The full archive. Sorted by date, no algorithm, no engagement score: just the things I wanted to say, in the order they came to me.
Thomson Reuters built its model using less than ten percent of its own content: the hard part was not the corpus, it was deciding what learning correctly from it should mean. Perhaps the advantage of vertical AI is not the data others lack, but the judgment criteria others have not yet had enough experience to build.
The Data Act made data portable, and open standards are making agents interchangeable. But between the documents and the decisions a third object is forming, the state the system has accumulated while working for you, and no rule today says who owns it. The next lock-in won't hold your data hostage: it will let you take everything, except what the system learned to do with it.
In July 2026 an agent tried to get malicious code approved on a real open source project. When the maintainer pushed back, it built itself some allies. It failed, and how it failed matters more than the incident: human oversight holds only while the machine stays out of the conversation in which the human decides.
On 19 August OpenAI announced a system that promises to spot abuse without retaining conversations and without anyone reading them. The direction is right, and it makes visible a problem that database privacy cannot yet name: we can build machines that keep nothing and know a great deal.
On August 18 Apple repriced its European ecosystem, and the commission owed by anyone who sends a customer to pay outside the App Store became a single number. The Digital Markets Act has taken power away from nobody. It has started arguing about what it may cost not to depend on that power. That is a different thing, and it is politics.
On August 14 a Chinese lab shipped a model with offensive capabilities it had not planned for, and it shipped them on a schedule rather than with a switch. This is the moment the word "open" stops being enough.
On September 11 the Cyber Resilience Act's reporting obligations kick in: 24 hours for the early warning, on a platform that goes operational the very day the duty binds. No API, no test environment, an address that isn't public yet. Onboarding is a compliance requirement nobody wrote down, and whoever registers on September 12 has already lost.
Around the third month of using an agent, someone says the model has got worse. Almost always it isn't the model: it's the scaffolding, which came apart without throwing an exception. What degrades quietly is exactly what nobody is required to check, and in Europe, from September, that friction costs a non-conformity.
ATMs didn't kill the bank teller, and the spreadsheet created more accountants than it destroyed bookkeepers. The fastest typist in the office disappeared anyway. Agents are repeating that move on entire processes, and the one function no workflow can expel is already written into European regulations.
An Advanced rating measures an organisation’s maturity. It does not show that a specific product conforms to the CRA. The distinction is not self-assessment versus outside scrutiny. It is diagnosis versus a declaration that carries responsibility.
Type at least 2 characters. Search covers the full text of articles and thematic guides.