The Anthropic Institute admits its models already write most of the code used to build the next ones. A critical reading of recursive self-improvement — without swallowing the hype.

The human musculoskeletal system works because there’s a skeleton underneath. Modern data and AI stacks have the same problem: many muscles —Pydantic, JSON Schema, SQL, GraphQL, RDF— pulling with no bone underneath. LinkML proposes that bone.

I started out wanting a lofi session to study to and ended up with six mythological agents arguing among themselves about how to chain tracks. This is the journey of ApolloAgents from the first script in February to version 3.1, which now DJs live inside the browser.

Two days ago we published a benchmark of our medical AI for World IBD Day. A single scenario was bleeding: 0.510 points below a bare GPT-4o. We could have published it anyway. We didn’t. This is the autopsy of the four bugs we found, in series, covering for one another.

The first three idols of augmentation attack your freedom. The fourth — the Genic Guidance — attacks your conscience: it turns refusing to redesign the human into a moral failing. And then, the way out of the labyrinth, hand in hand with Teilhard de Chardin.

There’s a moment when you stop walking toward something and start kneeling before it. That’s how one ends up before the three idols of augmentation: the Designable as Inevitability, Augmentation as Neutrality, and Fusion without Remainder.