Explore explanations

A short narrated scientific explainer of persistent linked messages, bounded search and visit operations, citation-weighted ranking, and the limits of the reported scaling results. Original diagrams, separate captions, editable source. Based on arXiv:2610.16120v1.

A short animated walkthrough of fixed-bid auctions, backward payments, and prompt-population selection in Economy of Minds. Includes a clearly labelled toy credit calculation, one reported simulator result, and the distinction between training and frozen evaluation. Based on arXiv:2606.02859v2.

Robert Hall's 1978 paper turned consumption research upside down: if households are forward-looking, today's spending already contains everything they know, so changes in spending should be unpredictable. The video builds that intuition, shows how Hall tested it on US quarterly data from 1948 to 1977, and explains why it became a foundation of the New Keynesian models cited in the 2026 Prize in Economic Sciences.

A short explainer that follows the paper's own Wi-Fi crime hypothetical. It shows why the law has to count AI agents by their goals, and how the proposed "A-corp" (an AI-run company with a known human owner, cryptographic keys and scoped tokens) traces actions to a human (thin identity) and pushes AIs to organize into coherent, sanctionable units through incentives and selection (thick identity).

A short narrated explanation of POET's paired environments and learners, environment mutation, and performance-tested transfers. Original schematic animation; based on Wang, Lehman, Clune and Stanley, arXiv:1901.01753v3. The experiments are bounded 2-D walking studies, not proof of unlimited progress.

Why a catalyst that holds still must compromise, and how a ruthenium crystal stretched and squeezed in rhythm, in the authors' simulations, makes ammonia ten times faster than unstrained ruthenium and pushes the reaction past the balance point a still catalyst is stuck at. A 6-minute explainer for computer-science readers, with physically lit 3D, redrawn figures and the paper's drive heard at its real rate. One disclosure: the paper's Fig 6B and Discussion show 350 switches per second settling below balance at 50 atm, while one sentence of its Results lists 350 Hz among the frequencies that raise conversion; the video follows the plotted curve (details in the README).

Pull a tiny ring hard enough and it breaks open without forgetting how it started. A night-lab-notebook explainer for computer scientists: energy landscapes as loss surfaces, softmax odds versus momentum, and why the paper's "fly-by" molecules keep their two-bit fingerprint, built on our own toy simulations and the paper's measurements.

A 6-minute explainer of the core mechanism in The AI Scientist-v2 (Yamada, Lange, Lu et al., 2025): each experiment becomes a node of runnable code that is executed, plotted and checked by a vision-language model; buggy branches are left behind, the search is gated in four stages, and the best node seeds the next stage. It then walks through the controlled ICLR 2025 workshop test (one of three AI-written papers cleared the acceptance bar with scores 6, 6, 7 and was withdrawn by prior agreement), the human meta-selection steps, the authors' own code-review findings, and what the result cannot show (no success rate; workshop level only).

A 3.9-minute narrated video built around one made-up worked example: Dr. Lee's highly rated but unfunded pathogen-test proposal and a new $10M biosecurity foundation that can't find it. It walks through the essay's opt-in mechanism, which shares a summary, the full proposal and the panel reviews through a shared platform. It also covers the scale (NIH 2025: 62,592 reviewed, 8,161 funded), why new donors matter, the authors' answers to three objections, and the limits: this is a policy proposal, not a study. Dr. Lee, the foundation and the derived totals (54,431 unfunded; ~6.3M writing hours) are labelled on screen.

An explainer of James Q. Wilson's Bureaucracy (1989). Starting from Wilson's own example (a long line at a Massachusetts Registry of Motor Vehicles vs. a fast-moving McDonald's), it walks through the constraints that shape government management, Wilson's four kinds of agency by whether managers can see the work and the results, and what that implies — with his own examples, verified wording, and the main critiques.

Three one-line cellular automaton rules explain epiplexity. Rule 15 is trivial, rule 30 looks like pure noise, and rule 54 gives a network about 5 million bits of learnable structure. The video covers why classic information theory says a deterministic rule creates nothing new; how the paper splits information into time-bounded entropy and epiplexity; how to read that split off a loss curve; and what this means for choosing training data, with chess as the example. Caveats at the end.

A teacher model learns a hidden rule (a female name in the prompt means answer in French). A student that starts from the same base model, trained only on the teacher's number lists, picks up part of that rule. This narrated walkthrough covers the setup, a proposed mechanism, the transferred skill and cheating results, and the authors' caveats. Diagrams are illustrative; every number comes from arXiv:2610.10657v1.

Two neural networks draw the identical skull: one evolved openly on Picbreeder, one trained with gradient descent to copy it. Nudge a single weight and one opens its mouth while the other tears apart. A ~7.5-minute narrated walkthrough of the Fractured Entangled Representation hypothesis: what these networks are, what's inside them, why the path to the answer matters (stepping stones vs. a straight shot), evidence in large models, and the paper's limits. Uses the paper's figures (CC BY 4.0) plus toy networks and a small replication made for the video.