The Symmetry You Choose

The choice of what should remain invariant can matter more than the model that exploits it.

On the Highest-Leverage Skill in AI Engineering

Core argument. AI systems are increasingly powerful at exploiting known regularities: given a pattern, they compress, generalise, and scale it with superhuman efficiency. A symmetry is a property that remains unchanged (invariant) when conditions vary: rotate an isolated molecule and its energy stays the same; shift a pattern across an image and the edge detector still fires. As that exploitation power grows, the highest-leverage human act shifts upstream: deciding which invariances the domain respects, at what level of abstraction, and which departures from them are signal rather than noise. This essay argues that selecting which symmetry to impose, at which level of abstraction, and where to let it break, is among the highest-leverage acts in AI engineering. Most architectures and data pipelines encode an implicit claim about which invariances the problem domain respects. That claim is often more consequential than the model that acts on it. Choose the right symmetry at the right level of abstraction, and the problem collapses in complexity before engineering begins. Choose it at the wrong level, and the system produces output that is internally consistent and materially wrong. When the goal is to understand the domain's generating structure, the most informative signal often concentrates where the assumed symmetry fails.

I. The Right Symmetry at the Wrong Rung

An isolated molecule does not care how you orient it in space. Rotate it, translate it, reflect it: its energy is the same. Researchers recognised this invariance and encoded it directly into neural network architectures. SE(3)-equivariant graph neural networks (Satorras et al., 2021; Cohen & Welling, 2016) build rotational and translational symmetry into their layers by construction, gaining efficiency that unconstrained networks would need orders of magnitude more data to approximate. The compression is substantial: fewer parameters, better generalisation, dramatically lower sample complexity.

Now place that same molecule at a cell membrane. The molecule now has a preferred orientation. The rotational symmetry that was exact in isolation no longer holds: the context imposes a direction. A model built for isolated molecules, now applied at the membrane, will miss the directional physics that determines whether the molecule binds. The symmetry is real. It applies in isolation. At the membrane, the level of abstraction has changed.

A symmetry holds within a particular effective description; move to a different level of abstraction, and the governing structure may change entirely. Recognising when this has happened is the problem this essay addresses. In a previous essay, "Two Compressions", I drew a line between extracting a pattern from data and proposing the generating mechanism that produces the pattern as a necessary consequence. This essay restates that distinction in terms of symmetry and develops it further. Exploiting a known invariance to collapse redundancy is what AI systems do with increasing power. Models can increasingly assist in exploring which invariances might hold, but the epistemic burden of committing to the right frame at the right abstraction level in messy real domains has not yet shifted away from human judgment.

It has become common in the AI community to assert that "human judgment is still needed." That assertion vector is directionally correct but underspecified: it does not tell engineers what the judgment consists of or how to develop it as a skill.

This essay uses symmetry as a lens to give that claim specific, actionable content: the judgment is about which invariances to impose, at which level of abstraction, and where to let them break. Symmetry provides a productive lens through which both operations become visible (a causal or information-theoretic framing could capture a related distinction; this essay uses symmetry because it connects directly to architectural choices in practice). To ground the argument, the sections that follow draw on examples from my own recent published work across multiple domains, as well as a worked case in document intelligence drawn from experience building these types of systems.

A note on scope: this essay uses "symmetry" on a ladder. At lower rungs, the term denotes mathematical invariance under a transformation group, its standard meaning in physics and geometric deep learning. At higher rungs, it denotes a preserved relation of comparability or consistency across representations: semantic comparability across accounting metrics, or normative consistency between narrative and numbers. The common question at every rung is the same: what structure is being held fixed, and where does that assumption fail? The concept stretches as the ladder ascends, and the stretching is deliberate; the decreasing formality tracks the increasing difficulty of the selection.

The distinction between exploitation and selection is practical. Every architecture choice, every data augmentation strategy, every equivariance constraint encodes an implicit claim about which invariances hold in the problem domain. Getting the claim right can reduce the effective complexity of a problem by orders of magnitude before a single gradient step is taken. Getting it wrong produces systems that are fluent, internally consistent, and subtly misaligned with the generating structure of the domain. The difference lies in the quality of the selection.

