Sami Belhadj

Claude watermarks everything it writes now. Read the limitations page first. On 2 August 2026 the transparency obligations of the EU AI Act became applicable. Anthropic signed the European Commission’s Code of Practice on Transparency of AI-Generated Content and switched on machine-readable marking across every Claude surface, worldwide, not only in the EU. The news traveled fast and the…

Claude marque désormais tout ce qu’il écrit. Commencez par lire la page des limites. Le 2 août 2026, les obligations de transparence de l’AI Act européen sont entrées en application. Anthropic a signé le Code of Practice de la Commission sur la transparence des contenus générés par IA, et a activé le marquage lisible par machine sur l’ensemble des surfaces Claude, dans le monde entier et pas…

Comment un petit modèle apprend à faire un travail aussi bien qu’un grand, ce que cela coûte, et quand c’est une mauvaise idée. Guide conceptuel et opérationnel. Dernière revue : août 2026. Ce que couvre cet article La distillation est la technique derrière la plupart des petits modèles qui dépassent leur catégorie, et derrière une bonne part des réductions de coût obtenues après une première…

How a small model learns to do one job as well as a large one, what it costs, and when it is the wrong idea. Conceptual and procedural guide. Last reviewed August 2026. https://open.spotify.com/episode/7wii0EMLIvHNZsVdhnp2MB?si=mA_c2hj7SzyRDEzbvgctww What this covers Distillation is the technique behind most small models that punch above their weight, and behind a large share of the cost…

Three days ago I argued that the fix for long-context failure is to send the model less. That was right, and it was incomplete. It tells you to shrink the context. It does not tell you what to keep. The default answer to that second question has been recursive summarization. Every few turns, compress the interaction history into a natural-language summary, carry the summary forward, discard the…

Vous avez récupéré les bons documents. Vous les avez tous mis dans le prompt. Le modèle est quand même passé à côté de la réponse. Cet article explique le mécanisme derrière cet échec, et ce qu’il faut faire. En résumé : un transformeur ne traite pas toutes les positions de son entrée de la même façon. Le rappel d’information suit une courbe en U. Ce qui se trouve au début et à la fin du…

You retrieved the right documents. You put them all in the prompt. The model still missed the answer. This article explains the mechanism behind that failure “Lost in the Middle” and what to do about it. The short version: transformers do not treat all input positions equally. Recall follows a U-shaped curve. Information at the start and end of a context is used well; information in the middle…

Un benchmark pratique de Spec Kit, OpenSpec, Superpowers, BMAD Method, GBrain, des skills de Matt Pocock, et de quelques challengers. Quelque part entre « l’IA va remplacer les développeurs » et « l’IA écrit n’importe quoi », une histoire plus discrète s’est jouée : les développeurs ont arrêté de débattre des modèles pour débattre du processus. En 2025, la question était « quel agent utiliser ».…

A hands-on benchmark of Spec Kit, OpenSpec, Superpowers, BMAD Method, GBrain, Matt Pocock’s skills, and a few challengers. Somewhere between “AI will replace developers” and “AI writes garbage,” a quieter story has been playing out: developers stopped arguing about the models and started arguing about the process. In 2025 the debate was which agent to use. In 2026 it’s which framework you wrap…

HANDS-ON REVIEW  ·  26 JULY 2026 Claude Opus 5 Anthropic’s fourth model in two months arrived with a pitch that sounds modest: near-frontier intelligence at half the price. The pitch undersells it. One score in this release moved twenty-fold, on the one benchmark specifically designed to resist the tricks labs use to inflate numbers. This is the full review: what shipped, what the numbers…

The bottleneck moved Coding agents write features now. They fix bugs, scaffold services, and open pull requests that a reviewer can plausibly approve. Whether that is a good development is a separate argument, and a real one. It is also not the argument most teams are having on a Tuesday morning. The question in front of them is narrower and more awkward: why does output quality vary so much…

On July 9, OpenAI moved the GPT-5.6 family to general availability, two weeks after a limited preview. There is no single “GPT-5.6.” There are three: Sol, Terra, and Luna. If you write code for a living, this release is worth ten minutes of your attention, and a bit of skepticism. Why it matters The headline is not a benchmark. It is economics. OpenAI’s pitch is that GPT-5.6 gets more useful…

A 5-part series — Part 5/5 The most common mistake isn’t picking the wrong discipline. It’s investing in a layer above where your team actually sits. People want to build a sophisticated harness when they don’t even have a test bench for their prompts. They stack autonomous agents on foundations that are taking on water. After four parts of definitions, here’s the part that’s useful at work:…

A 5-part series — Part 4/5 – Harness engineering When the Claude Code source leaked in late March 2026, people could finally measure something practitioners had suspected without proof: across roughly 512,000 lines of TypeScript, the model interaction itself is a tiny fraction. Everything else, the vast majority, is the harness. That sums up the outermost layer of our diagram. The model sits at…

A 5-part series — Part 3/5 Here’s the most counterintuitive idea in the whole series: giving a model more information can make it worse. It’s tempting to picture the context window as a big bag where you cram everything that might help. In practice it’s a budget. Every token you add spends another, and past a certain point the model degrades. Context engineering is the discipline that manages…

A 5-part series — Part 2/5 We buried prompt engineering a little too fast. The discipline isn’t dead, it’s been demoted. It now sits at the innermost layer of the diagram from Part 1: how you phrase a request, inside a single call. That’s small. It’s also the first place everything can go wrong. In this part I separate what actually works from what’s folklore. Because between 2022 and now, we…

A 5-part series — Part 1/5 Three years ago, knowing how to “talk” to a model was a rare and slightly magical skill. People traded screenshots of prompts that worked and passed around phrasings like cooking recipes. Andrej Karpathy summed up the era in a line that stuck: the hottest new programming language was English. Today nobody serious introduces themselves as a prompt engineer. The word…