A history made of documents, not a company
“Mechanistic interpretability” is a specific research programme with a datable birth, not a synonym for “understanding AI” in general. It is the attempt to reverse-engineer the algorithms a trained network has learned, at the level of individual units, weights and their compositions, checked against the network itself rather than against intuition. That programme has a lineage distinct from the broader project of AI alignment: alignment asks how to make a system pursue the goals its operators intend; mechanistic interpretability asks what a system is actually computing, a narrower and more mechanical question that alignment work sometimes draws on and sometimes ignores entirely.
This article is a history of that narrower field — who published what, when, under what institutional roof, and how the field’s own vocabulary and organisational footprint changed over roughly fifteen years. It is not an assessment of whether the field’s current methods work; that is a separate question with a separate literature. It stays close to dated primary sources: the papers themselves, the essays that proposed a research agenda, and the founding of the specific teams that turned the agenda into a line item on an organisational chart.
Seeing units before there was a field to name them
The idea that you could learn something by looking at what makes a single unit inside a trained network fire predates the modern field by more than a decade. Erhan, Bengio, Courville and Vincent’s 2009 technical report, presented that year at an ICML workshop on learning feature hierarchies, introduced what later became called activation maximization: rather than inspecting a unit’s weights directly, synthesize an input that drives the unit to fire as strongly as possible, and read the result as a picture of what the unit detects. The report compared several such techniques across stacked denoising autoencoders and deep belief networks and found that qualitatively legible interpretations of higher-layer units were achievable at the individual-unit level [1]. That is the founding move of the entire visualization lineage this field grew out of, five years before the architectures it would eventually be applied to existed in their modern form.
Zeiler and Fergus took a complementary approach in 2013, published at the European Conference on Computer Vision the following year: rather than synthesizing an input from scratch, their deconvolutional-network method traced a real image’s activations backward through a trained convolutional network to the pixels responsible for them. The paper is remembered for a methodological result as much as a visualization one — using the technique diagnostically, the authors identified architectural weaknesses in an existing ImageNet model and used the diagnosis to build a model that outperformed it [2]. Seeing inside the network was not just illustrative; it was already, in 2013, informing design decisions.
The same idea reached language models, in a different architecture entirely, in 2017. Radford, Jozefowicz and Sutskever trained a byte-level LSTM language model to predict the next character across a large corpus of Amazon reviews and discovered, without ever training for the task, that a single unit inside the network tracked sentiment: its activation alone was a strong, data-efficient sentiment classifier [3]. This result mattered historically for a specific reason. It showed that an individual unit corresponding to a human-legible concept was not a curiosity confined to vision networks; it happened in a recurrent language model trained on nothing but next-character prediction, years before circuits or attention heads had a formal literature of their own.
By this point activation maximization had moved from a single technical report to a formal, citable method. Olah, Mordvintsev and Schubert’s 2017 Distill article “Feature Visualization” laid out the optimization objective, the regularization tricks needed to keep synthesized images legible, and the tooling around the method, turning what had been a one-off report into something a lab could adopt as standard practice [4]. Two years later, Carter, Armstrong, Schubert, Johnson and Olah’s “Activation Atlas” pushed the method from single units to combinations of them, using feature inversion to visualize millions of activations from an image classifier and arrange them into an explorable map of what the network’s internal directions, not just its individual neurons, actually represented [5]. That paper is also an early institutional data point in its own right: its five authors split between OpenAI and Google, evidence that by 2019 more than one organisation had researchers whose job was specifically to look inside trained networks, years before any of them had a team with “interpretability” in its name on an org chart.
Circuits: the essay that gave the field its name and its bet
A decade of visualization work established that individual units and directions could be legible. It did not yet claim that the network’s computation as a whole could be decomposed and understood. That claim, stated as a deliberate research bet rather than a proven result, arrived in March 2020.
