Writing · Platform & cloud-native

Notesfromthecluster

Long-form technical writing on Kubernetes, service mesh, CI/CD and the developer experience — worked examples, diagrams, and the occasional algorithm.

Engineering

Latest·08 AUG 2026·11 min read

Twelve free models just walked into our benchmark — three of them beat the frontier

We wired OpenRouter's free tier into our 7-task LLM harness, registered 13 models with full metadata, and ran a fair 5-iteration sweep across all of them. Ling 3.0 Tiny, Laguna XS 2.1 and Gemma 4 26B posted averages above 98 on a board that Kimi K3 leads at 90.5 — and the entire run cost us nothing.

softwaremachine-learningllmbenchmarkingopenrouter

Read post→

Software19 JULY 2026· 8 min read

The delta rule: linear attention for a million-token context

Full attention pays an n² bill that a 1M-token context can't afford. Linear attention swaps the bill for a memory you write to — and the delta rule is what makes that memory smart. Kimi calls K3's KDA a 'hybrid linear attention mechanism'; this is the family it belongs to, from the kernel trick to gated delta updates.

softwaremachine-learningllmsequence-models

The delta rule: linear attention for a million-token context

Software19 JULY 2026· 7 min read

We pointed our own benchmark at Kimi K3 on launch week

Our 7-task harness renders (or shows) what models actually generate, live and sandboxed. Running Kimi K3 against K2.7, Gemini and Codex broke the harness three different ways before it produced a fair table — here's the data, and what K3 is actually good at.

softwaremachine-learningllmbenchmarking

We pointed our own benchmark at Kimi K3 on launch week

Software19 JULY 2026· 7 min read

How Kimi K3 works: 2.8 trillion parameters, 16 experts awake

Kimi K3 is the first open 3T-class model: 2.8T parameters, a 1M-token context, and only 16 of 896 experts active per token. We take the architecture apart: Kimi Delta Attention, Attention Residuals, and Stable LatentMoE, plus what each one buys.

softwaremachine-learningdeep-learningllm

How Kimi K3 works: 2.8 trillion parameters, 16 experts awake

Software15 JULY 2026· 8 min read

A neural network just learns a shape

A classifier doesn't memorise your data, it learns a shape: watch one hidden layer bend a straight decision boundary into the curve between two interleaved spirals, and see what breaks when you starve it of units or crank the learning rate.

machine-learningneural-networksclassification

Physics13 JULY 2026· 8 min read

How a magnet makes up its mind

Cool a magnet through its critical temperature and a billion spins spontaneously agree on a direction nothing outside ever chose: the 2D Ising model, spontaneous symmetry breaking, and the universality that ties a magnet to boiling water and neural memory.

physicsstatistical-mechanicsphase-transitions

Mathematics12 JULY 2026· 2 min read

Chaos hides in one line of algebra

Turn one dial on a one-line population model and it forks from a steady state to a 2-cycle, a 4-cycle, then chaos: the period-doubling cascade, the Feigenbaum constant, and the same fingerprint hiding in the weather and a swinging pendulum.

mathschaosdynamical-systems

Software07 JULY 2026· 12 min read

DSpark turns speculation into a scheduler

DeepSeek's DSpark keeps speculative decoding fast by pairing a parallel drafter with a confidence scheduler that verifies only the prefix worth paying for.

aisoftwareinferencellm

DSpark turns speculation into a scheduler

Software07 JULY 2026· 5 min read

Backpressure is the system saying no

Queues, retries, rate limits, and stream demand signals are all ways of saying the same thing: a healthy system needs a controlled way to refuse more work before overload turns into collapse.

softwaresystemsqueuesdistributed-systems

Electrical Eng07 JULY 2026· 7 min read

How DRAM remembers a bit

The second memory-chip-series post goes inside one DRAM access: a capacitor leaks, a bitline is precharged, a row is activated, sense amplifiers turn a tiny voltage nudge into a full bit, and the row has to be restored before the data is safe again.

hardwaresemiconductorsmemorysystems

Engineering06 JULY 2026· 14 min read

The Memory Stack: Why AI Turned DRAM Into Strategy

Memory chips look like commodity parts until the whole computer has to wait for them. This is a tour of the memory ecosystem: the history, fabrication, architectures, HBM versus GDDR stacking, 2026 supply pressure, AI workloads, and the breakthroughs trying to move or shrink the bottleneck.

