Skip to content

Updates.

Essays and news from Axiomatic.

  • A language model does not reasonIn 2025 I announced “reasoning” as a feature. The word was wrong. A language model picks the most likely next text, astonishingly well, and does not know what it does not know. So: the model writes, the sources decide what is true, code does what must be exact, the professional decides.14 September 2026
  • “The professional decides” is a boundary, not a shieldReading on human oversight of AI in health care ends in one conclusion I share: a signature proves presence, not supervision. “The professional decides” has to mean two things at once. A line Inora does not cross, and a design duty to put in front of her what she needs to decide without redoing the work.25 August 2026
  • Oversight is designed, not staffedLast month the BMJ called the clinician in the loop a liability sink. Being present is not the same as judging. The authors concede that oversight works on tasks that are bounded and checkable, so that is the design target: make the task checkable, put approvals where someone has time and authority, and test whether people still catch an error.2 June 2026
  • Read in parallel, decide in one placeDo not build multi-agent systems, said one lab in 2025; multi-agent systems are working, said the same lab in 2026. Both are right once you separate reading from writing. In care the implicit decisions behind an answer are specific, and the line between reading and deciding has to be drawn in the harness, not left to the model.12 May 2026
  • Sell the work, keep the judgementInvestors now say the next big software companies will sell the work and not the tool. The split they use, between intelligence work and judgement work, is useful. Where I part ways: the line between the two is drawn by the system around the model and by law, and what compounds is not data pooled across customers.17 March 2026
  • Cheap intelligence makes the rest scarceThe model makers themselves say capability is ahead of use, and that a given level of it gets cheaper every year. If so, what is scarce is not intelligence. It is the sources with owners, the checks, and the work of making an answer fit one organisation. What follows for a company that builds on language models.20 January 2026
  • Variation is not noiseThe case for replacing professional judgement with algorithms stands or falls with one assumption: that a brain is a computer. Where a judgement has no right answer apart from the judgement itself, difference between professionals is information. Give them the same sources, not the same conclusion.10 December 2025
  • An evaluation set makes “any model” trueSince the rebuild I say Inora is tied to no single language model. The sentence is only true if we can show it. Without the organisation’s own questions and answers, swapping a model is a leap of faith; with them it is a test run. Why I count evaluation as a precondition, and where I part ways with the standard advice.19 November 2025
  • Rebuilt on open building blocksSix weeks in, we are rebuilding Inora from the ground up on open-source building blocks. A small company does not win by building everything itself, or by letting one platform make its choices. What we take, what stays ours, and what it costs.18 October 2025