Live from the research notebook
The research notebook
The full archive from bretkerr.substack.com — 23 posts, newest first.
Archive
- Nov 4, 2026Read Against the Grain ↗
Hassabis read Gödel, Escher, Bach at 17 and took the opposite lesson from the one Hofstadter wrote. It became DeepMind, a Nobel, and Alphabet's chief scientist. The author is terrified.
- Nov 3, 2026Everything Converges ↗
AI’s convergence isn’t a philosophy story — it’s a cost-of-scale story. Two papers find the same small structure inside models, worth 42× fewer training samples and 100× less storage.
- Nov 3, 2026The Whisperer and the Stack ↗
OpenAI hired a person to win over Hollywood. Anthropic built a stack. One buys access but bets on dinner invites; the other buys credibility but skips the room where licensing deals get made.
- Nov 3, 2026Tent 39 ↗
McDonald's table tent is a $16.50 part corporate mandates and franchisees buy one checkout at a time. Table service fixed the crowd McDonald's kiosks made, years before Chick-fil-A went national.
- Nov 3, 2026The Head Table ↗
The midterms can change who gets to ask Anthropic questions, not who gets to answer them. Every lever that binds ANT — a 2–1 appeals-court ruling, a DOJ appeal, an FTC demand, export authority — stays
- Nov 3, 2026The Frenemy Loop ↗
AI labs are building a second channel beside the big consultancies, and the consultancies bought in. AI kills the analysis third of the fee. The name the board trusts just got more valuable.
- Nov 2, 2026Math Went First ↗
Kaplan's physics clock, Tao's brakes, and the bottleneck nobody priced. A correlation of The New Yorker (Galchen, Nov 1, 2026) and Quanta (Wolchover, Jan 26, 2026).
- Nov 2, 2026The Bridge Toll: what Anthropic's amended S-1 actually priced ↗
At 2:29 on Monday afternoon, a LinkedIn post from the angel investor Guillermo Flor was four hours old and climbing.
- Nov 1, 2026Filling the Hull ↗
A Harvard physicist and Claude found dozens of fields sitting on problems another field had solved. Finding the match is now cheap. Judging which one matters is the human bottleneck, and it tightens.
- Nov 1, 2026The Ledger and the Lunch Table ↗
On Wednesday, October 28, an admin at any Claude Enterprise customer could open Analytics, click “New report,” pick a team, set a window of up to the last 28 days, tick chat, Claude Code and Cowork, and walk away.
- Oct 31, 2026The Frame and the Filing ↗
Who moves Anthropic’s market? Podcasters and testers frame the story; launches, scoops, and filings move the price. Its 80-person comms team feeds the first group when the second makes news.
- Oct 31, 2026The house has no floors. ↗
Claude can write the chip, but verification, not drafting, is the bottleneck, and better design is new supply. At ~$500M vs. a $125.2B lease, the chip program is a 0.40% bet on the spend it can move.
- Oct 31, 2026Kaplan's habit: find where the law breaks ↗
Jared Kaplan's 2017 quantum gravity work and the 2020 scaling laws share one habit: fit the exponent, find where it breaks. Real but moderate continuity, and one Kaplan has never claimed.
- Oct 29, 2026Who Gets to Say No? ↗
Anthropic wrote the constitution. Palantir built the operating system. The Pentagon wanted control of the switch. Their alliance exposed the unresolved sovereignty problem at the center of agentic AI.
- Oct 29, 2026Two clocks to AGI: Ant Douglas & Marwell ↗
AGI’s real threshold isn’t a benchmark—it’s how long an agent can do expert work unaided. Even if reached by 2028, compute, institutions, safety and income may delay its impact into the 2030s.
- Oct 29, 2026Stop engineering the conclusion. ↗
Learned segmentation, shared weight geometry, and clear neurons hint at deep structure. But they don't prove one mechanism. The real breakthrough will be the experiment connecting all three.
- Oct 28, 2026The Human-in-the-Loop Mandate: Scaling Editorial Autonomy Without Yielding to AI Slop ↗
Unvetted AI text creates review bottlenecks. Requiring creators to curate model output turns passive slop into high-leverage agentic orchestration.
- Oct 27, 2026The Card is the Cheap Part ↗
Forget WeChat: X Money trades perks like 6% yields for your financial relationship. The metal card is cheap bait; the real gamble rests on X’s user retention and data playbook.
- Oct 27, 2026Ten Sentences, Two Readings ↗
Anthropic’s S-1 promises an “age of miracles,” then names the same engine as a risk. Across ten revealing phrases, “closing the loop” appears twice: as the moat on page 13 and a liability on page 30.
- Oct 26, 2026The Blanks and the Shepherds ↗
Same day, two disclosure regimes: Anthropic's S-1 blanks close under securities law; OpenAI's three claimed breakthroughs wait on eight outside shepherds. The timing is unproven. The contrast isn't.
- Oct 26, 2026The Meter and the Mortgage ↗
Anthropic sells intelligence on a 28-day meter and buys it on contracts that run to 2045. Its S-1 reveals a profitable core and suppliers who double as its lenders, sales channel, and rivals.
- Oct 25, 2026Poison, gates and the price of a pirated library ↗
Creators have won leverage through gates, contracts and courtrooms, not through demonstrated corruption of production models. Whether Glaze and Nightshade-style poison works at frontier scale is unproven in both directions.
- Oct 25, 2026Who Trains the Director? ↗
In the fall of 2024, in six river wards of Kogi State, Nigeria, people climbed trees and onto rooftops to take a phone call.