Archive

In the last days of October, fleets of AI agents came back with claims on theorems and with formalized proofs. The history of the great explorations suggests what comes next: charts, trade, corsairs, and inhabited continents.

In 1936, an engineer counted airplanes and found a curve that seems to price AI tokens today. What governs technology, the control plane, was built for a slower pace and may need a redesign of its own.

What laptop should a company buy its people for the next four years? Microsoft and Nvidia share a stage this week to answer that question—but most of the answer may have already quietly shipped since June, while agentic AI might have weakened the foundations of Microsoft's historic rent.

A paper on the economics of information, a theory of disruptive innovation and a bet the frontier AI labs are about to make all describe the same thing: value lives in a gradient, and using it is what spends it.

To explain the 2026 AI investment supercycle, Satya Nadella pointed the industry toward 1873, not 1969: the year railroad financing, much of it circular, collapsed under its own weight. The panic destroyed the capital. It never touched the railroad—the distinction this AI cycle keeps collapsing into one claim.

A paper published in Nature quietly repairs a defect Socrates complained about 2,400 years ago: a text, once written, can't answer back. It arrives just as Google starts angling to become something like an editor again—deciding, fragment by fragment, what a text is allowed to say.

Machines have spent the last decade getting radically better at learning. Between PISA's falling scores and the Fields Medallists' warning about mathematics itself, have we spent it quietly getting worse at learning ourselves?

What Fermat teaches us, and a dive into the abyss of mathematical abstraction.

Every major relationship in AI right now looks the same: each side uses the other to get through an exponential capex phase, while quietly building the option not to need it later.

Two things happened the same week, and together they may say something real about local AI. A second signal, an idle-Mac inference network, points toward something more interesting: compute distributed rather than centralized.

Babel bâtit une tour sans jamais dire qui l'habite. Néhémie répare une muraille, famille par famille. Au milieu du chantier technologique de notre époque, le second nous invite à préserver tout autant nos murailles intérieures.

Before the boulevards, the crossroads. The current agentic moment looks like exchange taking shape before anyone has thought to organize it.

Codex Sites compresses the chain from idea to software. The final artifact is no longer code—it's a working application, generated from a conversation.

AI may be triggering a Copernican revolution in enterprise information systems—not by replacing ERP and CRM, but by reducing their exclusive gravitational monopoly over how work gets done.

A large part of the training substrate of modern AI comes from a broad cognitive commons—scientific work, books, code, archives, public discourse. What happens when value extracted from it is captured faster than it is redistributed?

Dario Amodei's technological optimism and the Vatican's Antiqua et Nova arrive from worlds apart, yet converge in surprising ways—both asking what makes us truly human as machines grow stronger.

Sora and Gemini's image generator collided with our perception of truth in the same week—both raising the same question about the sources that fed them, and what we're actually looking for when we ask a model to show us reality.

Reconstruire Notre-Dame à l'identique convoque l'archéologie, l'ingénierie et la data science — et rappelle que la technologie n'a de sens que mise au service d'un dessein humain.

Every gold rush ends the same way: enthusiasm first, regulation later, and a bill someone eventually has to pay.