Skene Agentic
The library

22 explainers, each one running in the page.


Built one at a time, mostly to find out whether I understood the thing. They are grouped by the cell they belong to on the hub.

‹ Back to Hive

Agentic Systems

jsx

Recursive Agent Optimization

One policy acting at every node of a dynamic delegation tree (arXiv 2605.06639), with an interactive execution tree and a lambda slider.

Run it →
jsx

Pharma QC Pipeline

A working six-agent document-review app for pharma copy that ingests .docx and scores findings by severity.

Run it →
html

Under the Hood — a Single Co-authoring Turn

The five-layer architecture beneath one chat turn, drawn out.

Run it →
html

Agent Ecosystem

A browsable interface over the fifteen catalogue entries, filterable by kind and category.

Run it →

Attention

jsx

One Token’s Journey Through the Stack

Vaswani et al. (arXiv 1706.03762) followed end to end — one token from embedding through positional encoding, eight heads, the residual stream and cross-attention, with a live attention playground.

Run it →
jsx

Query, Key & Value

An interactive walk through the three self-attention projections, from intuition to the full scale-and-softmax pipeline.

Run it →
jsx

Attention Sinks Induce Gradient Sinks

Stepped paper explainer (Chen & Yao, 2026): attention mass piling on the BOS token creates a matching sink in the backward pass.

Run it →

Deep Learning

jsx

Deep Learning in Neural Networks

How stacking layers builds a feature hierarchy from edges to objects, and why depth beats width.

Run it →
jsx

Deep Learning, Explained by a Chicken, a Pug, and You

A field guide comparing shallow and deep “brains” to explain what the word deep actually buys you.

Run it →

Agent Harnesses

jsx

Meta-Harness

End-to-end optimisation of model harnesses (arXiv 2603.28052), with feedback compression set against OPRO, TextGrad, GEPA and AlphaEvolve.

Run it →
html

Quantum Maze — Stage 2: Fork / Join Harness

Replays blind and shared fork/join agent traces through the maze harness, side by side.

Run it →

Memory

jsx

Learning to Memorize at Test Time

Five tabs on Google’s Titans (arXiv 2501.00663): three memory systems and a surprise-driven rule for what gets remembered.

Run it →

World Models

jsx

Learning to Minimize Surprise

Friston’s free energy principle and active inference in five tabs, with a live Bayesian-inference simulation.

Run it →

RL

jsx

1000-Layer Networks for Self-Supervised RL

NeurIPS 2025 (arXiv 2503.14858): residual blocks, layer norm and Swish preventing signal collapse in 100× deeper RL critics.

Run it →

AI

jsx

Nemotron-Cascade 2

NVIDIA’s 30B/3B-active MoE reasoning model (arXiv 2603.19220): the post-training cascade, the benchmarks, the medals.

Run it →

Biology

jsx

The Beaver Keystone Species

How beavers reshape water, vegetation, biodiversity, firebreaks and carbon — ecology as a systems argument.

Run it →

Physics

jsx

The Physics of Gymnastics

Angular momentum, moment of inertia and projectile flight paths, read off real gymnastics skills.

Run it →

Coding

jsx

Bash, From Zero — Chapter 0

The four layers between you and the computer, the anatomy of a command, and a preview of the thirteen chapters after it.

Run it →

Emergence

html

Ants — Two Colonies Competing

A pheromone-trail simulation where foraging strategy emerges from purely local rules.

Run it →

Loop Engineering

html

Quantum Maze — Stage 1: Classical Agent

A single classical agent solving the maze — the baseline loop everything else is measured against.

Run it →

Tools and Connectors

html

Quantum Document Toolbox — Team 2 Review

A prototype executive-review interface driven off a mock HCP sales-aid deck.

Run it →

Robotics

html

A Robot Learns to Speak in Shadows

Yan, Wang, Liu & Chen (Duke, 2026): a 21-DoF soft-skinned hand learns a differentiable model of the shadow it casts, then performs hand-shadow puppetry against sixty-one target silhouettes.

Run it →