Olah, Cammarata, Schubert, Goh, Petrov and Carter’s Distill essay “Zoom In: An Introduction to Circuits” proposed three claims, each explicitly labelled speculative by its authors: that networks contain features, meaningful directions or units of computation; that features are wired into circuits by the weights connecting them; and that analogous features and circuits recur across different networks trained on different data and tasks, a property the authors called universality [6]. The essay’s worked examples — a curve detector built from earlier edge detectors, a dog-head detector assembled from curve and texture detectors, a car-detector neuron and a cat-face-detector neuron sharing machinery — were all from vision networks, and the paper’s title supplied the metaphor the whole subsequent field organised itself around: understanding a network by zooming in on its parts, the same way you understand a mechanism by taking it apart on a bench. This is the paper most often credited with naming “circuits” as a research programme rather than a scattered set of visualization techniques, and it is the reason “circuit” is now the field’s default noun for a checked, causal account of a specific piece of a network’s computation.
The programme needed a home and a target architecture to become more than an essay. Anthropic was founded in 2021 by a group of researchers that included Olah, and in December of that year its “Transformer Circuits Thread” published its founding document: Elhage, Nanda, Olsson and colleagues’ “A Mathematical Framework for Transformer Circuits” [7]. Where the 2020 essay had been programmatic and largely visual, this paper was formal, and specific to the architecture that by then dominated language modelling. It showed that a zero-layer transformer’s behaviour is readable directly off its weights as bigram statistics; that a one-layer attention-only model behaves as an ensemble of bigram and skip-trigram predictors; and that composition between heads only becomes possible at two layers or more, at which point qualitatively new algorithms appear. Its decomposition of each attention head into a query-key circuit, which determines where attention is directed, and an output-value circuit, which determines what gets written once attention lands there, became the piece of formal apparatus nearly every subsequent transformer-circuits paper builds on. The essay had named the bet; this paper gave it a mathematics and a publication venue of its own.
Induction heads: the field’s most thoroughly triangulated result
A framework is not yet a discovery. The field’s first landmark discovery, and still the case most often pointed to as an example of what a fully supported mechanistic claim looks like, followed three months later.
Olsson, Elhage, Nanda and colleagues’ “In-context Learning and Induction Heads” described a specific, simple mechanism: a pair of composed attention heads that finds an earlier occurrence of the current token in the context and copies whatever followed it, implementing the pattern “if A was followed by B earlier, predict B after A now” [9]. What made the result a landmark was not the mechanism’s simplicity but the volume and diversity of evidence assembled behind it. The authors reported six largely independent lines of support: induction heads form abruptly during a narrow phase change in training that coincides with a sharp jump in a model’s ability to use its context; architectural changes that shift when induction heads can form shift that same jump to match; directly ablating the heads substantially degrades in-context learning; the heads implement recognisable general behaviours including copying and simple translation; the mechanism generalises naturally from the small models where it was found to larger ones; and the relevant behaviours change continuously, not discontinuously, across that range of model sizes. The study behind those six lines covered thirty-four transformers tracked across training and more than fifty thousand individual attention-head ablations [9].
The paper’s authors were explicit that the weakest of the six strands was the extrapolation from small models, where the mechanism had actually been verified, to the large models where it was only inferred to hold. That caveat is itself part of why the result became the field’s reference case rather than merely its first widely cited one: it demonstrated what the evidentiary bar for a mechanistic claim ought to look like, including which parts of the claim remained open, rather than presenting a single striking correlation as settled. Four years and hundreds of subsequent interpretability papers later, no other single mechanism in the field’s literature has been supported by a comparably large and varied body of evidence, which is why induction heads remain the example most often taught first.
Superposition names the obstacle, sparse dictionaries become the answer
The circuits programme assumed that features are the network’s basic unit of computation, and that a network’s own neurons would be a reasonable place to look for them. That assumption ran into a specific, well-documented obstacle within two years of the framework paper, and the field’s dominant current tool exists because of how that obstacle got named.
Elhage, Hume, Olsson and colleagues’ 2022 “Toy Models of Superposition” studied small networks simple enough to be understood completely, where the ground-truth features were known because the researchers had constructed them. They found that when a network needs to represent more features than it has dimensions, and those features occur sparsely, the network can represent many more features than its dimensionality would naively allow by assigning them to non-orthogonal directions and accepting a controlled amount of interference between them — a phenomenon they termed superposition, with the network’s behaviour shifting through distinct geometric regimes as feature sparsity increased [8]. The practical consequence was uncomfortable for the circuits programme’s founding assumption: if a network’s features are packed into superposition, an individual neuron is not, in general, the right unit to call a feature. The obvious diagnostic that a real neuron often responds to several unrelated concepts — a behaviour the field calls polysemanticity — had already been documented in the original circuits work two years earlier; the superposition paper supplied the leading explanation for why it happens.