hardwaresemiconductorsmemoryai

Physics04 JULY 2026· 12 min read

Simulating a black hole

Ray tracing Schwarzschild geodesics in a fragment shader — how one differential equation produces the shadow, the photon ring, and Interstellar's lensed disk.

physicsgeneral-relativitysimulationwebgl

Hiking30 JUNE 2026· 12 min read

The Walker's Haute Route — Le Châble to Zermatt

The Swiss half of the great Chamonix-to-Zermatt traverse, walked hut to hut over seven days and seven passes: the Sentier des Chamois, the Prafleuri boulder-fields, the Pas de Chèvres ladders, Lac de Moiry, the Forcletta and the Augstbordpass, to a Matterhorn finish at Zermatt.

hikingtrail-guide

The Walker's Haute Route — Le Châble to Zermatt

Hiking29 JUNE 2026· 10 min read

The Overland Track in winter: a trail guide

Tasmania's great alpine walk in the quiet of winter: eight days from Cradle Mountain to Lake St Clair, walking through snow, summiting side peaks, and staying in the empty huts.

hikingtrail-guide

The Overland Track in winter: a trail guide

Software24 JUNE 2026· 6 min read

Robots that act by denoising

The same denoising trick that paints a cat from TV static can drive a robot arm. Instead of refining pixels you refine a trajectory of future actions, sculpting a random tangle of motions into a plan conditioned on what the camera sees. We work out why plain regression freezes a robot between two valid moves, why diffusion handles that naturally, and how the idea scales into the vision-language-action models on the frontier.

softwaremachine-learningroboticsdiffusion

Physics24 JUNE 2026· 8 min read

Remembering a qubit that forgets

Every physical qubit forgets within microseconds and can smudge in two directions at once, yet you can weld many forgetful qubits into one logical qubit that remembers, and spending more of them lowers the error rate, but only once each qubit is good enough. The trick is to measure the neighbours' parity and never the data. We build the surface code by hand, watch errors become chains whose endpoints light up, and see why Google's Willow result mattered.

physicsquantumquantum computingerror correction

Software24 JUNE 2026· 8 min read

The loop that beats attention

Attention won by reading every token at once, and the bill was quadratic compute plus a cache that grows forever. State-space models like Mamba go back to the recurrent loop everyone thought was dead: a fixed-size summary carried forward in linear time. We work out why the loop came back, the one change that made it competitive, the trick that keeps it parallel, and why its little finite memory is both the win and the catch.

softwaremachine-learningllmsequence-models

Software24 JUNE 2026· 9 min read

A frozen model gets smarter if you let it think longer

Freeze every weight in a language model so it can't learn a single new thing, then let it spend ten times as much compute thinking before it answers. It gets more questions right. The gain isn't knowledge, it's search and verification over the model's own working — and you can train the model, with nothing but a right/wrong signal, to do that thinking for itself.

softwaremachine-learningdeep-learningreasoning

A frozen model gets smarter if you let it think longer

Cosmology24 JUNE 2026· 18 min read

The Universe Might Be Inside a Black Hole

Several serious models put our universe inside, or on the horizon of, a black hole in a larger parent cosmos — and the Big Bang as that black hole forming. A survey of the real physics, then one speculative step further.

cosmologyastrophysicsblack holesrelativity

The Universe Might Be Inside a Black Hole

Software23 JUNE 2026· 5 min read

Backprop is just the chain rule

Training a neural network sounds mystical, but the engine underneath is one idea from first-year calculus: the chain rule, applied backwards through a computation graph and reusing its work. We trace a forward and backward pass through a tiny graph, see why we run it in reverse, and connect it to the downhill step that actually does the learning.

softwaremachine-learningcalculus

Software23 JUNE 2026· 5 min read

How to paint with noise

Image generators start from pure TV static and end with a photo. The trick that makes it possible is wonderfully sneaky: don't learn to paint, learn to remove a little noise, then run that backwards from static. We build the forward noising process step by step, see the signal-versus-noise schedule, and work out why predicting noise is such a clever thing to train.

softwaremachine-learningdiffusiongenerative-ai

Software23 JUNE 2026· 5 min read

Three rules build a computer

Conway's Game of Life has three rules and a grid of on-or-off cells, and nothing else. From that, gliders crawl, guns fire, and: astonishingly: you can build a working computer. We play with the rules, watch structure emerge that nobody designed, and follow the thread to its unsettling end: a system this simple can be impossible to predict.

