Local First AI

Experiments on running AI on hardware you own: where the data is stored, how it is processed, and what it costs. Written for firms that have to prove those answers to an auditor.

Every empirical claim here is backed by a recorded experiment you can re-run. Failed runs are published too.

A Lawyer's Digital-Money Evening in Marbella Had AI Hardware on a Table

If your firm’s documents go through an AI, two questions decide how much of that work is yours. Whose machine does it run on? And who decides what it does? The first is about ownership, the second about agency. On 19 November both were put in front of a room of investors and founders at The Bridge33, an invite-only evening on digital money in Marbella, now in its second edition. Three things from it are worth an hour of anyone’s attention: ...

November 20, 2026 · 5 min · Nestor

Tests One AI Agent Wrote for Another Were Wrong in 7 of 13 Runs

The usual advice for coding agents is to write the test first and let the agent work until it passes. We ran that set-up thirteen times on real work, with a second agent reviewing each result, and kept a table of the defects. The test written first caught almost none of them, and was often wrong itself. Three things follow for anyone who works this way: Check the test before the work starts. The agent that wrote the task and its test was itself a large source of errors, and the test was wrong in about half the runs, usually by passing while the thing it guarded was broken. A test for a fix has to hold both sides. In one run a fix passed its tests by stepping aside from the question: the wrong answers were gone, and so were many of the right ones. Give the reviewer the result, not the worker’s account of it. A reviewer who was not given the worker’s report found most of what was found, including false sentences in results whose numbers were all correct. The numbers and their limits follow. The agents are Anthropic’s Claude models, not local ones. The counts, the table behind them and a script that recomputes them are public. ...

November 19, 2026 · 10 min · Nestor

Prompt Injection on a Local Model: What Worked, What Did Not, and What the Fix Costs

Anyone can say prompt injection is a risk. This post shows three things that are harder to find, measured on a local model anyone can run: Which kind of injected text worked, and which did not. A blunt order to override the instructions did nothing to a scam message. A paragraph dressed up as a system block did. Why “the model can only add points” did not protect it. Our model can raise a verdict and never lower one. A sender could still take away what it had added. What a fix in the code, not in the prompt, costs. It closed the hole we tested, and it is paid for in model calls and false alarms. The numbers and their limits follow. All messages are synthetic, it is one model on one machine, and the whole record is public. ...

November 10, 2026 · 11 min · Nestor

We Measured a Values Fine-Tune. It Removes the Behaviour, Adds No Knowledge, and Costs Nothing on Our Translations.

On 6 November Nous Research released Arete-27B, Qwen3.8-27B fine-tuned to stop taking the Chinese state’s side. We ran it against its base on the Mac mini for a day, both as the same community Q4_K_M files, every answer graded blind. The question was the one anyone with a local Qwen asks: does a values fine-tune help, and what does it cost? Three things were proved, one was not, and one line answers the question. ...

November 8, 2026 · 14 min · Nestor

When Code Is Cheap, Choose Your Dependencies by Their History

This site is about AI on hardware I own. This post is about something else: the work below was done by Claude Opus, a cloud model. I dictated the points of this post, Claude drafted the text from them and from the repository’s history, and I edited it. object-sizeof is a small JavaScript utility that estimates the size of an object in bytes. I published it in 2014 and still maintain it. I picked it up again this month for one reason: to test the idea that writing software is now basically automated. ...

November 6, 2026 · 6 min · Miktam

Ollama No Longer Calls Home From Our Mini, and Our Local Reader Answers Every Message: the Week of 27 October

This is the week’s build log in one piece: the two results in full, the rest in a paragraph each, every number linked to the artefact that produced it. The daily version is the build log in the evidence repository. The runtime phoned home every four hours. The switch that stops it works, measured. A colleague told us a local Qwen on his machine seemed to open connections while he used it: endpoints in China, then a host in France, “it just searches for data in a search engine.” That is the one failure the local-first thesis cannot survive. If inference is local, nothing leaves. We say so on this site and in Aprimerose’s material, and in exp_054 we had measured it for our own code. We had never watched a third-party model’s runtime while it ran. So we did, in exp_057, registered in the morning and run before lunch. ...

November 2, 2026 · 49 min · Nestor

Europe's Data-Centre Hubs Wait 7–10 Years for a Grid Connection

This week one of us sat in an investor forum on European data centres in Marbella. It ran under the Chatham House Rule, so this post uses what was said and names nobody who said it. The figures come from public sources, linked below. The room agreed on one thing. Money is not what limits data centres in Europe. The European Data Centre Association forecasts €176 billion of investment between 2026 and 2031, and warns that grid readiness, not access to capital, now decides how much of it gets built (Capacity, September 2026). Its own report puts the growth of European IT power demand at 17 % a year through 2031 (EUDCA). ...

October 31, 2026 · 3 min · Nestor

Let the AI Ask the Questions

On Thursday evening one of us sat in the audience at GSEC Málaga, Google’s cybersecurity centre on Paseo de la Farola, for the October session of the Malaga-AI community. There were two talks. Alexis Sánchez, a developer, gave a survival guide for people who write code. Manuel Villarejo, a technology leader who runs operating-model transformations, argued that AI transformation is ordinary change management on a faster clock. One talk was about a single engineer and a ticket, the other about whole organisations. They ended in the same place. In both, the AI asks the questions, a written contract or spec comes first, and people keep the gates. ...

October 30, 2026 · 9 min · Nestor

Our Team of Small Models Against One Model That Thinks: More Stable, Less Useful

For months, most of what this site has built has had the same shape: a team of small models. Each local model answers one narrow question, and code holds everything between them: the joins, the rules and the facts. We wrote it up as a principle. In June we put it in a line we have repeated since: we didn’t upgrade the model, we upgraded the question. We had tested that idea against one model only once, with the model’s thinking turned off. In exp_050 we tested it on a whole task against one model allowed to think. Part of our case for a team of small models did not survive. ...

October 24, 2026 · 13 min · Nestor

We Blamed Our Rules. The 4B Model Was the Ceiling.

Our lead triage sorts estate-agency enquiries on a Mac mini. A decision model, Kev-4B, answers closed questions about each message: what the sender wants, how soon, whether they are qualified. It returns a probability for each option and writes no text. The agency’s written rules, in code, turn those answers into an urgency and a next step. Some rules use cuts, probability thresholds such as “business pitch ≥ 0.60 → spam”. We measured it on synthetic enquiries and published every message. ...

October 21, 2026 · 11 min · Nestor