# When Uncertainty Matters - The Case for Uncertainty Aware Intelligence [Draft]

:writing_hand: Tim D. - November 2025

In expert decision making, a pillar of elite intelligence is the ability to weigh risk against reward using information that is diverse, incomplete and often conflicting. This balancing act is not optional. It underpins success in unfamiliar situations, in environments with little prior data, and in domains where the consequences of error are life or death. Even at the highest levels of expertise, what we value is not brute recall or speed, but judgement under uncertainty: the capacity to integrate signals, experience, intuition and the plausibility of different outcomes to reach balanced, defensible decisions.

If we want intelligent systems that truly assist or augment experts, then reasoning under uncertainty cannot be an afterthought. It must be the core operating principle.

Yet uncertainty has not been placed at the centre of modern AI. Deep learning has dominated the landscape, with architectures inspired by biological neurons only at the most superficial level. This resemblance has created a comforting illusion of biological similarity. In reality these models are data-hungry, computationally expensive and fundamentally naïve about uncertainty. They behave nothing like the small, elegant, deeply reasoning 1.5kg organ that navigates the world with remarkable efficiency at a peak power of 25W.

A more natural foundation lies in the Bayesian statistical tradition. Here, uncertainty is treated as a first-class object. An initial belief — the prior — is updated as new evidence arrives, yielding a posterior that captures both knowledge and confidence. This is a clean, principled and mathematically grounded way to handle uncertainty, and mirrors how humans reason.

Attempts to bolt Bayesian ideas onto deep learning have been incremental rather than transformational. They increase computational cost and complexity, but in pursuing expressiveness they lose the simplicity and interpretability that make Bayesian reasoning powerful in the first place.

Meanwhile, elegant classical methods such as Gaussian Processes have existed for decades. They quantify uncertainty naturally, work well in sparse or noisy data, and remain small and efficient. They are not brain-like, but they share many of the characteristics we associate with expert intelligence. In fact, Gaussian Processes power some of the most sophisticated decision systems in the world today — often in places far from public view.

My experience aligns with this strongly: for the right classes of problems these methods are not academic curiosities, but the natural choice.

But their broader adoption has been limited by three challenges:

* **Challenge 1. Scaling.** They do not scale naturally to high-dimensional inputs or outputs. As dimensionality grows, data requirements explode and computational cost becomes prohibitive.

* **Challenge 2. Language.** It is unclear how to apply them directly to modern language-based models, whose characteristics differ entirely from traditional quantitative systems. Yet this is essential since language-driven intelligence is now redefining work and driving the emerging agentic economy.

* **Challenge 3. Accessibility.** They require careful model design and specialist training. Barriers to entry remain high.

Recognising the breadth and value of uncertainty-aware methods, the challenge became clear: democratise advanced uncertainty-aware AI without requiring specialist modellers or exotic infrastructure.

Over the last two years we carried out sustained customer discovery across some of the most demanding, high-consequence environments in the world. This revealed not only solutions to the mathematical challenges, but also the additional layers required for real adoption. The barriers were not only model-level. They were product, platform, operational and organisational.

These lessons shaped our direction. We learnt what needed to exist for uncertainty-aware AI to move from niche to normal. Where foundations needed strengthening. Where mathematical tooling had to evolve. And how product design would determine whether these ideas reached real users.

This frames the rest of the paper: what we have learnt, what we have resolved, where we are focused, and what becomes possible as uncertainty-aware AI becomes central to mission-critical systems.

## Why Attention Is NOT All You Need — What About Representation?

The publication of "Attention Is All You Need" created a compelling narrative: scale and data alone allow attention mechanisms to internalise the full structure of a domain. Semantics, syntax, causal relations and admissible behaviours were assumed to emerge implicitly. Representation became a secondary concern.

This interpretation gives attention far too much credit.

Attention is not a generative model of the world. It does not define prior structure, specify the space of plausible behaviours or encode constraints and correlations. What it does is precise but limited: attention reweights information within an existing representation. In Bayesian terms, it performs a local posterior update, refining beliefs in light of context, without ever creating structure that was absent at the start.

This is the critical point.

Representation provides the prior — the geometry, assumptions and constraints that define what the model considers plausible.

Attention provides the posterior weighting — adjusting those beliefs given current input.