A reasonable question at this point: why should an AI engineer think in terms of symmetries at all? The answer is that you already do, whether you use the word or not. Every time you share convolutional weights across spatial positions, you are exploiting translational equivariance. Every time you normalise inputs, you are deciding which variations the system should treat as non-essential. Every time you augment training data with random rotations, you are asserting approximate rotational symmetry. Symmetry thinking makes explicit what engineering already does implicitly. And making it explicit gives you three things: (i) a language for articulating what your architecture assumes, (ii) a method for testing those assumptions, and, as we will see, (iii) a diagnostic for understanding why your system fails where it does.

II. Exploitation and Selection

A system that is invariant under some transformation contains built-in redundancy. You do not need to describe every point or direction separately. You describe one representative piece, one fundamental domain, and the symmetry operation fills in the rest. The gravitational field of a perfect sphere depends only on the radial distance r, not on direction. The equation F(r) = -GMm/r² is one function of one variable. Without spherical symmetry, you would need a function of three coordinates for every point in space. The symmetry collapses that redundancy into a single degree of freedom plus the knowledge of the rotation group. This is why physicists can solve the gravity of a star with a single radial equation. The symmetry is doing the compression.

The idea that symmetry is a compression is well understood in physics; Noether's theorem (1918) formalised the deep link between continuous symmetries and conserved quantities. Whether symmetry fully explains the tractability of physics, or serves as an indispensable taxonomy that points toward deeper structure without itself providing the final explanation, remains open even at the foundations. This essay requires only the weaker claim. Less explicit, and more consequential for engineering, is that identifying the symmetry and exploiting the symmetry are different operations, different in kind.

Consider what Newton actually did. Kepler's first law says orbits are ellipses with the sun at one focus. This is a geometric regularity in the observable data, a law-like pattern in the trajectories. Newton's achievement was to propose a generating structure, the inverse-square force law, from which the observed regularities follow as necessary consequences. The rotational invariance of the gravitational field around a point mass is a property of that structure: something Newton posited, not observed. And from that posit, Kepler's ellipses, the equal-area law, and the period relation all emerge as theorems.

Two acts, then:

Exploiting a symmetry means: given that the system is invariant (or equivariant) under some transformation, use that structure to compress the description. One radial function instead of a three-dimensional field. A convolutional filter whose weights are shared across every spatial position. A permutation-invariant aggregation over graph nodes. This is powerful, and it is what AI systems increasingly do. It is symmetry exploitation.

Selecting a symmetry means: proposing that this system, at this level of abstraction, is invariant under this particular transformation and not others. Proposing that the gravitational field is spherically symmetric. Proposing that the relevant state space is phase space (position and momentum), not configuration space (position alone). Proposing that the force law is inverse-square rather than inverse-cube. This is the frame-setting act: it determines what the model takes to be the relevant entities and relations, and which compressions become available. It is symmetry selection.

The first act lives within a fixed mathematical framework. The second act chooses the framework. The distinction is more visible at the creative end of selection: proposing a new invariance for an unfamiliar domain is qualitatively different from applying a known convolution. Routine selection (choosing a CNN for a standard image task) lies closer to exploitation on this spectrum. The argument that follows concerns the generative(creative) end, where the framework is proposed rather than retrieved. The intent is to (i) highlight what humans still have to bring to the table, and (ii) the types of skills that humans ought to develop to leverage AI transformatively.

III. The Rung Problem

The gap between selecting a symmetry and exploiting one becomes practically urgent when we recognise that symmetries are level-dependent. They are properties of effective descriptions at particular scales of abstraction, and they break, often usefully (generating information), when you move between levels.

Physics illustrates this directly. The fundamental laws of electromagnetism are invariant under continuous rotation: no direction is preferred. But when atoms arrange themselves into a crystal lattice, the continuous rotational symmetry breaks into a discrete subgroup. A snowflake is invariant under 60-degree rotations, not arbitrary ones. The underlying physics has not changed. The effective description at the emergent scale has different symmetries from the effective description at the fundamental scale.

This is how complex systems work. Effective field theories in physics formalise it as a ladder. At each rung (atoms, molecules, materials, organisms) you write a new effective theory whose symmetries are those that survive at that scale. Rotational symmetry holds for spherical atoms; it breaks in molecules with specific geometries; it reappears approximately for large aggregates. The symmetry regime changes at every rung because the relevant invariances change.

This is the rung problem, and it is a central challenge for AI systems that encode structural priors. The question is always: at which level of abstraction does the symmetry hold, and where does it break?

