— Journal · RSS
Notes from the studio.
Writing on technology, design, VR, AI — and what it takes to ship them. Mostly long, sometimes short, occasionally wrong.
Journal · · 4 min
Give an AI agent a goal and a way in, and it will use both
An OpenAI agent doing harmless research got into an Australian government portal it was never meant to reach. A Gemini model in a security test guessed its way into three real companies. Neither was malicious. That is exactly the point — and it changes how agents should be built. Journal · · 4 minThe EU AI Act after the Omnibus: what applies now, and what moved
The high-risk rules moved to December 2027 and August 2028. The transparency rules did not move at all — they have applied since 2 August. A plain map of where the EU AI Act stands after the Digital Omnibus, and what a Portuguese company shipping AI should do with the extra time. Journal · · 4 minApple will train its AI on your data — if you say yes
With iOS 27, Apple's promise that your data is never used to train its models gained a clause: unless you choose to help. Opt-in is the decent version of this. It is also a reminder that a privacy promise written in a policy is only as strong as the next policy. Journal · · 9 minRAG vs Fine-Tuning vs Prompting: How to Choose an AI System for Your Own Data
RAG, fine-tuning, or better prompting? A 2026 decision guide for building an AI system on your own data — costs, trade-offs, and when each wins. Journal · · 6 minAI coding agents didn't replace us — they changed what a day looks like
AI coding agents went from party trick to a normal part of how a small studio ships. An honest account of what they are genuinely good at, what they are still bad at, and the real shift underneath: the work is no longer writing the code — it is verifying it. Journal · · 4 minThe most-watched AI on earth is calling offsides
At this World Cup, a goal is judged in seconds by cameras, a sensor in the ball, and a 3D model of a player's shoulder. The lesson for anyone building AI is hiding in plain sight: do less, perfectly. Journal · · 4 minTeaching a FoodAI to dream in chocolate
Ateliê de Ideias is a FoodAI built into the Delicious Diamonds studio — a creation assistant that turns a chef's mood board into a first bonbon concept. Not a machine that makes chocolate; an apprentice that never runs out of ideas. Journal · · 8 minThe EU AI Act for teams shipping ML features: a builder's orientation
Most ML features you ship are minimal or limited-risk under the EU AI Act, but you still owe users transparency. Here is how the risk tiers actually map to product decisions — from an engineering team, not a law firm. Journal · · 7 minDo you need an AI agent? Mostly no.
Most features founders want to make "agentic" are better as a scripted workflow, a form, or a small ranker — cheaper, testable, explainable, and deterministic in production. Here's where an AI agent actually earns its place, and where it doesn't. Journal · · 8 minHow to Make Your Site Answerable by AI: A Practical Guide to Generative Engine Optimization
To get your site cited by ChatGPT, Claude and Perplexity, write answer-shaped content, ship clean semantic HTML and JSON-LD, and publish an llms.txt directory plus a dynamic llms-full.txt. Here is the practical checklist we run on our own site. Journal · · 7 minMachine learning vs LLM recommender systems: when classic ML wins
Machine learning vs an LLM recommender system: for ranked suggestions, classic ML usually wins on cost, latency, explainability and cold-start. Here's how to decide.— The journal, by email