New releaseThe Anthropic Book ↗

19 chapters  ·  corpus-verified  ·  audio narration

CONTEXT JAMMING

Field notes from inside the context window.

Hire me

Labs · 25 doors · 25 pieces

The AI-dvent Calendar

Twenty-five pieces from this site, each behind a door on a graphics card. Open them in any order; the board powers up as you go.

0 / 25 powered on

Opens every door in sequence.

D01 Day 1 of 25

Interactive Explainer

Transformer Atlas

Two GPT-2 small attention heads share a pattern but have opposite causal effects. Zero-ablate one and measure.

Read the piece

All 25 pieces

The same pieces as plain links, in day order.

  1. Day 1: Transformer AtlasInteractive Explainer

    Two GPT-2 small attention heads share a pattern but have opposite causal effects. Zero-ablate one and measure.

  2. Day 2: Scaling LawsInteractive Explainer

    Kaplan's 2020 finding: loss falls on nearly straight log-log lines as model size, data or compute grow.

  3. Day 3: The Physics OriginThe Anthropic Book · N°01

    In 2000, seventeen-year-old Dario Amodei was cut from the US Physics Olympiad team. The book starts there.

  4. Day 4: Jared KaplanFounderFiles · N°006

    A holographic-gravity physicist ran dimensional analysis on the loss landscape, co-wrote the scaling laws, and co-founded Anthropic.

  5. Day 5: Constitutional ClassifiersInteractive Explainer

    Anthropic's jailbreak defense moved into the generation stream and got about 40× cheaper by watching the whole exchange.

  6. Day 6: The Ghost in the MachineOral History · GPT-3

    Fourteen voices from inside OpenAI recount how the GPT-3 paper broke AI orthodoxy and started the scaling era.

  7. Day 7: The Geometry TradeDispatch · N°008

    A December 2025 paper says networks collapse into the same low-rank geometry. Chip design becomes cartography.

  8. Day 8: Universal Weight SubspaceResearch Build

    Filter training data by whether each example's gradient points into the shared weight subspace, negative finding included.

  9. Day 9: Chris OlahFounderFiles · N°002

    Olah got hired on a blog, mapped circuits and features, and now asks what we owe the systems.

  10. Day 10: The Frame and the FilingResearch · Who Moves Markets

    Podcasters and newsletter writers shape how Anthropic is read; launches, scoops and filings are what move the price.

  11. Day 11: The Whisperer and the StackDispatch · N°009

    OpenAI hired a Hollywood whisperer; Anthropic built a stack. One bets on dinners, one skips the licensing room.

  12. Day 12: Claude ShannonFounderFiles · N°068

    A classified 1945 memo held the first statistical model of English. Frontier models still play Shannon's guessing game.

  13. Day 13: Kinematic SpaceInteractive Explainer

    A 2015 Stanford paper argues MERA tensor networks discretize kinematic space, where each geodesic becomes a single point.

  14. Day 14: Thought MoleculesLabs · Tool

    Local-first RAG over your Drive research docs that returns only verbatim, substring-verified passages, ready to paste back.

  15. Day 15: Juan MaldacenaFounderFiles · N°009

    His 1997 paper put a region's whole interior on its boundary. It's high-energy physics' most-cited paper.

  16. Day 16: Concrete Problems in AI SafetyInteractive Explainer

    Break the pipeline behind an office-cleaning robot and find five failures, from reward hacking to distributional shift.

  17. Day 17: Chinchilla ScalingInteractive Explainer

    Split a fixed compute budget between parameters and tokens, and see why 70B beat 280B.

  18. Day 18: Ilya SutskeverFounderFiles · N°005

    He scaled the transformer until scaling broke, then left OpenAI to build Safe Superintelligence Inc.

  19. Day 19: The Scaffolding Is the ProductDispatch · N°006

    Same GPT-5, three scaffolds, a 2.7× gap. Stanford's ARTEMIS beat nine of ten human pentesters at $18.21/hr.

  20. Day 20: Persona VectorsStudy

    Sycophancy, hallucination and malice are linear directions in activation space: measurable, controllable, partly preventable.

  21. Day 21: MathematicaFeature · The Eames Exhibit

    IBM asked the Eameses to explain mathematics. They built a machine room where it performs itself.

  22. Day 22: Demis HassabisFounderFiles · N°037

    Chess prodigy, game designer, neuroscientist: Hassabis turned game systems and brain science into DeepMind's agent architectures.

  23. Day 23: Tech Tarot CardsLabs · Deck

    A tarot deck for builders: founders, context windows and scaling laws as archetypes. Pattern recognition, not mysticism.

  24. Day 24: Dario AmodeiFounderFiles · N°024

    One move, repeated: make an opaque system legible before you trust it. Retina, scaling laws, Constitutional AI, Anthropic.

  25. Day 25: The Anthropic BookThe Anthropic Book

    A primary-source-anchored book on Anthropic's first five years, chapter by chapter, with PDF, EPUB and audio.