Most problems come with a sample size of one.

ἅπαξ λεγόμενον. The thing said once.

A hapax legomenon is a word that occurs only once in a body of text. With no second instance to compare it against, its meaning has to be reasoned out from the context around it.

The problems that reach an outside adviser usually have the same structure. Each arrives in one specific configuration, with nothing closely comparable to check it against, so the answer has to be worked out from the evidence available.

Luigi Acerbi is Associate Professor of Artificial and Human Intelligence at the University of Helsinki, where he leads the Machine and Human Intelligence group.¹ His doctorate in computational neuroscience and machine learning is from the University of Edinburgh. He has published thirty-eight peer-reviewed papers, sixteen of them at NeurIPS, ICLR and ICML, with others in Nature and Nature Neuroscience,² and wrote BADS and VBMC, open-source inference software used in research groups worldwide.³ Those papers have been cited more than two thousand times.² His research area is uncertainty quantification, which in practical terms means knowing how far a model’s output should be trusted. At Hapax he is responsible for technical design, quality assurance and the advanced teaching.

Eva Martin runs the firm. She leads the advisory engagements and delivers the training, teaching the advanced sessions alongside Luigi. She has an MSc in Data Science and AI from Goldsmiths, University of London, where her thesis found that modern vision models significantly outperform the published CNN benchmarks for crisis-image classification.⁴ Before that she spent eight years in financial planning and analysis at Procter & Gamble and Coty, latterly responsible for planning at the €1.3B Wella Professionals unit. That meant distributor analytics covering more than forty markets, a €2B division’s forecasting through a divestiture, and a replacement forecasting system in Oracle built from scratch. She was also Coty’s data-analytics trainer for Europe, designing and delivering the programme across the region, work close to the training Hapax does now, and she spent those eight years on the client’s side of this table.

The work falls into three parts, each with its own page. Advisory covers feasibility studies, proof-of-concept builds and deployment. Training is closed and in-house, either a standard programme or one designed around your organisation. Counsel is a retainer of a few days a month, for clients who want an outside view while a decision is still open.

The scope and fee are agreed in writing before we begin. Whatever we build belongs to you when we leave, and engagements end with a handover so you can run things without us. There are two of us, so we take a small number of engagements at a time.

Three commitments shape every recommendation we make.

The first concerns fairness. Models learn from records of past decisions, and those records carry the biases of the people and processes that produced them. A hiring dataset remembers who used to get hired. A credit dataset remembers who used to get credit. Left unexamined, a model reproduces those patterns at scale. When we build or assess a system that touches people, we test how it behaves across groups before anyone relies on it, and when the data cannot support a fair answer we say so.

The second concerns transparency. A model that cannot explain its reasoning should not be making consequential decisions, and a system whose behaviour cannot be audited should not be deployed. We build and recommend systems that can be examined, and we document how they work in language their operators can actually read. Where a model is genuinely opaque, we say that plainly, because a client who knows what they cannot see is in a better position than one who assumes they can.

The third concerns scale. We recommend the smallest system that does the job, measured against what your organisation is now and where it is credibly going. An oversized platform costs more to run than it returns and ages badly. It also has a fuel bill: training and running large models consumes significant energy and water, and the gap between what a task requires and what gets deployed is, in aggregate, an environmental cost. This occasionally means advising a client to buy less than they had budgeted for.

Some of what we learn gets written up. At the moment we are working on a paper on when to trust a language model’s output, which will appear when it is finished. Ten case studies are published, each built on data with a held-out answer key, so every figure in them has been marked against that key.

The notes