CONTEXT JAMMING

Field notes from inside the context window.

Context Jamming / Explainers

Interactive arXiv Explainers

Famous papers from AI and physics, rebuilt as interactive research instruments. Manipulate a core diagram from each paper, then open the full explainer for the argument, equations, and source notes.

  • N° 01 · 2020arXiv:2001.08361
    Interactive sampleSelect bottleneck
    Power-law scaling frontierThree descending curves represent parameter, data, and compute limits. The selected bottleneck is emphasized and moves the frontier point.POWER-LAW FITLOG RESOURCE →LOSS →
    ACTIVE BOTTLENECK · CLOSS ≈ 1.88PPL ≈ 6.6 · FIT-DERIVED

    Empirical Laws of AI

    Scaling Laws for Neural Language Models

    How model size, data, and compute collapsed onto a small set of power laws—and reorganized the logic of training language models.

    Open full explainer
  • N° 02 · 2025arXiv:2412.14014
    Interactive sampleChange dimension
    Dimension-dependent gravitational phaseThe hand rotates among one, i, minus one, and minus i for a bare sphere. Adding the observer fixes the result at minus i.1i−1−i−1D = 4
    GRAVITYi^(6) = −1PARTICLE(−i)^(3) = iPRODUCTobserver not included

    Euclidean Gravity

    Real observers solving some imaginary problems

    Why de Sitter’s state count rotates through i—and why the observer must enter the calculation before we decide what is being counted.

    Open full explainer
  • N° 03 · 2015arXiv:1512.01548
    Interactive sampleDrag endpoints
    Boundary interval and bulk geodesicTwo draggable boundary endpoints define an interval and a curved bulk geodesic inside the disk.BULK GEODESICBOUNDARY INTERVALKinematic coordinate planeA single point at theta and alpha represents the complete interval and geodesic, not a location in the disk.POSITION θSCALE αKINEMATIC POINT
    u = 0.45 · v = 2.00θ = 1.22 · α = 0.78ONE POINT = ONE COMPLETE GEODESIC

    Kinematic Space

    Tensor Networks from Kinematic Space

    How a tensor network became a map of every possible way to probe a universe.

    Open full explainer
  • N° 04 · 2026arXiv:2601.04603
    Interactive sampleSet risk, then run
    Adaptive classifier pipelineA harmless synthetic risk score passes through the protected model and activation probe. Higher scores recruit an exchange classifier before an allow or refuse decision.EXCHANGERISK 34PROTECTEDMODELACTIVATIONPROBEPASSEXCHANGECLASSIFIERALLOW0.55 PROBE + 0.45 CLASSIFIER
    PRODUCTION CONTEXT · 94.5% PASS / 5.5% ESCALATEREADYADAPTIVE COMPUTE · SYNTHETIC ROUTE

    Constitutional Classifiers++

    Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks

    How Anthropic made a stronger jailbreak defense approximately 40× cheaper by moving classification into the generation stream.

    Open full explainer