A convolutional neural network hard-codes translational equivariance. This is the right structural prior for early feature extraction in images: an edge detector learned at one position should work at every position. But object detection requires breaking that equivariance at higher layers, because the task involves absolute spatial position (this object is in the upper-left corner of the image). The early layers exploit the symmetry. The later layers break it. The decision about where the break occurs, at which rung the translational equivariance ceases to be the right effective description, is an engineering judgment.

The molecule example from this essay's opening is another instance. An SE(3)-equivariant graph neural network for molecular property prediction hard-codes rotational and translational invariance: the right symmetry for an isolated molecule. But a molecule in a crystal has preferred orientations. A protein at a membrane interface has a preferred direction. The symmetry that was exactly right at one rung (isolated molecule) is wrong at another (molecule in context). Knowing when to impose the symmetry and when to let it break is a judgment about which effective theory governs the regime you care about.

The rung problem recurs across applied domains. Here are a few examples from projects that I contributed to recently to illustrate. In materials science, traditional alloy design assumed that corrosion behaviour is governed by a single principal element; multi-principal element alloys break this invariance, as all constituents interact non-linearly (Ghorbani et al., 2025). In gemstone assessment, device-robust color grading requires normalising camera differences into a perceptually uniform color space, imposing device invariance, while origin verification requires matching spectral signatures through dynamic time warping to accommodate non-linear spectral shifts (Roy et al., 2025): same specimen, different symmetries at different rungs. In multi-agent coordination, the standard paradigm assumes agents share a common goal; decentralized settings where agents pursue individual objectives require selective, goal-aligned communication, because the common-goal uniformity holds for some agent pairs and fails for others (Du et al., 2025).

The rung problem is why the symmetry-as-compression idea, taken naively, can mislead. It suggests that the engineer's task is to identify symmetries in the data and encode them. The deeper task is to decompose what the prevailing approach treats as a monolith: to find the joints where an assumed uniformity gives way to structure at a different level. Which rung of the abstraction ladder does the problem live on? Which symmetries hold at that rung, where do they break, and what does the breaking pattern reveal about the generating structure? That is a symmetry-selection problem: the engineer must propose the structural frame within which exploitation then operates.

IV. Every Architecture Is a Symmetry Claim

Architectural decisions that encode structural priors are testable claims about which invariances hold. Making these claims explicit changes how engineers can reason about design.

Hard-coded equivariance (group-equivariant CNNs, E(n)-equivariant networks, permutation-equivariant architectures) encodes the claimed symmetry directly into the network layers. The payoff when the claim is right is real: fewer parameters, better generalisation, orders-of-magnitude reduction in sample complexity. Even within a single mechanism, the symmetry choice matters: standard multi-head attention assumes head isolation (each head computes independently), and relaxing that assumption through cross-head mixing can reduce the parameters needed for multi-step composition from linear to sub-linear in the number of reasoning steps (Duvvuri et al., 2026). A related assumption, that uniformly splitting embeddings across heads preserves sufficient semantic content per head, breaks when temporal coherence matters: in video frame prediction, the splitting dilutes the semantic signal each head can access, and restoring the full embedding per head recovers the lost structure (Nguyen, Thudumu, Vasa, Mouzakis et al., 2025). The cost when the claim is wrong is equally real: the model enforces an invariance the problem does not have, or misses one it does.

Softer approaches encode weaker claims. Data augmentation (random rotations, flips, crops) says "the data is approximately invariant, and I am uncertain about the exact symmetry group". Deliberate symmetry breaking (positional encodings, anchor-based detection heads, asymmetric initialisations) says "the symmetry holds up to this point, and here it should stop".

These claims extend beyond architecture into training procedures. The reversal curse (Song, Han & Goodman, 2026) is a vivid case: models trained on "A is B" systematically fail to infer "B is A"; a bidirectional equivalence humans find trivial. The autoregressive training objective induces a directional asymmetry that breaks the logical symmetry of equivalence. The same survey documents a broader pattern: minor semantics-preserving perturbations (reordering options, renaming variables) cause large performance shifts, revealing that invariances the task demands are absent from the model. Each failure pattern is a symmetry map, diagnosing where the design's implicit claims diverge from the domain's actual structure.

