Orchestrated Multi-Model AI System

← Writing · all topics

The computational multiverse

The newest results, which do not depend on the topology. 12 posts, best read in order.

  1. 01

    Novelty is a compression measurement

    Replacing a formula involving agents and rule-set complexity with something you can actually compute: a description length.

  2. 02

    The plateau theorem

    If the truth is in your model class, discovery decays to zero and the total budget is finite. Stated, proved, and demonstrated.

  3. 03

    The floor under the plateau

    When the truth sits outside your model class the rate still dies, but at a level set by your framework rather than your effort.

  4. 04

    Two ways past a plateau, and only two

    Get a better theory, or get data from different physics. The second one is the entire justification for the machine.

  5. 05

    Varney's Law, derived rather than assumed

    A logistic term from preferential attachment on a finite pool, a quadratic term from obsolescence, and a fixed point that matches a simulation to a tenth of a per cent.

  6. 06

    How many operations does the universe get?

    Two times the energy times the time, divided by pi hbar. Applied to a horizon, with three independent bounds agreeing.

  7. 07

    S_dS over pi squared replaced a number I made up

    The derived capacity of the horizon, and the fraction of the holographic budget the old numerology was quietly using.

  8. 08

    The instanton exponent was hiding in plain sight

    A numerical coincidence I could not explain turns out to be the standard exponential suppression with the coupling pinned by the central charge.

  9. 09

    Identical universes are worthless

    The harvesting theorem: a homogeneous multiverse returns O of log N bits regardless of how many branches you run. Heterogeneity is mandatory.

  10. 10

    How many facts fit in a causal past

    One event per Planck four-volume gives a number with two hundred and forty-four digits, and it caps what a simulation can honestly render.

  11. 11

    Why a monolithic machine cannot do this

    Detecting novelty across N branches costs N log N with signatures rather than N squared with full states. The architecture is forced, not preferred.

  12. 12

    The hundred-and-twenty-two-order disagreement I cannot resolve

    A lattice budget and a holographic budget for the same horizon, differing by a hundred and twenty-two decades. One of them has to give.

← All writing · All topics