If the representation is unstructured or excessively flexible, attention must work disproportionately hard to impose order. Much of the scale and computational burden of modern foundation models stems from this: attention is being used to compensate for the absence of a well-formed prior.

Representation is the foundation of intelligence. Attention is the inference mechanism operating within it.

## Representation Is a Modelling Choice

Simpler representations are data-efficient and interpretable, but restrictive. Richer representations are expressive but data-heavy. Deep learning resolves this by pushing expressiveness to its limit: models with vast flexibility that require enormous data and compute. Classical models do the opposite: careful, deliberate construction by experts, performing brilliantly when well-designed but hard to democratise.

Real organisations end up with the worst combination: limited data, no internal modelling expert, and decisions that matter.

So why are humans such extraordinary representation engines? Because from birth we are conditioned by structure. Physics constrains our world. Language organises our thought. These structures prune the space of representations we even consider. Humans are powerful not because we compute more, but because we represent better.

## Goal-Oriented, Structure-Constrained Learning

Our position begins with a simple observation: intelligent systems work because the world is structured. Physical processes obey laws; engineered systems follow constraints; language exhibits grammar and semantic regularity. Human cognition exploits this structure relentlessly.

We take the same approach: design representations aligned with domain structure, not leave structure to be discovered implicitly. When structure is strong, representations become compact, data-efficient and robust. When structure is weak, uncertainty is explicit.

This avoids the paralysis common in modern AI workflows. Instead of awaiting perfect datasets, we begin with a structured representation of what is plausible, express uncertainty over what is not, and tighten the representation as understanding improves. This leads directly to models that are small, specialist and efficient.

## What We Have Built — Initial Focus on Quantitative Models

Our work began where systems are naturally quantitative: regression, simulation surrogates, computer vision and structured prediction. Here, representation matters acutely and uncertainty is essential.

Representation enters a model in two places:

* **Inputs:** how we organise features.

* **Outputs:** how we represent admissible behaviours.

Modern approaches often decouple these, allowing them to expand freely. Autoencoders optimise reconstruction, not reasoning. The representation preserves everything, not what matters.

**We do the opposite:** goal-oriented representations. Preserve only what is needed to predict the quantity of interest and capture uncertainty. The target becomes the organising principle of the representation.

Technically, this aligns with probabilistic manifold learning and active subspace discovery — but crucially, we quantify the uncertainty in the reduction itself. When structure is clear, subspaces are sharp; when data is sparse, the representation signals its own uncertainty. The system highlights what it does not know and where new data is most valuable.

Physical constraints are encoded directly into the representation. A model should not waste capacity learning conservation laws or monotonicity. These should be present from the outset.

Small, structured, constrained representations — equipped with explicit uncertainty — are far more elegant and efficient than compensating for missing structure through scale.

## Our Current Research Frontier — SLIMs

As our work expanded into language-driven reasoning, the limits of the foundation-model paradigm became stark. Large models are built on unstructured, unbounded representations. They rely on attention to impose coherence on a latent space that does not encode domain semantics or ontology. This is why they drift, hallucinate and behave inconsistently: attention is compensating for missing structure.

We chose the opposite path.

SLIMs — Specialist Lightweight Intelligence Modules — are small, domain-grounded intelligence nodes designed to operate as autonomous components inside larger decision systems. A SLIM is not a general-purpose model wrapped in instructions. It is a risk-aware Bayesian decision-maker, shaped explicitly by domain priors, curated evidence and ontological structure.

At their core, SLIMs replace “scale first, specialise later” with prior fitting. Using Prior-Data Fitted Networks, a SLIM approximates a Bayesian posterior predictive over tasks drawn from a structured prior. At inference time, it conditions on a small set of expert examples and produces decisions that reflect both knowledge and uncertainty.

SLIMs are transparent by design. Every prediction decomposes into prior basis, data-driven update and divergence between them. Divergence becomes a diagnostic for anomaly, domain shift and extrapolation.

Where quantitative models encoded physical constraints, SLIMs encode semantic and ontological constraints. Ontology-aware attention anchors the transformer’s latent space to domain structure, producing sparse, semantically aligned attention patterns. SLIMs learn faster, reason more coherently and remain within boundaries without relying on massive parameter counts.

SLIMs also operate in swarms — extraction, reasoning, verification, reporting, anomaly detection — mirroring expert systems. Each is auditable. The multi-agent behaviour that emerges is lightweight, resilient and interpretable.

