorded.
Not yet determined.
Not yet determined.
The canonical public address of the artefact itself, rather than of a paper describing it.
Where the work appeared, which is often not the same as who produced it.
Time-based linear media.
Models that process static images.
Observations on the artefact's visual and interaction design, including what makes it work or fail.
The shape the picture takes: what you would call the graphic if describing it to someone who could not see it.
The object being made visible, stated in plain terms a non-specialist could follow.
One line stating the artefact's significance in terms a reader could dispute.
A tag is a filter, not an opinion; claims about significance belong in Why it matters, where a reader can dispute them.
Organisation, artefact type, sector and open source are already columns; tagging them again would create two sources of truth that drift apart.
The publication year held separately, so that entries sort chronologically without parsing dates.
The artefact's own source code is publicly available under a licence.
Confidence and source information are part of every entry. Unknown information stays unknown, and a plausible inference is not treated as a verified fact.
Public submissions and discovered candidates are reviewed outside the Ledger. Nothing becomes a published entry until it has been approved, researched from primary sources and validated against the controlled vocabularies.
The structured dataset is not yet published as a public download. The Atlas currently publishes the validated entries and their evidence through the website.