Between editions

Articles

What the news leaves out, what a paper actually says, what a model can and cannot do. The English pieces live on Substack: free, dated, no paywall. The Italian archive goes out in English one piece at a time.

Subscribe on Substack
  1. 12 October 2026The money in AI goes to those you cannot replaceWho actually profits from AI? The summer 2026 results show operating margins above 60% in some links of the chain. A map from minerals to apps: where profits sit today, and where they may move.
  2. 4 October 2026Buying without buyingFive times in under two years, a large AI group has licensed the technology of a smaller company and hired its founders. None of the five was bought by that group, and none remained the company it was.
  3. 27 September 2026My book is out: how AI really works, and how to hand it real workThirteen chapters for people who already use AI at work and want to know when to trust it.
  4. 26 September 2026Jev, a model that doesn't write a wordTypeSafe AI has released a model that returns probabilities instead of text. How it works, what it promises, and which numbers hold up.
  5. 19 September 2026The summary the agent writes to itselfWhen the context window fills up, the agent hands over to itself: writes the note, opens the new window, picks up from there. Nobody reads that handover.
  6. 19 September 2026The price per token is half the billEvery ranking compares what a million tokens cost. None of them tells you how many tokens the job takes, and that is where the bill is decided.
  7. 7 September 2026The Only Winning MoveA 1983 film and a 2026 survey measure the same thing: how much human is left in the loop.
  8. 6 September 2026We are debating plagiarism while a generation outsources what it believesCognitive surrender, belief offloading and algorithmic monoculture: what we hand over to AI without noticing
  9. 5 September 2026Europe made AI literacy a legal duty. Now someone has to teach itSince February 2025, organisations that deploy AI have had to make sure the people operating it understand it. Schools are organisations. Almost nobody is asking who trains the teachers.
  10. 5 September 2026Nvidia bought the shelfHugging Face is where AI models are found, compared and downloaded. That is why Nvidia is paying 86 times its revenue.
  11. 2 September 2026When AI pursues a goalWhy checking individual actions is no longer enough with long-running agents
  12. 1 September 2026LLMs and numbers: oil and waterYou can spot a wrong sentence. You cannot spot a wrong number - and what to do, in practice, to trust the answers
  13. 31 August 2026The model is the engine, but you drive the carWhy the difference between AI assistants no longer lies in the model - it lies in what gets built around it
  14. 30 August 2026The loop: what turns a model into an agentAnd why, in analysis - unlike in code - the agent has to build its own judge
  15. 29 August 2026A bigger window does not buy more attentionWhy a larger memory is not enough for an AI system - and how agents learn to split the attention among themselves
  16. 28 August 2026Three values are enough to run a language modelEvery weight can be -1, 0 or +1. The model barely notices, and the electricity bill does.
  17. 27 August 2026Winning by letting goWhy Anthropic gave away its most successful invention - and why giving it away was the only way to make it win
  18. 26 August 2026The memory an agent needs is not your profile. It is the state of the workFrom assistants that complete a task to assistants that carry the work forward
  19. 25 August 2026When AI agrees with you just to keep you happy: the problem of sycophancyHow language models learn to please you, and why that makes them unreliable
  20. 24 August 2026Claude Cowork: when the AI assistant steps out of the chat and works alongside youFrom conversational interaction to operational collaboration: how an agentic system that acts on your desk works
  21. 23 August 2026A robot has to imagine the consequences before it movesWorld models, simulation, and why the hardest part of physical AI is getting the data
  22. 21 August 2026How to make your AI assistant actually remember youPersistent memory is the only part of an AI system you can open, read and rewrite. Almost nobody does
  23. 20 August 2026Your assistant did not forget. It can no longer see.How to spot a saturated chat, and how to hand the work over to a new one
  24. 19 August 2026You are not choosing a model, you are designing a systemFive questions to ask before adopting a large language model - and the cases where the answer is not to use one
  25. 18 August 2026A RAG system works by ignoring almost everything you gave itRetrieval-augmented generation, step by step: why the quality of the answer depends on what the system decides not to read
  26. 17 August 2026An AI has four memories, and they live in four different placesWhat a model actually remembers, where it keeps it, and what happens when the conversation runs too long
  27. 16 August 2026Agentic AI: when the model stops answering and starts doingThe shift from systems that produce text to systems that produce outcomes - and what it takes to keep them on a leash
  28. 14 August 2026Chunking and the context windowThe art of metering information into an AI system
  29. 13 August 2026Physical AI: when artificial intelligence gets a bodyFrom models that generate content to machines that sense, decide and act in the real world
  30. 12 August 2026The machine learning that came before the neural networks
  31. 12 August 2026Why a book about AI has to keep rewriting itselfWhat a living book is, and why a print date is not a truth date