Artificial Intelligence I changed my chunking from 800 to 1,200 tokens, my hit@1 jumped +12%, and for thirty seconds I thought I’d found the optimum. In reality, my questions about the MSS60 were now being evaluated against the MSS54 index. The routing between corpora had been broken from the start, without anything in the metrics flagging it. A rising metric produced by a broken eval pipeline looks exactly like a rising metric produced by a genuine improvement. Three silent traps, multi-corpus routing, the absence of versioned history, and hit@k and MRR with no LLM judge, mean that most RAG teams make their decisions on numbers that measure nothing.
Artificial Intelligence Most AI SVG generators cheat: they hand you a base64-encoded PNG dressed up with a .svg extension. The test takes ten seconds, open the file in a text editor. In 2026, Recraft V4 is the only consumer tool that reliably produces real native SVG from a prompt, with Adobe Firefly Text-to-Vector as a legally better-protected alternative. Three trades collapse along the way: junior UI icon designer, low-end freelance logotype, vectorizer. What gains value is the upstream (brand design, design-system architecture) and the downstream, the technical integration.
Artificial Intelligence On 8 May 2026, Thariq Shihipar, an engineer on the Claude Code team at Anthropic, published a manifesto that’s been circulating ever since: drop markdown, switch to HTML for your specs, implementation plans and reports. The piece is clever, well written, and structurally an act of strategic marketing dressed up as personal experience. On a narrow subset (throwaway editors, interactive playgrounds, prototypes with sliders), the author is right; on everything else, his blanket enthusiasm conceals six blind spots. Token cost waved away, semantic HTML flipped into presentational HTML, loss of reviewability, the maintenance paradox that traps the user inside the production loop, reading that gets skimmed rather than scrutinised, an attack surface opened up by indirect prompt injection. An analysis of a poorly disguised strategic signal, plus a pragmatic decision matrix for sorting out where HTML deserves to be adopted and where markdown remains, by design, the right pivot format.
Artificial Intelligence Every session, Claude Code reopens your codebase like a visitor who has never set foot in it. It unfolds the README, fires off its greps, opens whole files to reconstruct an architecture it had already pieced together the day before. Understand-Anything, an open-source project released in March 2026, offers another path: index the codebase as a local knowledge graph, versioned like a lock-file, queried on demand by the agent instead of being re-read on every turn. This isn’t a token optimisation, it’s an architectural shift that extends the “RAG is dead, long live the Agent” thesis right into the code itself. Multi-agent pipeline, comparative economics, the limits of a project just six months old: what this break changes for Claude Code and for your workflow.
Artificial Intelligence Converting a technical PDF into RAG-ready markdown isn’t just a matter of running LlamaParse over it. On a 51-page chapter of the Funktionsrahmen MSS60, the Siemens documentation for the engine control unit of the BMW M3 E92 and M5 E60, I discovered that the interesting work begins after the automatic parsing. The semantic decisions that really condition retrieval quality, preserving the source language rather than translating, breaking dense tables into individual subsections, supplementing diagrams with French prose and pseudo-code, no parser will make them for you. This LLM-assisted work is done only once per document, its marginal cost per query is zero, and its upfront investment pays for itself in user-hours from the very first week of use. Field notes from two evenings of conversion that turned my understanding of what really makes a RAG pipeline worthwhile on its head.
Artificial Intelligence We spent two years debating the model. GPT-5 versus Claude 4.5, Gemini versus Mistral, million-token context versus prefix cache. Meanwhile, the elephant in the room stayed invisible: 90% of production RAG pipelines are swallowing a mush of badly extracted pixels, then act surprised when the LLM hallucinates. The problem isn’t in the brain. It’s in […]
Artificial Intelligence In 2026, generating a photorealistic image costs a cent, indistinguishable from a real photograph. Four professional markets are collapsing in silence: stock photography, e-commerce packshots, editorial illustration, routine retouching. Beyond the markets, it’s the social institution of photographic proof, a hundred and eighty years old, that is cracking: courts, insurance, journalism, elections, privacy. A cold mapping of the models that matter, of copyright laundering act II, and of what regains value when every image becomes suspect by default.
Artificial Intelligence Since April 8, Grok has been automatically translating all content on X, without you even noticing. No more clicking “Translate”: the algorithm now folds Japanese, Brazilian, or Nigerian posts straight into your feed. This invisible automation changes everything: language silos blow apart, creators reach global audiences, dissidents bypass censorship through instant translation. Of course, disinformation travels just as easily as information. But for the first time in history, billions of people can communicate with no language friction, and nobody seems to have grasped the scale of the earthquake.
Artificial Intelligence Anthropic first cut off unmetered access to agents, then, in the same breath, launched its own infrastructure service billed by the hour. The move is calculated, but it’s also honest. The real issue isn’t pricing: it’s what “managed” actually implies. Handing your agent over to an infrastructure you don’t understand means trading weeks of plumbing for a silent cognitive debt. Automating what you don’t understand isn’t automating; it’s outsourcing your incompetence with a better service contract.
Artificial Intelligence The 2018 U.S. CLOUD Act allows federal authorities to require any American company to hand over data hosted anywhere in the world, including in Europe. OpenAI, Anthropic, Google, and Microsoft are all subject to it, regardless of where their servers physically sit. Choosing Mistral isn’t enough: as long as access runs through Azure, the model stays hosted on American infrastructure under that same jurisdiction. Sovereignty begins the moment the model runs on infrastructure you control: a dedicated server in France, or Mistral’s own data center in Essonne. In 2026, with Ollama and quantized Mistral Small models, this architecture is no longer reserved for large corporations; it’s within reach of any team that already administers Linux servers.