softwareemergencecellular-automata

Electrical Eng23 JUNE 2026· 5 min read

What JPEG throws away

A JPEG is often a tenth the size of the raw image and you can't see the difference. Where did ninety percent of the data go? We follow an 8×8 block through the discrete cosine transform and quantisation to find out: JPEG sorts a picture by spatial frequency, then quietly bins the fine detail your eyes barely register.

electrical-engineeringsignalscompression

Mathematics23 JUNE 2026· 6 min read

Why everything becomes a bell curve

The bell curve turns up in heights, measurement errors, exam marks and sensor noise, and that is not a coincidence — it is a theorem. We work through the central limit theorem with a Galton board and a random walk: why adding up enough independent things, almost whatever they are, always lands on the same curve, and when it doesn't.

mathematicsstatisticsprobability

Physics23 JUNE 2026· 5 min read

Why time only runs forwards

The microscopic laws of physics run the same forwards and backwards, yet eggs never unscramble and gases never un-mix. We resolve the paradox with a box of mixing particles: the arrow of time isn't a force, it's a counting argument — there are overwhelmingly more messy arrangements than tidy ones, so disorder wins by sheer weight of numbers.

physicsthermodynamicsentropy

Electrical Eng23 JUNE 2026· 5 min read

Why your shower temperature oscillates

The scalding-then-freezing shower dance isn't you being bad at taps — it's a feedback loop fighting a time delay, and it has a name. We meet the PID controller behind thermostats, cruise control and showers, tune one live, and find out why a little lag turns sensible corrections into oscillation.

electrical-engineeringcontrolfeedback

Software22 JUNE 2026· 11 min read

A transformer reads everything at once

The transformer's one real trick is reading every token at once and letting each decide what matters. We put the whole machine on the bench — embeddings, positions, the residual stream, the feed-forward step — and work out why reading everything at once was such a departure, and why something so architecturally dull keeps getting smarter the more we feed it. With an interactive animation for every piece.

softwaremachine-learningdeep-learningtransformers

Software21 JUNE 2026· 7 min read

Repair is just solving in disguise

I gave a 160M-parameter Go model a reflect-and-fix loop so it could bootstrap past its own ceiling. It fixed zero bugs. The reason runs deeper than the code, and it taught me my model had been memorising all along.

machine-learningllm

Electrical Eng19 JUNE 2026· 6 min read

Filters from Poles and Zeros

A filter is two polynomials, and the roots of those polynomials are the whole story. Place a few points in the complex plane and you can read the entire frequency response straight off the geometry — no calculus required at the point of use.

eesignalsfilters

Software19 JUNE 2026· 5 min read

Learning by Rolling Downhill

Every neural network you've ever used was trained by the oldest trick in calculus: to minimise a function, walk downhill. The whole story of modern optimisers is a list of the specific ways plain downhill walking fails, and the patch for each.

softwaremachine learningoptimization

Cosmology19 JUNE 2026· 13 min read

The Fine-Tuned Universe

Nudge a handful of the constants of nature by a fraction of a percent and you get a sterile cosmos: no stars, no atoms, no chemistry. That apparent fine-tuning is either luck, selection from a multiverse, or a sign we are missing something deep.

cosmologyastrophysicsfine-tuningmultiverse

Physics18 JUNE 2026· 5 min read

Quantum Tunnelling You Can See

Quantum tunnelling isn't an energy loophole: a wave decays through a wall instead of stopping. Watch a packet split, and see the exponential that keeps the Sun lit.

physicsquantumquantum mechanics

Mathematics18 JUNE 2026· 5 min read

How the Leopard Got Its Spots

Two chemicals that do nothing but react and spread can break a blank sheet into spots, stripes, and mazes. Turing's last great idea was that diffusion, the great smoother, is also where pattern comes from.

mathsnonlinear dynamicspattern formation

Software18 JUNE 2026· 12 min read

Four ways to shrink a KV cache

A transformer's KV cache is a four-dimensional tensor, and every compression trick — quantisation, eviction, cross-layer sharing, linear attention — attacks one of its axes. Here is the tour, and the cautionary tale of a tiny code model whose accuracy fell 20 points because a smoke test never exercised the one axis that bites.

softwaremachine-learningllmtransformers

Software17 JUNE 2026· 5 min read

A* Search, Visually: the Heuristic Is the Whole Game

A* is not a clever algorithm so much as Dijkstra plus a bet about the future. The same code becomes Dijkstra, greedy best-first, or A* depending on one term in the priority key — and admissibility is the single property that buys optimality.