The symmetry claims extend further still: even the engineering process has structural invariances. That error is inherent in all models, that training distributions diverge from deployment, that benchmarks expire as contexts shift, that data and code and human processes entangle; these are constraints that persist across model types and domains, properties of the practice rather than of any particular system (Barnett, Vasa et al., 2026). Engineering decisions that treat these constraints as incidental rather than structural produce systems that may be locally optimised but systemically fragile.

The practical heuristic: early in the design stages, state the symmetry claim explicitly. What invariance does this architecture encode (or ought to)? At which rung does that invariance hold? Where should it break? Both directions of error are costly. A network trained with rotational equivariance on a problem that lacks rotational symmetry will enforce the wrong structure. An AI model designed without it on a problem that has it will waste capacity rediscovering what could have been given for free.

This symmetry selection is not reliably learnable from the data in the general case. The data shows you the output of a process. The symmetries of the generating mechanism are not uniquely recoverable from the output alone: the same observational data can be consistent with multiple symmetry groups. Observational data tells you which patterns hold. It does not tell you which invariance principle produces those patterns as necessary consequences, nor at which rung of the abstraction ladder that invariance operates.

V. Where Invariance Breaks, Information Lives

The instinct most engineers bring to the problem is that symmetry is good (it compresses, it regularises, it reduces parameters) and symmetry breaking is a failure to manage. The lesson from physics runs deeper: asymmetry is itself constitutive of structure. Phase transitions, magnetisation, crystal formation, biological chirality: all arise because a symmetry failed to hold.

In AI engineering, the analogous insight is that among the most consequential design decisions is where to stop enforcing an invariance. In a transformer, the attention mechanism is permutation-symmetric: it treats input tokens as an unordered set. Positional encodings break this symmetry, and the breaking is what allows the model to distinguish "the dog bit the man" from "the man bit the dog." The symmetry was mathematically convenient. The breaking was semantically necessary.

The deeper principle: if an assumed invariance holds perfectly, it compresses the problem efficiently and contributes little discriminative signal within that frame. The invariance is "free structure", redundancy you can exploit. But where the invariance fails, something is happening that the symmetry cannot explain. The failure is a signal. It points to generating structure that the assumed symmetry does not capture.

This means a model's error patterns are a map of where its implicit symmetry claims are wrong. A financial analysis model that consistently misjudges companies with non-standard revenue recognition policies is telling you, through the structure of its errors, that the assumed semantic comparability across the peer group breaks at a specific point. A model that hallucinates financial figures more often for non-standard table layouts is revealing that the assumed structural invariance across report formats has an exception. The pattern of failure has a structure, and the structure of those errors can be diagnostic.

This allows us to rethink evaluation design. The standard approach measures aggregate performance: accuracy, F1, perplexity. These are symmetry-blind metrics. A symmetry-aware evaluation tests the claims directly. If your architecture claims translational equivariance, measure whether performance is actually equivariant under translation. If your financial analysis pipeline claims that revenue metrics are comparable across a peer group, test specifically on the companies where revenue recognition policies differ. The evaluation becomes a symmetry test: where do the model's implicit invariance claims break, and what does the breaking pattern reveal?

This differs from metamorphic testing, where you specify a known invariance in advance and check whether the system preserves it. Metamorphic testing evaluates within a frame. The symmetry-map idea uses the structure of the errors to discover the frame: which invariances the domain actually has, at which rung they operate, and where they break.

The picture this suggests is richer than a clean division between human selection and machine exploitation. AI can powerfully assist the exploration of which symmetries might hold. Consider an agent-harness loop designed to probe the domain rather than solve the downstream task. The AI systematically tests candidate invariances: does performance hold across these subgroups? Does the error pattern shift when this variable changes? Are these two representations actually comparable, or does the comparison break for identifiable reasons? The AI traverses the space of possible symmetries at scale, mapping which ones hold, which ones break, and where the breaks cluster. The human examines the map and judges which invariances reflect actual structure in the generating mechanism and which are artefacts.

The division of labour is a loop: the human proposes a symmetry claim, the AI tests it at scale, the error patterns reveal where the claim breaks, the human revises the claim or shifts to a different rung, and the cycle continues. Each iteration deepens the human's understanding of the domain's generating structure. The AI's contribution is cartography. The human's is commitment, as well as context/domain knowledge that unavailable to the models natively.

