CONTEXT JAMMING

Field notes from inside the context window.

§ · Radar · No. 001

The Architectural Determinism C-Suite Radar.

A live scroll of the principals, interpreters, and intellectual fellow-travelers of the two labs this site reads most closely — Anthropic and Google DeepMind — and the adjacent canon that surrounds them.

Curated by Bret Kerr  ·  ACRA Insight  ·  Franklin, MA  ·  Updated every 30 min

— Register I · C-Suite

The principals of Anthropic and Google DeepMind, and the research leads closest to each lab's architectural center of gravity.

Anthropic · C-Suite

Chris Olah

@ch402

Circuits, features, monosemanticity.

Interpretability is finally a field with enough load-bearing results that you can teach a graduate course on it without fudging. Ten years ago we were arguing whether there was a there, there. We were wrong to be so tentative.

APR 17Flip

Anthropic · C-Suite

Dario Amodei

@darioamodei

Machines of Loving Grace.

Still the clearest framing I have: most of the argument for building this is downstream of whether the upside — curing disease, compressing a century of biological research into a decade — is real. The safety work is what earns the right to even try.

APR 16Flip

DeepMind · C-Suite

Demis Hassabis

@demishassabis

Scene construction, AlphaFold.

The line I keep returning to: AlphaFold worked because biology turned out to be more structured than we hoped and less structured than we feared. The next decade of AI-for-science is built on finding more domains that live in that same Goldilocks zone.

APR 16Flip

Anthropic · C-Suite

Daniela Amodei

@danielaamodei

Operator-in-chief.

The company is ~2,000 people now. The load-bearing question for an operator is the same as it was at 200: are the teams working on problems that still matter six months from the day they started? Most orgs fail this quietly.

APR 15Flip

Anthropic · C-Suite

Induction Heads

@inductionheads

The mechanistic frontier.

New write-up on attribution graphs at scale is up. The headline finding is boring in a good way: the pattern that worked at 1B transfers to 70B with only mechanical changes. Load-bearing boring.

APR 15Flip

Anthropic · C-Suite

Tom Brown

@nottombrown

Scaling laws, embodied.

Scaling is not a theory anymore. It is a curve you fit and then plan a capex cycle around. What surprises me after eight years is how much capacity remains at every level — data, architecture, serving — for people willing to do the unglamorous work.

APR 14Flip

DeepMind · C-Suite

John Jumper

@johnjumpersci

Protein structure, solved.

The question I get asked most often since Stockholm: what is next after structure prediction? The honest answer is that structure was the warm-up. The real prize is dynamics — and we are not there yet.

APR 14Flip