softwarealgorithmscomputer scienceai

A* Search, Visually: the Heuristic Is the Whole Game

Electrical Eng17 JUNE 2026· 12 min read

AM, FM, QAM: A Tour of the Modulation Zoo

Every modulation scheme is the same act — painting information onto a carrier — and they differ only in which property of the carrier you paint on. Plotted as a constellation, AM is a line, FM is a circle, and QAM is a grid.

electrical engineeringcommunicationsmodulationdsp

AM, FM, QAM: A Tour of the Modulation Zoo

Software17 JUNE 2026· 14 min read

B-Trees vs LSM-Trees: The Two Religions of On-Disk Data

Every database you use bets on one of two storage engines: B-trees (read-optimised, update-in-place) or LSM-trees (write-optimised, append-and-compact). The choice isn't about speed but about which kind of amplification you're willing to pay.

softwaredatabasesdata structuressystems

B-Trees vs LSM-Trees: The Two Religions of On-Disk Data

Physics17 JUNE 2026· 12 min read

Band Gaps Are Just Bragg Reflection

The forbidden energy gap that makes a semiconductor a semiconductor is not a quantum mystery. It is the exact same wave interference that makes an X-ray diffract off a crystal or a soap film show colour.

physicscondensed mattersolid statequantum

Physics17 JUNE 2026· 7 min read

Every Qubit Gate Is a Rotation

Quantum computing is taught as alien linear algebra. On the Bloch sphere it collapses to one homely fact: a qubit is an arrow on a ball, and every operation you can do to it is a rotation of that arrow.

physicsquantumquantum computing

Mathematics17 JUNE 2026· 12 min read

Every Wave Is a Circle: Fourier Series as Epicycles

We are taught Fourier as an integral to memorise. The geometric truth is older and stranger: any periodic signal, however square or spiky, is drawn by a stack of spinning circles — the same epicycles Ptolemy used for the planets.

mathematicsfourieranalysissignal processing

Every Wave Is a Circle: Fourier Series as Epicycles

Electrical Eng17 JUNE 2026· 7 min read

Every Wire Is an RLC Circuit: Why Your Digital Signal Rings

There is no such thing as a digital signal at the physical layer. The clean trapezoid you draw is a fiction; every trace is a distributed RLC network, and the ringing and reflections are the lumped RLC step response playing out at picosecond timescales.

electrical engineeringsignal integritycircuitsrf

Every Wire Is an RLC Circuit: Why Your Digital Signal Rings

Software17 JUNE 2026· 14 min read

How Python Dicts Really Work: Open Addressing, Probing, and the djb2 Lie

The dict is the most-used data structure in Python and almost nobody knows it uses open addressing, not chaining — so a single deletion leaves a tombstone, and a simple hash table is really a careful dance between load factor, probe sequences, and cache lines.

softwaredata structurespythoncomputer science

How Python Dicts Really Work: Open Addressing, Probing, and the djb2 Lie

Mathematics17 JUNE 2026· 11 min read

The Lorenz Attractor and the Limits of Prediction

Determinism does not imply predictability. Lorenz's three-equation toy weather model is fully deterministic yet unknowable past a horizon you can compute — a geometric fact, not an engineering failure.

mathematicschaosdynamical systemsphysics

The Lorenz Attractor and the Limits of Prediction

Physics17 JUNE 2026· 13 min read

Normal Modes: Why Coupled Things Beat, and When They Become Chaos

Two pendulums linked by a spring look hopelessly complicated. But there is a change of coordinates in which the same system is just two independent pendulums that never talk to each other. The mess is an illusion of the basis you chose.

physicsmechanicsnormal modeschaos

Normal Modes: Why Coupled Things Beat, and When They Become Chaos

Mathematics17 JUNE 2026· 4 min read

Phase Portraits: See a Differential Equation Before You Solve It

Most nonlinear ODEs have no formula, yet their fate is readable from the vector field. Poincare's trick: classify fixed points, limit cycles, and chaos by geometry.

mathematicsdynamical systemsdifferential equations

Phase Portraits: See a Differential Equation Before You Solve It

Electrical Eng17 JUNE 2026· 12 min read

Build a Software-Defined Radio in 100 Lines

A radio is no longer hardware. Once you sample the antenna fast enough, every classic RF block — mixer, filter, demodulator — becomes a few lines of arithmetic on a stream of complex numbers. The antenna is the last analog component.