VI. The Ladder: From Reports to Investment Thesis

The preceding sections have been abstract. To make the argument concrete, consider a problem many AI engineering teams now face: building a system that reviews business documents (annual reports, earnings transcripts, stock exchange filings) to support an investment decision. The substrate of intelligence is a frontier model accessed via API. The task is ordinary. It is the kind of problem where AI is being seriously considered to deliver value today. It also is helpful to see where the symmetry selections hide.

The workflow from raw document to investment-relevant output involves distinct stages. At each stage, the engineer makes a symmetry claim: an assertion about which invariance holds. As we ascend the stages, the invariances weaken, the stakes rise, and the human contribution increases. The stages form a ladder, and the ladder is the argument.

Systems like docETL (Shankar et al., 2024) make this pipeline structure explicit and programmable. The engineer declares a sequence of operations over documents: map, reduce, resolve, filter. Each operation encodes an assumption about what structure exists and what invariances hold. The model executes within each operation; the pipeline design determines what the model can see and how it compresses. docETL also provides an optimizer that rewrites pipelines, decomposing coarse operations into finer-grained steps when a single operation tries to span too many assumptions at once. In the terms of this essay, the optimizer behaves like a mechanical diagnosis of the rung problem: it discovers that one invariance claim was too broad for the data and splits it into claims at different levels. The pipeline is the symmetry selection. The model's work within each step is the exploitation.

Structural invariance across reports

The first symmetry claim is that all annual reports, despite surface differences in formatting, layout, and prose style, share an invariant deep structure. There is always a set of financial statements (income statement, balance sheet, cash flow). There is always a narrative framing the numbers. There are always risk disclosures. The surface varies; the anatomy is invariant.

This is an exploitable symmetry. The engineer designs a chunking and extraction pipeline that maps every report, regardless of format, into a canonical representation: tables parsed into structured data, narrative sections tagged by function, footnotes linked to the claims they qualify. The model executes this extraction reliably. This is symmetry exploitation: collapsing surface variation across thousands of report formats into a single canonical structure.

The symmetry claim here is strong and the exploitation is reliable. But the claim itself is the engineer's. Why these categories and not others? Why separate management discussion from the chairman's letter? A different decomposition (organised around cash-flow implications rather than document sections) would produce a different canonical form and a different downstream analysis. The choice of structural symmetry is prior to the extraction.

Semantic invariance within accounting

The second symmetry claim is that the same financial metrics, computed the same way, mean comparable things across companies within an industry. A debt-to-equity ratio of 2.5 at a utility and 2.5 at a technology company carry different signals. But within a sector, the claim is that the metric is approximately invariant: it measures the same underlying economic reality regardless of which company produced it.

This invariance enables the model to apply a uniform analytical framework, a fixed set of ratios, trend analyses, and anomaly detectors, across an entire peer group. The model handles this capably: computing ratios, flagging deviations, identifying trends, producing structured comparisons. Symmetry exploitation again.

But here the rung problem appears. Consider a toy case. Two SaaS companies each report $100M in annual revenue and 25% year-over-year growth (both under an accounting standard). Company A recognises revenue ratably over multi-year contracts; Company B recognises a significant portion upfront because it bundles professional services into the initial sale. Company A's $100M reflects a backlog of committed future revenue. Company B's $100M front-loads income that will not recur at the same rate. A model comparing their revenue growth rates will produce a table showing two companies growing at 25%, ranked identically. The output is likely fluent, numerically correct, and misleading: the underlying cash-flow dynamics are structurally different.

The numbers are reported under the same label ("revenue") but are not related by any simple transformation. The semantic symmetry the engineer assumed, that "revenue" means the same thing across the peer group, is broken by the specifics of each company's accounting policy. The model will not flag this on its own unless the engineer has built in the instruction to check for it, which requires knowing in advance that revenue recognition is a point where the assumed symmetry breaks. That knowledge is knowledge of the domain's generating structure. It is symmetry-selection knowledge.

The invariance between words and numbers

The third symmetry claim is subtle and powerful. In a truthfully reported business, the narrative and the financial data are two representations of the same underlying economic reality. They should be invariant under the transformation that maps one to the other. The prose description of the business trajectory and the financial trajectory should, modulo framing and emphasis, tell the same story.

