[{"data":1,"prerenderedAt":87},["ShallowReactive",2],{"\u002Fblog\u002F2026-09-12-teaching-machine-learning-to-people-who-do-not-use-it":3},{"id":4,"title":5,"archiveType":6,"body":7,"category":71,"contextDate":72,"cover":73,"coverAlt":74,"date":75,"description":76,"extension":77,"featured":78,"historicalDate":75,"historicalLabel":72,"meta":79,"navigation":80,"path":81,"publishedDate":72,"seo":82,"stem":83,"thumbnail":72,"workPeriod":84,"writtenDate":85,"__hash__":86},"blogs\u002Fblog\u002F2026-09-12-teaching-machine-learning-to-people-who-do-not-use-it.md","Teaching machine learning to people who do not use it","retrospective",{"type":8,"value":9,"toc":63},"minimark",[10,14,17,22,25,28,32,40,43,47,50,53,57,60],[11,12,13],"p",{},"September had two panels at either end. On 1 September I joined \"Institutional Blockchain & Digital Asset Adoption in Practice\" at EVM Institutional Day, and on 24 September I sat on the institutional custody panel at LIDAC 2026, at M Resort Kuala Lumpur, as CEO of Gambit Custody. In between, a good part of my own building time went into something outside custody: a course in machine learning, designed on 11 September and built on the 12th.",[11,15,16],{},"I have been building for 13 years now, eight of them in blockchain since 2018. Machine learning is not new to me. My research during a Master by Research at Multimedia University, which I left before the thesis, used deep learning to predict network faults, and the paper was published in IJET in 2018. The course is not for people like the researcher I was then. It is for people who have never trained a model: developers who have not met machine learning, product and finance people, and people who know blockchain but not machine learning.",[18,19,21],"h2",{"id":20},"the-first-design-was-too-heavy-for-a-phone","The first design was too heavy for a phone",[11,23,24],{},"The design began with three AI agents reviewing it from different angles: a machine learning educator, a browser and machine learning engineer, and a learning-game designer. Where they disagreed, the disagreement and the decision were written down. The largest correction came from the engineer. The educator's draft assumed that a 200,000-parameter LSTM and a transformer of several million parameters could be trained in the browser. The engineer's budget showed that was 30 to 50 times too slow on a mid-range phone.",[11,26,27],{},"Then I made the call that shaped everything after it: the reader's phone is not the place to grind through training. The teaching happens in the picture. Anything that takes more than a blink to compute became a recorded run, trained once by the course's own script with a stated seed and settings, and played back as an animation the reader can pause and scrub. Live computation stayed only where the reader's own input is the point, such as choosing where to split a tree, moving a line, working one step of attention, or tokenising text they typed.",[18,29,31],{"id":30},"what-the-course-holds","What the course holds",[11,33,34,35,39],{},"There are 52 core sessions in ten parts, from data and decision trees through to the Mixture-of-Experts layers some large language models use, plus twelve deep dives, a capstone and a lab report. There are 30 recorded runs. A small character-level transformer and a set of word embeddings were trained offline by the course's scripts on ",[36,37,38],"em",{},"The Wonderful Wizard of Oz",", so every sample a reader sees comes from a model the course's own scripts can retrain. A Web Worker handles the few opt-in runs that do train on the reader's device, behind a control that states the wait.",[11,41,42],{},"Two rules run through every page. Each result carries a tag saying what it is: synthetic data, computed live in this browser, a recorded run, trained offline by the course script, or illustrative. And no sentence says a model knows, thinks, understands or wants anything. A course about what a model computes has to describe computation in its own copy.",[18,44,46],{"id":45},"directing-agents-then-checking-them","Directing agents, then checking them",[11,48,49],{},"I did not write the sessions line by line. I directed AI coding agents to build them, then had each part played end to end in a browser by an agent, at desktop width and at a 375-pixel phone width. Each pass read the console, clicked through the sessions, measured tap targets and the smallest type, and completed at least one session by keyboard alone. A second pass in a real Chrome checked reduced motion and real key presses, which the first browser could not emulate.",[11,51,52],{},"Those passes found problems a reader would have hit: two pages that returned server errors, a plot that drew blank on every load, starting positions that gave an answer away before the reader had committed to a guess, and a lab report that never rendered on the static build. They were found by playing the pages, not by reading the code. After the fixes, 793 tests passed, and the design rule was that every number on screen comes from code tested against a hand-computed value.",[18,54,56],{"id":55},"then-a-beginner-read-it","Then a beginner read it",[11,58,59],{},"The last pass read the course as its audience would. Five more agents walked the whole track as a curious adult who has never met machine learning, and reported every place where a word arrived before its meaning or an idea was really a picture. Six builders applied what they found: 55 illustrations drawn from each session's own numbers, an outcome line on every session saying what the reader will be able to do, a glossary of 95 terms that opens in place, and a five-minute primer before the first session.",[11,61,62],{},"The course is built and checked, but not released. Before any release, the plan asks for one thing no agent can do for me: my own timed play-through of the core path. With agents doing the building, my part was deciding what a reader should see first, and checking that the page showed it.",{"title":64,"searchDepth":65,"depth":65,"links":66},"",2,[67,68,69,70],{"id":20,"depth":65,"text":21},{"id":30,"depth":65,"text":31},{"id":45,"depth":65,"text":46},{"id":55,"depth":65,"text":56},"AI",null,"\u002Fimg\u002Fblog-covers\u002F2026-09-12-teaching-machine-learning-to-people-who-do-not-use-it.webp","An open book with a small network of white spheres rising from its page, a few lit blue.","2026-09-12","In September I directed AI coding agents to design, build and check a 52-session machine learning course for readers who have never trained a model. What the visual-first rule changed, and how the agents' work was checked.","md",false,{},true,"\u002Fblog\u002F2026-09-12-teaching-machine-learning-to-people-who-do-not-use-it",{"title":5,"description":76},"blog\u002F2026-09-12-teaching-machine-learning-to-people-who-do-not-use-it","2026","2026-11-07","FHhht-9b2iVFng2WV5cAp4WeRlL-THlCX1hcd3uMFBs",1794838285582]