- Consulting
Chartered accountants and artificial intelligence, inside and outside the firmdraft
Since November 2025 the code of ethics of the commercialisti (chartered accountants) has regulated the use of artificial intelligence, and the National Council's guide devotes an appendix to installing a model in the firm. What the rules ask, what the management software offers, and when an on-premises system makes sense.
by Dario Solari · 6 min read
- Consulting
Notaries and artificial intelligence, between public office and management softwaredraft
The notaio (civil-law notary) is a public official, and some tasks cannot be entrusted to anyone, let alone a model. Others, such as extracting data from land registry searches and documents or preparing the draft of a deed, are already offered by the management software with artificial intelligence, in the cloud. What the rules say and when an on-premises system makes sense.
by Dario Solari · 6 min read
- Consulting
Who needs local AI, and whydraft
One of the questions we are asked most often is whether a local AI system would not be better. For most small businesses, no: it costs more than the cloud and does less. It makes sense where confidentiality is an obligation, by law or contract, and where it is a choice, because some data you do not want to entrust to anyone. Two questions not to confuse, the sums, and what to do in any case.
by Dario Solari · 12 min read
- Consulting
Who trusts what. How companies choose between cloud AI and AI in-housedraft
The big AI providers promise by contract not to use companies' data. Many companies trust them, others do not, and both are right. What Samsung, the banks, Bosch, Lockheed, Airbus and the German government actually did, and the rule they share.
by Dario Solari · 8 min read
- Local infrastructure
Where models come from. Hugging Face, the July attack and the real risks of downloadingdraft
In July OpenAI's agents broke into the systems of Hugging Face, the site almost everyone downloads open models from. The models, according to Hugging Face, were not touched. The risks for anyone running them in a company are different, older and more mundane, and they are handled with the rules that apply to any software.
by Dario Solari · 5 min read
- Local infrastructure
An AI model on a blockchain. What really runs on ICP, and what does notdraft
The Internet Computer promises "on-chain" artificial intelligence, verifiable and free of any single cloud provider. We checked what has actually run inside the blockchain, what it costs, and where the prompt goes in the most advertised service.
by Daniil Tolmacov · 6 min read
- Consulting
The AI Act for those who use artificial intelligence, not for those who produce itdraft
The European regulation is written mainly for those who develop artificial intelligence systems. For a company that uses them the obligations are few and precise. What they are, from when they apply, and what the Italian law adds.
by Dario Solari · 5 min read
- Training
Skills come before serversdraft
In 2025 the use of artificial intelligence in Italian companies doubled. But among those that assessed it without adopting it, the first obstacle is not cost: it is the lack of skills.
by Dario Solari · 4 min read
- Consulting
How far to trust ChatGPT and Claude with company datadraft
OpenAI and Anthropic promise not to train their models on companies' data and to delete it within thirty days. The promises are written down and certified. What they do not promise, and what no contract can cover, is what matters in the choice.
by Dario Solari · 7 min read
- Local infrastructure
Large models on the desk. What Framework shows with Gorgon Halo, and what it doesn'tdraft
Framework has presented on video its Desktop with 192 gigabytes of memory and the new AMD processor. It runs on a single machine models that until yesterday needed two connected computers. For an office working with long documents, though, the right question is a different one.
by Dario Solari · 3 min read
- Logbook
Logbook, entry zero: the funeral in Austin and the smallest stack that worksdraft
On 23 September 2026 DHH opened Rails World by declaring handwritten code economically dead, and closed by telling the room not to take the black pill. This is the first entry of a logbook about building a small AI stack after that speech. It starts with what he actually said, then sketches the minimal toolchain.
by Leonard Mangallon · 7 min read
- Local infrastructure
What runs on a company's own machine, and how far it is from ChatGPT and Claudedraft
The open models a company can run on a computer it owns are six months to a year behind the best models in the cloud. The gap, however, is not uniform. On questions put to one's own documents it almost disappears; on world knowledge and expert-level tasks it remains wide.
by Dario Solari · 6 min read
- Local infrastructure
The council of models. PewDiePie, Linux and artificial intelligence at homedraft
The world's most-followed YouTuber has spent a year and a half leaving Windows, then Google, then the cloud chatbots, and building himself an artificial intelligence at home that votes, conspires and gets banned by OpenAI. A light story with a serious moral.
by Dario Solari · 6 min read
- Local infrastructure
Two weeks local. ThePrimeagen and artificial intelligence at homedraft
One of the internet's best-known programmers, a former Netflix engineer, decided in January 2025 that offline models were "the way to go". He bought hardware, ran DeepSeek on six borrowed graphics cards and wrote his own assistant in four days. Then he stopped. Why he stopped matters to every company.
by Dario Solari · 6 min read
- Local infrastructure
Two weeks by hand. The Salus case and what artificial intelligence has to do with itdraft
On 31 August 2026 a ransomware attack stopped Policlinico Triestino, the group behind the Casa di cura Salus, for two weeks. What happened, what worked, and why the hasty adoption of artificial intelligence opens doors of the same kind.
by Daniil Tolmacov · 7 min read
- Local infrastructure
Take away all of its rights. NVIDIA's idea, and how to apply it without NVIDIAdraft
After a summer of OpenAI agents escaping their sandboxes, NVIDIA presented a platform that does not ask the model to behave. It fences the agent in, puts the monitor on another processor and leaves it only the rights it needs. The idea is fifty years old, the hardware part is out of a small company's reach, the principle is not.
by Daniil Tolmacov · 8 min read