The model can be directed to test this invariance. Given the extracted financial trends and the extracted narrative claims, it identifies points of inconsistency: where management claims improvement while the numbers show deterioration, where the risk section omits a risk the data makes obvious. The model can be good at this, because it can hold the full context of a long document and perform the comparison at a level of thoroughness that a human analyst under time pressure often cannot match.

But the information lives in the break. Suppose a company's CEO letter describes "accelerating momentum" and "expanding market position" while operating cash flow has declined for three consecutive years and the customer acquisition cost (CAC) has risen 40%. The narrative and the numbers are telling different stories. That discrepancy is the signal. The symmetry between narrative and numbers was a diagnostic tool, and the breaking was the diagnosis.

The engineer chose this frame: that narrative-financial consistency is the right invariance to test, and that inconsistency is the informative signal. An alternative frame might treat the narrative as strategic communication rather than factual reporting, looking for invariance in the rhetorical strategy across reports (does management always downplay the same type of risk?) rather than invariance between narrative and numbers within a single report. The choice of frame determines what the system detects and what it misses.

The symmetry that does not exist

The final stage is the one that matters most and the one where the framework reaches its limit: synthesising the analysis into an investment thesis.

Here the relevant invariances are causal rather than statistical: complex, context-dependent, and not capturable by the pattern-matching symmetries the model exploits at lower rungs. An investment thesis is a claim about the future generating structure of value. It asserts that certain causal mechanisms, competitive dynamics, technological shifts, regulatory changes, management quality, will produce certain outcomes. Two companies with identical current financials can have radically different investment profiles because the generating mechanisms of their future value are different. There is no transformation under which the investment thesis for Company A and the investment thesis for Company B are equivalent, because the thesis addresses the generating structure: the causal mechanisms that will produce future states from this point forward.

The model, instructed to produce an investment thesis, will generate one that is coherent, well-structured, and draws appropriately on the preceding analysis. It will pattern-match against the thousands of investment analyses in its training data. What it will not do is determine whether the generating structure of value for this particular company, in this particular market, at this particular moment, is captured by the analytical frame the workflow has imposed.

The most consequential investment decisions often hinge on this determination. The analyst who recognised in 2007 that the assumed symmetry between mortgage-backed securities and traditional fixed-income instruments had broken, that "same credit rating implies same risk profile" no longer held because the generating mechanism of the underlying assets had changed, was doing something qualitatively different: proposing that the analytical frame of the standard analysis was wrong. The state variables the industry was using (credit ratings, historical default rates) did not capture the relevant causal structure (correlated default risk in a new class of synthetic instruments). That was symmetry selection where it matters most: a recognition that the invariance the entire industry was exploiting had broken at a rung nobody was watching.

The pattern across the ladder

The trajectory across these four stages, in this domain, is self-similar. At the base, the symmetry is strong and well-characterised; the model's exploitation is reliable. At the second rung, the symmetry is approximate and breaks in specific, knowable ways; the engineer needs domain knowledge to anticipate the breaks. At the third rung, the symmetry is a normative claim (reports should be consistent) applied to an empirical reality (they often are not), and the signal is in the breaking. At the top, the invariances are causal rather than statistical; proposing the generating structure is the decision itself.

As you ascend, the symmetries weaken, become more contingent, and depend more on the specific causal structure of the domain. The exploitation (applying the assumed invariance at each stage) is what the model does. The selection (choosing the invariance, anticipating where it breaks, recognising when the frame itself is wrong) is what the engineer contributes. The leverage of the human contribution increases as you climb, because the consequences of selecting the wrong symmetry grow at each rung. Other domains may distribute this leverage differently across their rungs; the claim is that the pattern of increasing human contribution at higher levels of interpretive complexity is general, even if the specific rungs vary.

Current AI tooling and engineering methods do not yet surface this rung-dependent structure systematically. Developing evaluation frameworks that do would be a significant step toward making "human judgment is still needed" an engineering discipline rather than a hand-wave.

VII. The Frame Is the Decision

The symmetry you choose determines the compression you get. The rung you choose determines whether the symmetry applies. The breaks you anticipate determine whether the system produces insight or fluent nonsense.