The tool the field converged on to work around this obstacle emerged in an unusually compressed window in late 2023. Cunningham, Ewart, Riggs, Huben and Sharkey’s paper, submitted in mid-September, showed that training a sparse, overcomplete autoencoder on a network’s activations and reading its learned dictionary elements as candidate features produced substantially more interpretable results than reading off individual neurons or principal components — and, importantly, supported finer causal identification of the components behind a specific behaviour than earlier methods had managed [10]. Three weeks later, working independently at Anthropic, Bricken, Templeton, Batson and colleagues published “Towards Monosemanticity: Decomposing Language Models With Dictionary Learning,” applying the same basic idea — a sparse autoencoder trained on a one-layer transformer’s activations — and reporting that a substantial fraction of the recovered features corresponded to clean, human-legible concepts where the underlying neurons had not [11]. Two separately run teams filing close variants of the same method within roughly three weeks of each other is itself a data point about the field at that moment: the superposition obstacle and the dictionary-learning fix it implied were, by the second half of 2023, obvious enough to more than one group of researchers working from the same public literature that neither needed to see the other’s paper first.
What followed was a scaling race, and it is documented closely enough to date to the month. In May 2024, Templeton and colleagues at Anthropic published “Scaling Monosemanticity,” training sparse autoencoders with up to thirty-four million features on the middle-layer residual stream of Claude 3 Sonnet, a production-scale model rather than a one-layer toy, and reporting that the recovered features were multilingual, generalised to images despite being trained on text, and included directions corresponding to safety-relevant concepts such as deception, sycophancy and power-seeking [12]. One month later, in June 2024, Gao and colleagues at OpenAI published “Scaling and evaluating sparse autoencoders,” introducing k-sparse autoencoders to control sparsity directly, establishing scaling laws relating autoencoder size and sparsity to reconstruction quality, and training a sixteen-million-latent dictionary on GPT-4 activations over forty billion tokens [13]. Two months after that, in August 2024, Lieberum, Rajamanoharan, Conmy and colleagues at Google DeepMind released “Gemma Scope,” an openly published suite of more than four hundred sparse autoencoders trained across nearly every layer and sub-layer of the Gemma 2 model family, released specifically so that researchers outside the small number of labs able to afford training their own large dictionaries could work with production-scale interpretable features directly [14]. Three laboratories, each running its own separate research programme, shipped a large-scale sparse-dictionary result within a four-month span of the same year. That is not evidence the method is correct — separate papers from within this same broader literature have since raised sharp questions about what a dictionary’s reconstruction objective actually certifies — but it is unambiguous evidence of which tool the field had, by the second half of 2024, converged on as its default.
From one team’s side project to a standard organisational line
The clearest way to see mechanistic interpretability’s growth as a field, distinct from its growth as a set of results, is to track where it sits on an organisational chart over time, and the record here is more legible than for most research areas because the same handful of people moved between the organisations involved.
Chris Olah’s own career traces the field’s earliest institutional path. He began doing feature-visualization work at Google Brain, moved to OpenAI in 2018 specifically to lead its interpretability research, and left in 2020 to co-found Anthropic the following year, where he now leads interpretability research as a dedicated function rather than a side project attached to a capabilities team [19]. Anthropic’s Interpretability team states its mission plainly on its own research pages: to discover and understand how large language models work internally, as a foundation for AI safety, on the grounds that it is very difficult to reason about the safety of a system without understanding what it is doing — and the team’s own account of its lineage explicitly traces its methods back through the circuit-based interpretability and scaling-laws work its members did before Anthropic existed [17].
