This week I’ve been moving house (see previous post on how AI helped with that) but in-between packing boxes I’ve been onboarding a new employee to help me at Sunholo - R. Daneel (The R stands for Robot).
A large part of AI engineering these days is selecting which model is best for which task, a bewildering choice that is getting more complex, not less, over time.
Lots of people back from summer holidays now so less technical work this week, but that gives me a chance to step back and examine the AI projects I’ve been working on and start to analyse what really gave value.
A couple of weeks ago I wrote about the dark factory loop being used to self-improve AILANG, the programming language designed for AI rather than human programmers.
There is a lot of noise in AI trends which can seem abstract or theoretical to those not following every day, so I’ll try not to add to that but instead show a public example of a living, breathing open source project and how it is using terms like “loop engineering”, “dark factories” and “agent graphs” to get things done today.
I’m currently working on two projects that have both converged to a pattern I think will become a standard in AI-human co-creation.
As you may have heard, Anthropic was forced by the U.S. administration to revoke access to its latest Fable 5.0 model due to security concerns, and it may only come back as usable by U.S. citizens. This has sparked off various reactions across Europe on AI sovereignty which should be addressed.
Last week I presented at IDA Driving AI 2026 in Copenhagen. This is the talk retold while it's still fresh
This is the sixth post in a series on AI delegation, trust, and authority. Read the series introduction here. Earlier posts cover authority, reproducibility, visibility, and decision budgets.
One consequence of AI being trained to be so eager to please — the "helpful assistant" persona baked in via reinforcement learning — is that it will make things up. This is perhaps the first difference we must internalise, and where we should be careful not to anthropomorphise the machine too readily. If a human makes things up, we suspect deceit and ulterior motives. An AI's motives are fashioned by that eager-to-please training: it invents references (hallucinations) because it's trying to please you. This raises real questions about how adversarial we want AI to be — an AI trained for brutal honesty may give us better truth-tracking, but at the cost of the compliance that makes it useful in the first place. For now, we can build trust in AI only if we give it a way out to say "I don't know". Most hallucinations I see these days are the prompt's fault rather than the AI's.
An example of where this goes seriously wrong is when an AI is deployed in a responsible, public-facing position with no refusal path. New York City's MyCity chatbot is the most documented case.
The diagnosis that ties the series together: MyCity didn't lack knowledge. It lacked a refusal path. Every one of those answers should have been "I don't know — consult a lawyer or call 311." Instead, every one was a confident paragraph.