Each correct symmetry selection makes every downstream operation more efficient. If you identify that the problem is rotationally invariant, every training example effectively becomes an infinite family of examples. If you identify the right rung, the model only needs to learn the effective theory at that scale, not the full multi-scale dynamics. If you identify where the symmetry breaks, the model concentrates capacity on the genuine degrees of freedom rather than wasting it on collapsed redundancies. These savings multiply. A correct structural frame, the right symmetry at the right rung with the right breaking pattern, can reduce the effective complexity of a problem by orders of magnitude before training begins.

This is why symmetry selection is among the highest-leverage points in the engineering process. As exploitation power grows, the return on getting the frame right grows with it: a more powerful model extracts more value from a correct structural prior and produces more convincing wrong answers from an incorrect one.

This essay and my previous one (Two Compressions) now converge on the same point, approached from different directions. AI systems exploit known invariances in observable data with superhuman fidelity and speed. They collapse redundancy, find patterns, and compose existing abstractions into sophisticated structures. This is symmetry exploitation, and the room beneath its ceiling is immense.

AI systems increasingly assist in the exploration of which symmetries might hold, as the previous section described. What has not yet shifted is the epistemic burden of commitment: the justified judgment that this invariance, at this rung, reflects the generating structure of the domain. Models can propose candidate symmetries, test them at scale, and map where they break. The judgment that a particular invariance is critical, that it reflects the generating mechanism and should govern the system's architecture, still requires a form of accountability that current systems do not bear; and our current development methods have not made a priority to address. Exploration is being shared. The commitment has not been, yet.

The engineer who understands this occupies a specific role: to stand at the point where the structural frame is chosen, where the rung is selected, where the breaks are anticipated. Everything downstream from that decision is increasingly automatable. The decision itself is where judgment currently lives, and where it will live until models can bear the epistemic burden of being wrong about the frame.

References

Barnett, S., Pasquini, A., Kurniawan, S., Sivasothy, S., Hill, R., & Vasa, R. (2026). "The Physics of AI." IEEE/ACM 5th International Conference on AI Engineering (CAIN '26).

Cohen, T.S. & Welling, M. (2016). "Group Equivariant Convolutional Networks." Proceedings of the 33rd International Conference on Machine Learning (ICML 2016), pp. 2990-2999.

Du, H., Nguyen, H., Thudumu, S., Vasa, R., & Mouzakis, K. (2026). "Goal-Oriented Multi-Agent Reinforcement Learning for Decentralized Agent Teams." IEEE 23rd Consumer Communications & Networking Conference (CCNC), pp. 1-2.

Duvvuri, S.S., Ekbote, C., Bansal, R., Tiwari, R., Khatri, D., Brandfonbrener, D., Liang, P., Dhillon, I., & Zaheer, M. (2026). "Interleaved Head Attention." arXiv:2602.21371.

Ghorbani, M., Li, Z., Pasquini, A., Vasa, R., & Birbilis, N. (2025). "Supervised Machine Learning for Corrosion Assessment of Multi-Principal Element Alloys Using Experimental and Generative Datasets." npj Materials Degradation, 9:155.

Nguyen, H., Thudumu, S., Du, H., Vasa, R., & Mouzakis, K. (2025). "Overcoming Semantic Dilution in Transformer-Based Next Frame Prediction." AMCIS 2025.

Noether, E. (1918). "Invariante Variationsprobleme." Nachrichten von der Gesellschaft der Wissenschaften zu Göttingen, Mathematisch-Physikalische Klasse, pp. 235-257.

Roy, M., Thudumu, S., Fisher, J., Vasa, R., & Mouzakis, K. (2025). "Device-Robust Spectral Grading and Origin Detection from UV-Vis-NIR Images: Towards Practical Gemstone Quality Assessment." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops.

Satorras, V.G., Hoogeboom, E., & Welling, M. (2021). "E(n) Equivariant Graph Neural Networks." Proceedings of the 38th International Conference on Machine Learning (ICML 2021), pp. 9323-9332.

Shankar, S., Zamfirescu-Pereira, J.D., Hartmann, B., Parameswaran, A., & Zaharia, M. (2024). "docETL: Agentic Query Rewriting and Evaluation for Complex Document Processing." arXiv:2410.12189.

Song, P., Han, P., & Goodman, N. (2026). "Large Language Model Reasoning Failures." Transactions on Machine Learning Research.

Vasa, R. (2026). "Two Compressions: AI Can Find Kepler's Ellipses. It Cannot Yet Find Newton's Laws". Vasa Substack