Blog.

Cursor recreated SQLite using an agent swarm, without source code. Automated clean room, model economics, and infra bottlenecks.

The real question isn't the price of the subscription, it's how big a task an agent can hold on its own.

The math nobody does on local models: RAM, bandwidth, throughput. The core problem is physical, not software.

How we cloned the useful part of ProductBoard with agents and a tree of markdown files.

The new speech-to-text actually works. And what that says about the balance of power between software makers and users.

My first test: paying my supplier invoices on its own, from Gmail to Qonto.

The specifics of GenAI projects I wish I'd known two years ago.

Anthropic ships the first building block of an AI that uses a computer like a human. And it's available via API.

Forcing a LLM to answer exactly in the expected format: how it works and why it's key.

Text, images, sound: what multimodality concretely changes in use cases.

Three tips to de-risk a GenAI project before writing a single line of code.

Web search is shifting from the search engine to the answer engine. What that implies.

The real-time alert pattern we built on Agicap's AI Assistant.

Letting the AI refuse to answer and guide users to clearer questions.

A prompt pattern that turns vague questions into precise ones.

Let the AI ask the questions. A prompt pattern for proactive assistants.

First steps with Deno on a REST API, using Alosaur and TypeORM.