In short, SLIMs extend our original philosophy — structured representation, Bayesian updating, explicit uncertainty — into language.

Early deployments in fusion, healthcare and cyber are validating the approach. They demonstrate that when representation carries structure and uncertainty is explicit, intelligence can be lightweight, auditable and deeply effective.

## The Start of the Product Journey

The turning point came during our last funding round. It was the moment we accepted that mathematical elegance alone would never deliver the impact we wanted. The ideas were powerful, but without productisation they would remain academic artefacts.

If we were to matter inside critical national infrastructure, engineering, and high-consequence decision systems, we needed the product that carried these ideas into the real world.

Most organisations struggle with real AI adoption. Despite tools like ChatGPT, workflows remain dominated by Excel, shallow prototypes and a lack of understanding about how to turn problems into deployable, auditable AI workflows. In regulated environments the barriers are even higher.

The gap between theoretical models and operational implementation is vast.

This forced a personal shift — leaving the comfort of academic idealism behind. Elegant ideas are irrelevant if they cannot reach real users. The work was no longer only mathematical; it was organisational, cultural and product-led. That transition became the culture of digiLab.

We now sit in a deliberate tension between academic simplicity and pragmatic product design. We reject elitism, reject complexity for its own sake, and focus on what works, what is usable and what can be deployed. This tension matters because of the scale of the opportunity: if we succeed in weaving reasoning systems into the core of critical infrastructure, we build something that lasts decades.

This was the moment the mission became clear — and the real journey began.

## Product: The Iceberg Beneath the Mathematics

Most people do not code. At EDF, ninety-seven percent of the workforce do not write code. The remaining three percent carry almost all analytical responsibility.

This creates a deep asymmetry: those who can experiment build trust through experience; those who cannot see only abstraction. Even with no-code interfaces, trust is limited without the ability to explore or interrogate the workflow itself. You cannot trust what you cannot touch.

Graph-based builders like n8n.io and Langflow popularised a visual UX pattern for interacting with AI. But they are automation layers, not environments for rigorous, uncertainty-aware workflows in safety-critical settings.

## Product Philosophy — One Platform for Two Worlds

To work inside real organisations, our platform must serve two audiences simultaneously.

- For the 97%, it must be intuitive: a visual workflow builder where AI workflows can be seen, explored and trusted.

- For the 3%, it must be fully developer-first: custom nodes, code injection, model building, integration, simulation, data engineering.

If either group is excluded, adoption fails. When both are supported, something important happens: the technical three percent become internal champions. They build high-quality workflows that the rest of the organisation can use safely. The asymmetry becomes an accelerant.

This duality — no-code on the surface, developer-first underneath — is one of the reasons our platform succeeds in complex environments.

## Graph Architecture — The Representation of AI Solutions

Graphs are the natural grammar of AI workflows. They reveal structure, encode dependencies and allow strong typing.

> "A toaster that only accepts bread is a safer toaster." *How I ended up describing strongly typed to a CEO recently.*

Unlike most workflow builders, we do not use global state. We pass information intentionally between nodes. This provides two major advantages:

- **Security by design.** Nodes see only what they need. Workflows are compartmentalised.

- **Full asynchrony and deployment flexibility.** Each node is a self-contained API, enabling heterogeneous deployment across local machines, GPUs or secure on-prem environments.

The deeper significance connects back to representation. A workflow graph is a structured representation of how a problem is solved. It captures transformations, dependencies, uncertainty propagation and validation. It externalises reasoning explicitly — the opposite of foundation models, which internalise reasoning implicitly inside billions of parameters.

This makes the workflow itself a machine-readable representation of operational intelligence.

And once the representation is explicit, something powerful happens: the platform becomes a data flywheel for expertise. Every workflow built or executed contributes evidence about how organisations solve problems, how tasks decompose and which structures are universal.

Over time, the system can learn the mapping from problem statement to workflow structure, allowing us to automate not just inference but the construction of rigorous AI workflows. This begins to productise the work of AI consultants — the design and orchestration of intelligent processes.

This is the quiet breakthrough.

By grounding workflows in structured representation, encoding uncertainty end-to-end and enabling deployment from cloud to secure on-prem, we are not just building tools. We are building the operating system for applied AI.

It is where representation, uncertainty and product meet; and it is the foundation for everything we are building.