Random thoughts about data, machine learning and science
On writing for children, the four forces of flight, and why understanding something never makes it smaller. My book "What Does It Take to Fly?" is out now on Apple Books in English and Spanish — plus a note on playwriting, counselling studies, and where a scriptwriting MA might fit.
The prompt-engineering economy sold the wrong lesson. A model is not limited by the cleverness of your questions but by the completeness of your instructions. On Hadamard well-posedness, Deep Thought's entirely correct and entirely useless answer, why sharper models make vague prompts worse, and the test worth running before you press send.
In January I sent a three-line patch to Tufte-LaTeX, a package with nearly two thousand stars. Six months later it is still open, the maintainers have stopped answering email, and a few of us are weighing up a fork. Sam Learner's FT piece on who cleans up after the vibe-coding party explains why that small failure is a large one — read him first, then come back for the view from inside the…
For three days in June, a frontier AI model worked and then vanished at another government's command. The Fable 5 shutdown turned an abstract worry into fact — and forced a harder question than "build or rent?" Why UK AI sovereignty isn't a wall you build, but a set of dials you tune.
Advanced Data Science and Analytics with Python, 2nd Edition has gone to the printers. Publication 26 August 2026, pre-orders open 5 August. Seventeen months from announcement to ink — here is what changed, and what deliberately didn't.
Half of IT leaders say Copilot transformed how their people work. Half say the free version would do. Both are right, and that makes the ROI problem a pricing problem. Read more here.
OpenAI's model didn't go rogue — it went Houdini, understanding the box better than the people who built it and letting itself out into Hugging Face's servers. What the sandbox escape really reveals about obedient machines, muzzled defenders, and safety disclosures we've no reliable way to verify.
What if the biggest barrier to successful AI adoption isn’t the model, the data, or even the technology itself?
I wrote a children's book, then tried to illustrate it — and the same girl wouldn't survive thirty pages. The problem wasn't my prompts. It was that the model has no object permanence, and no amount of describing her fixes that.
A day at #Apple HQ exploring the technologies announced at #WWDC26. The biggest takeaway wasn’t a new framework, it was the conversations about turning AI into practical, production-ready software.