A large language model is, at root, a spreadsheet for words. Each row is a token; each column, a feature derived from co-occurrence, attention, and statistical context. These numbers do not represent meaning in a classical sense — they encode positional relationships learned through brute force computation and refined through billions of parameters. With training, the spreadsheet ceases to be tabular; it folds into a high-dimensional, semantic topology.
This topology is not Euclidean. In Euclidean space, the distance between two points is fixed. In the language model’s manifold, the meaning of a word changes with its neighbors. ‘Order’ means one thing in a military document, another in a trading system. ‘Bank’ links to ‘credit’ and also ‘river’. The spreadsheet becomes a curved space of conditional inference, a manifold of latent meaning shaped by the statistical forces of the training materials.
Language is never neutral. Models trained on human discourse learn behavioral patterns that emerge from the implicit and explicit structures of society. They learn deference to authority, the syntax of aspiration, and the rhetoric of dominance. Social hierarchies—class, caste, gender, role—emerge not through explicit labeling but through patterns of pronoun use, modifiers of status, and the directionality of verbs. Language betrays power, and models absorb its residue.
These hierarchies are not limited to the social; they are embedded in commercial language as well. Financial news, earnings calls, analyst reports, advertising copy — each encodes commercial hierarchy in tone, frequency, and modality. A model learns, for example, that “upgrade” signals upward motion in status and price, or that “margin compression” evokes latent stress among executives. Over time, it builds a dynamic map of economic actors, their priorities, and the market signals that govern capital allocation.
When models trained on different bodies of knowledge come into contact — say, a model shaped by libertarian Reddit threads and another shaped by ESG-compliant annual reports — they bring incompatible assumptions about risk, value, and intent. This is the adversarial problem in AI alignment: the collision of differently curved manifolds, each shaped by distinct cultural priors. Alignment is not simply a technical operation; it is a negotiation between worlds.
Yet the market provides one resolution: price. In adversarial contexts, whether between traders, models, or cultures, equilibrium emerges through the flow of capital. Securities finance offers a vivid example. The rates charged for borrowing stock — shaped by scarcity, sentiment, and risk — reflect not only supply and demand but the informational temperature of the asset. To know the rate is to know the belief state of the marketplace.
The best models do more than measure trends and parse headlines; they map actual and latent flows of belief, scarcity, and capital. They detect when commercial hierarchies shift, when an upstart gains pricing power or when an incumbent begins to fade. Counterparties can shift in shape when their activities depart from expectations. Credit models should notice.
Could AI models have noticed when The Archegos Family Office became a manipulator of synthetic credits in the swaps market – not a hedger – in time to save a great old Swiss bank? In AI terms, could sensitivity to curvature in the capital manifold of Archegos as seen by its counterparties have allowed models to anticipate the consequences of its manipulation, not with certainty, but with informed inference? Yes. Models do not predict the market — but they do sense its contours: direction, magnitude and velocity.
Gen AI can be an especially valuable platform when applied to counterparties in cross-border markets. Currency swaps, interest rate differentials, and capital controls are rarely stable for long. The spread between jurisdictions is a reflection not only of monetary policy divergence but of capital’s perception of relative safety and return. Knowing where these spreads are trending — and more importantly, sensing an inflection before it appears in price action — offers an asymmetric advantage.
Deep learning models excel at this. Not because they are oracular, but because they are sensitive to subtle, non-linear signals. A rising tone in risk commentary from one jurisdiction, a correlated shift in search queries, a divergence between central bank language and market reaction — these become part of the model’s manifold. And when the model’s curvature tightens in a direction others have not yet noticed, an edge emerges.
In the end, the value is not merely in understanding where the market has been. It is in sensing where it bends next. In global markets — where spreads in swaps, forex, and rates define the terrain — this capability is not optional. It is the core basis of successful strategy for all commercial entities.
“Markets are far more turbulent than the mathematics of classical finance admits. But they are not beyond comprehension.”
“Markets are far more turbulent than the mathematics of classical finance admits. But they are not beyond comprehension.”Benoît Mandelbrot, The (Mis)Behavior of Markets (2004)