Technology

Technology is treated here as material to be taken apart, measured, and put back into service, not as a stream of announcements to comment on. This category holds both of the blog’s registers: the honest bench test, a keyboard, a vertical mouse, an OLED panel pressed into service as a monitor, a domesticated electricity meter, a server running FreeBSD rather than Debian, video encoding measured with VMAF rather than judged by eye; and the critique of what hardware and platforms do to their users, from the dongle that refuses to talk to its neighbor to the service you thought you owned that changes the rules mid-course. You will find pieces written after real, long-term use, often revisiting something that once seemed like a sound purchase, and a lasting interest in old ideas that still hold up, from the BGP protocol to the X Window System. Written by a practitioner who prefers a justified technical choice to a defended preference, and who provides the numbers when he has them.

RAG Chatbots: Your Visitor Speaks French, Your Corpus Speaks English

Artificial Intelligence

Your visitor asks a question in French, your documentation is in English, and your RAG chatbot either pulls up the wrong passages or answers in the wrong language. The embedding is rarely to blame; it handles the jump between languages remarkably well. The real culprits are the English-only reranker someone plugged in without a second thought, and a system prompt that ends up imposing its own language on the LLM. Here’s how to retrieve the right passages, rank them properly, and answer in the visitor’s language.

Jensen Huang Wasn’t a CEO. He Was Nvidia’s Operating System.

Economy

On November 12, 2026, Nvidia pushed out Jensen Huang, its founder and chief since 1993, and handed the interim job to its CFO. It’s tempting to see this as a simple change of driver. That ignores the fact that, with sixty direct reports and $2.86 million in profit per employee, Huang was the company’s operating system. Nortel, BlackBerry, Nokia, and Intel have already paid for the illusion of the parachuted manager, and Nvidia has just added its name to the list.

Fable, Opus, Sonnet, Haiku: Anthropic’s Guide Talks You Out of Its Own Top Tier

Artificial Intelligence

Anthropic has published a guide to choosing between Fable, Opus, Sonnet, and Haiku. Read with a calculator in hand, it turns into a confession: maximum effort, advisor mode, and the top-tier model cost a fortune for a marginal gain. The recommended default isn’t the most expensive model, and just two tasks out of twenty can blow up the bill. Here’s what to take away before you turn every dial to max.

France Is Building AI Sovereignty in Angers and Dismantling It in Parliament

Digital sovereignty

On November 3, the owner of Lidl teamed up with Bull, now state-owned, to build European AI infrastructure around the Angers factory. On November 8, Arthur Mensch warned that the Darcos bill, up for debate in the National Assembly on the 19th, will decide whether Mistral’s next gigawatt lands in the Essonne or somewhere else. Mensch is right about the mechanism, since proving a negative across billions of words is impossible, but his flat-rate levy would make copyright laundering official. Meanwhile, the French state is buying the servers with one hand and driving the workload away with the other.

Adobe Experience Manager: The CMS That Makes Websites Worse

Softwares

Every site I’ve watched move to Adobe Experience Manager has gotten worse, whether in usability, responsiveness, or raw speed. This isn’t the fault of a clumsy integrator: between ClientLibs, Adobe’s own scripts, and a Dispatcher cache that personalization quietly bypasses, the bloat is built into the platform. Adobe all but admits it by selling Edge Delivery Services as the way to finally get fast pages. And the people who sign the contract will never open the tool.

Ingesting Tables into a RAG: The Definitive Architecture Guide

Artificial Intelligence

Most RAG pipelines silently collapse when encountering dense, two-dimensional summary tables, turning explicit matrices into unparseable raw text. This guide outlines the exact architecture needed to split, structure, and vectorize complex tables without semantic loss. Discover how to leverage structure-aware chunking tools, contextual headers, and advanced embedding models to ensure precise retrieval at scale.

WordPress Object Cache: Why I’m Sticking with Redis Object Cache Instead of FlyingPress

WordPress & WooCommerce

FlyingPress can now manage a Redis object cache from its dashboard, and the single-plugin argument is appealing. Yet I still recommend Redis Object Cache, or Object Cache Pro for high-traffic stores, for three reasons: maturity, configuration, and coupling. An object cache loads before every plugin and touches every request, so it demands years of bug fixes, fine-grained configuration for multi-site hosting, and a life cycle independent of the page cache. Once FlyingPress’s integration has proven itself, the question will be worth asking again, and this post lays out the conditions.

Code Was Never the Product

Artificial Intelligence

Seven clones of Adobe’s software, written by AI agents, racked up 42,000 GitHub stars in eight days, yet their own reports rate them fit for barely a third of professional work. That isn’t a failure of the code; it’s proof that software is more than code. Legacy file formats, ecosystems, patented codecs, measured knowledge and long-term maintenance: none of it comes out of a prompt. When code becomes abundant, scarcity moves, and it settles precisely where these copies don’t go.

Abundance Will Have Owners, or It Will Have Masters

Economy

Elon Musk predicts that, thanks to AI and robots, money will barely matter by 2036 and everyone will receive a universal high income. On abundance, he is probably right; on the end of money, he has the wrong scarcity in mind, because land, energy, attention, and positional goods will remain scarce. Above all, an income is something that gets paid, set, and suspended: abundance received by bank transfer is not abundance owned. The real question is not how much the machines will produce, but how many people will own them.

Mistral Large 4: Le Chonk Doesn’t Settle Sovereignty. It Settles the Bill

Digital sovereignty

On October 6, Mistral unveiled Mistral Large 4, nicknamed “Le Chonk”: a trillion-parameter model trained on 3,800 GPUs in the company’s own European datacenters. It doesn’t catch up with OpenAI or Anthropic, and that isn’t its job. It draws level with the best Chinese open models, breaks into the global top five for cybersecurity (partly because American models refuse the task), and finally makes a question worth asking that nobody seems to raise: how much is France already exporting in tokens? Two weeks ago, I argued that Mistral had reduced sovereignty to a European access point. I stand by every word. But the picture needs completing.