That lineage runs through a second institution as well. Neel Nanda, a co-author on both the 2021 transformer-circuits framework and the 2022 induction-heads paper, worked at Anthropic as an interpretability researcher under Olah before moving to Google DeepMind, where he now leads that organisation’s dedicated mechanistic interpretability team — the group responsible for the Gemma Scope release described above [18]. By 2024, then, three of the organisations building the largest language models each had a specifically named interpretability research function, staffed in part by researchers who had trained under one another across those same three organisations — a small field, but one with real institutional continuity rather than isolated individuals repeatedly reinventing the same questions.
The field’s academic footprint grew in step with its industrial one. In July 2024, the first Mechanistic Interpretability Workshop was held alongside the International Conference on Machine Learning in Vienna, drawing ninety-three accepted papers to a single dedicated venue — a scale of submission that would not have been possible even two years earlier, when the entire field’s core bibliography numbered in the dozens rather than the hundreds [16]. That same year, Bereska and Gavves published the first field-wide review paper, “Mechanistic Interpretability for AI Safety: A Review,” a synthesis attempt that is itself a marker of maturity: a field does not get a review article until its literature has grown large enough that no single new paper can be read in the context of everything that came before it without one [15].
Predictions, with what would falsify them
These are forecasts, kept separate from the sourced history above. Horizon: 12 August 2031.
One. At least one additional frontier laboratory beyond Anthropic, OpenAI and Google DeepMind will stand up a named, dedicated mechanistic interpretability research function, rather than folding the work into a general safety or alignment team. Disconfirmed if, by 2031, interpretability research at labs outside these three remains an individual-researcher activity with no dedicated team or public team page.
Two. Sparse-dictionary methods will be superseded, partially or wholly, by a successor technique with a different training objective, because the field’s own 2023–2024 literature already documents specific cases where reconstruction-based dictionaries fail to identify the right causal units. Disconfirmed if, by 2031, the dominant published method for recovering features from a trained network is still a sparse autoencoder trained on the same reconstruction-plus-sparsity objective introduced in 2023.
Three. The induction-heads paper’s six-strand evidentiary structure — training dynamics, ablation, generalisation across architecture and scale — will be the template explicitly cited by later papers claiming a comparably well-supported mechanism, rather than being superseded by a faster, lower-evidence standard for publication. Disconfirmed if a majority of widely cited circuit-discovery papers published after 2028 support their central claim with only one or two of those six lines of evidence.
Four. The count of dedicated academic venues for mechanistic interpretability — workshops, tracks, or a standalone conference — will be higher in 2031 than the single ICML workshop that existed in 2024, tracking continued growth in submitted papers rather than a plateau. Disconfirmed if, by 2031, mechanistic interpretability work has been folded back into general interpretability or safety venues with no dedicated track of its own.
None of these predictions requires a capability breakthrough elsewhere in AI. They follow from the pattern already visible across roughly fifteen years of this specific history: a technique for looking inside one unit, then a name and a bet for looking at all of them together, then one thoroughly checked landmark case, then an obstacle to that bet’s founding assumption, then a fix three separate teams converged on within months of each other, then dedicated organisational and academic homes for the people doing the work.
What the history actually shows
Read as a sequence of dates rather than a single company’s product narrative, mechanistic interpretability is older than the term itself and more distributed across organisations than any single lab’s publications suggest. Activation maximization was demonstrated in 2009, applied diagnostically to convolutional networks in 2013, formalised as a method in 2017, and had already found an interpretable single unit inside a language model by the same year — all before “mechanistic interpretability” existed as a phrase. A 2020 essay proposed features, circuits and universality as a deliberately speculative bet and gave the resulting programme its name; a 2021 paper gave that bet a formal mathematics specific to transformers; a 2022 paper assembled, across thirty-four models and fifty thousand ablations, the field’s still-unmatched case study in what sufficient evidence for a mechanistic claim looks like. The same year identified the obstacle — superposition — that the field’s current default tool exists to work around, and two teams found that tool independently within three weeks of each other in 2023, before three separate laboratories each shipped a large-scale version of it within four months of one another in 2024. By then the field had dedicated teams at three major labs staffed in part by researchers who had trained under one another, a dedicated academic workshop drawing ninety-three papers in its first year, and its first field-wide review. That is not a technology’s marketing timeline. It is a genuinely young research field’s actual, checkable institutional history, written by the people who did the work, in public, as they did it.
