Talks
Workshops, conference talks and briefings since 2017, most recent first.
Who decides what an artificial intelligence should optimise? Advanced manufacturing is the case: the question is not what the technology can do but how the objective function gets written, and by whom — which is the decision that gets automated along with everything else.
A live demo takes one prompt from raw data to a finished dashboard, and the talk asks what it means that the product dissolved into the prompt. It closes on a sustained critique of Acemoglu’s Simple Macroeconomics of AI: the modest 1.2% follows from the shape of the model — Hulten’s theorem and an exogenous task taxonomy — and restoring the feedback loop he excludes overturns the conclusion. Choosing the formal framework is itself a political choice about what the model is permitted to find.
Seven primitives — CLAUDE.md, hooks, skills, subagents, teams,
commands and session management — argued as the unit of reuse in a
data-science practice, in place of prompts. Context rot is the
constraint the whole design answers to, and the talk ends on a minimal
setup and a maturity ladder to start from on Monday.
The thesis comes first: encode the norm — objectives and limits — as an executable harness, and evolutionary search over code can explore the policies and heuristics that maximise welfare under verifiable constraints. Built from Move 37 through AlphaFold to AlphaEvolve and tested against Hassabis’s conjecture that any pattern in nature can be modelled efficiently by classical algorithms, it closes on what that makes programmable, auditable and measurable in public policy, with the Corredor Interoceánico as the worked case.
Read through Lyapunov stability, Prigogine’s dissipative structures and Beer’s Viable System Model: a trained model reduces variety in Ashby’s sense while a language model preserves it, and that difference decides where each one belongs.
A survey for faculty across the whole stack — supervised learning, reinforcement learning, deep learning and generative models — organised around what each one makes possible rather than how it works, and worked through a live unemployment-prediction example with applications in health, environment and education.
Machine learning, reinforcement learning and generative models set against the decisions they can actually change, with centaurs — human-and-model pairings — as the working unit. The hard part is not the model but naming the decision and the action it enables.