Archive

Open vs. closed isn't the split anymore. The question is which layer you control—and whether you get to choose who learns from you.

Two statements, eleven days apart. Karp and Nadella describe the same movement from opposite ends of the value chain. The interesting question is what follows from taking their argument seriously—including toward their own companies.

An application is something we attach ourselves to—a specialised place for a specialised use case. The new workbenches do the opposite. Intent travels across use cases, invoking the tools and systems required along the way.

Coding agents multiplied commits by 1.8×. They multiplied shipped releases by 1.3×. A new NBER paper just measured the gap between RPM and kilometers traveled.

Frontier models can disprove a decades-old mathematical conjecture. The same models misread an analog clock half the time. Stanford calls this the jagged frontier. I think 'topography' better captures what we're actually discovering.

MCP solved 'how do I use this.' ARD is going after 'how do I find this.' Same playbook Google has been running since 2005.

Microsoft's reorganization reads like a leadership story. It may be something deeper—Conway's Law in action, at a company of 200,000 people.

The harness is the charging network. Without it, agent autonomy isn't viable—no matter how powerful the model behind it.

Agents are not applications. The managed Windows laptop most enterprise employees use was built around a security model that assumes the opposite.

The input layer is becoming almost paradoxically simple—a command, a sentence, an intent. The output layer is moving in the exact opposite direction: increasingly structured, visually rich artifacts. The complexity hasn't vanished. It has migrated.

What OMCR illustrates is a deeper pattern: the grammar of software development is being ported, almost intact, into the production of knowledge itself.

It was here, in a frugal technological environment, that I first discovered the vitality and the sheer potential of the open-source digital ecosystem. And, in many ways, decided to dedicate my career to it.

A team built like software, for years. But the more our teams embraced agentic AI, the more it felt like the wrong analogy—they don't feel like software anymore. They feel like an agent.

Without a shared memory, agents don't scale knowledge—they just scale repetition, unnecessarily burning tokens at machine scale while scaling inconsistency.

The real contest isn't to control the model—it's to command the runtime that executes it. Five players, five very different bets on where the agent should actually live.

The next key question may not be who has the best model, but rather who controls the agent runtime acting on behalf of the user.

Intelligence exposed through APIs gets observed, compressed, and re-injected into the next generation of models. The system starts feeding on its own outputs.

One visible surface, for humans—interface, narrative, persuasion. Another underground layer for agents—tools, schemas, capabilities. The web is increasingly stratifying into two.

What happens when open source is used more than ever, yet visited less than ever? AI didn't break the license—it broke the loop.

Interfaces become a native output of AI, not something designed upfront. In a post-search world, value no longer sits in ranking pages, but in owning the surface where intent is fulfilled.

The early large models, prone to hallucinations and black-box reasoning, quickly required the externalization of structure—functioning almost like an exoskeleton of guardrails, just as arthropods first solved structural weakness through external shells.

Alors que nos interfaces digitales tendent à rendre les expériences toujours plus uniformes, c'est pourtant dans la nuance, dans la pluralité des points de vue et des cultures, que naissent les idées les plus fécondes.

OpenAI and Google just rallied around MCP—more than a technical alignment, it may be a turning point in the standardization of agentic infrastructure.

Eight years of partnership, 500 websites, and a framework that stayed true to its design principles while everything around it kept changing.

You never see as far as when standing on the shoulders of giants—a lesson relearned in front of the largest truck ever built.

Artificial does not mean magical—on the contrary, it means made by our hands, crafted with art, as its etymology suggests. A year of exploring generative AI honored a certain craftsmanship more than it threatened it.

What was originally an outsourced digital program to manage turned out to be, over the years, a remarkable human and technological adventure.

L'enjeu réside tout autant dans le risque que l'IA générative se développe dans une direction sous-tendue par des considérations économiques et anthropologiques tout autres.

The only work that truly sets one free involves four stages: setting a goal, choosing the means, implementing them, and finally contemplating. Too often, contemporary machines trap us in the relentless execution of imposed tasks.

L'IA moderne et ses automates sont un héritage direct des intuitions de Pascal, qui nous invite à travers les siècles à ne pas perdre de vue la question du sens de nos aventures technologiques.

LaFabrique, c'est un modèle entrepreneurial qui met l'Homme au centre et valorise son savoir-faire, sans l'opposer ni à la technologie ni au matériau.