electrical engineeringsdrdspradio

Build a Software-Defined Radio in 100 Lines

Algorithms14 JUNE 2026· 11 min read

Attention, From the Inside Out

Attention is just a weighted average whose weights the data computes by asking itself questions. A worked tour through scaled dot-product attention, temperature and sampling, and what a representative 46B-active / 1T-total Mixture-of-Experts spec actually means — with live matrices you can poke.

technologyalgorithmsmachine-learningdeep-learning

Cosmology14 JUNE 2026· 13 min read

Why the Night Sky Is Dark

In an infinite, eternal, star-filled universe every line of sight should end on a star and the night should blaze like the Sun. It doesn't, and the darkness is hard evidence that the universe is neither infinitely old nor static.

cosmologyastrophysicsolbers paradoxredshift

Electrical Eng10 JUNE 2026· 8 min read

PLL Design from First Principles

A phase-locked loop is a control system with a phase detector instead of a summing junction. The intuition you can build with the lab above is more durable than the textbook derivations.

technologyalgorithmsphysicselectrical engineering

Electrical Eng10 JUNE 2026· 7 min read

The Smith Chart is Geometry

What looks like a chart for radio engineers is really a Möbius transform drawn on the complex plane. A visual essay on why impedance matching is a question of circles, lines, and rotations.

technologyalgorithmsphysicselectrical engineering

Cosmology09 JUNE 2026· 6 min read

Dark Energy and the Runaway Universe

The expansion of the universe is speeding up, which forces a strange conclusion: roughly 70% of everything is a smooth, invisible energy that pushes space apart and resists almost every explanation we have.

cosmologyastrophysicsdark energyrelativity

Algorithms08 JUNE 2026· 23 min read

The Neural Network Zoo, Revisited

A guided tour through the Asimov Institute's Neural Network Zoo — every architecture from the poster, with intuition for what each one is actually for and an interactive SVG diagram for the major families.

algorithmsmachine-learningdeep-learning

Cosmology03 JUNE 2026· 12 min read

Gravitational Waves: Hearing Spacetime Ring

When two black holes spiral together they shake spacetime itself. In 2015 a detector with 4 km arms felt the ripple as a fifth-of-a-second chirp, stretching its mirrors by less than a proton's width.

cosmologyastrophysicsblack holesrelativity

Gravitational Waves: Hearing Spacetime Ring

Cosmology29 MAY 2026· 3 min read

The first three minutes

Within minutes of the Big Bang the entire universe was a fusion reactor. The helium it forged then, about a quarter of all ordinary matter, is still here, and it matches the prediction to the percent.

cosmologyastrophysicsnuclear physicsBig Bang

Cosmology23 MAY 2026· 13 min read

The Cosmic Microwave Background: the universe's baby photo

The oldest light in existence is a near-perfect blackbody at 2.725 K with hot and cold ripples of one part in 100,000 — and those ripples are sound waves frozen into the infant universe at the moment it turned transparent.

cosmologyastrophysicscmbearly universe

The Cosmic Microwave Background: the universe's baby photo

Cosmology17 MAY 2026· 13 min read

How Stars Live and Die

A star's whole life is one long stalemate between gravity pulling in and fusion pushing out — and whether it ends as a white dwarf, a neutron star, or a black hole is settled almost entirely by the mass it was born with.

cosmologyastrophysicsstarsnucleosynthesis

Cosmology09 MAY 2026· 13 min read

The Cosmic Distance Ladder

We never measure a cosmic distance directly. We bootstrap from parallax to Cepheids to Type Ia supernovae, each rung calibrating the next, all the way out to the expansion of the universe — and a 5σ crack in the topmost rung.

cosmologyastrophysicsdistance ladderhubble constant

Cosmology04 MAY 2026· 13 min read

Dark Matter: the Unseen Gravity

Galaxies spin as though they weigh five times more than every star and gas cloud we can see. Either most of the universe is matter that emits no light, or gravity itself is wrong — and the evidence keeps landing on invisible mass.

cosmologyastrophysicsdark mattergalaxies

Physics28 APR 2026· 12 min read

The Universe on a Surface

The most information you can pack into a region is set by its boundary area, not its volume. In a deep sense the contents of a 3-D space are written on a 2-D surface, like a hologram.

physicsblack holesinformationquantum gravity

Physics16 APR 2026· 12 min read

Anatomy of a Black Hole

A black hole is mostly empty geometry. The only real surface is the event horizon, and every feature worth a name — photon sphere, ISCO, shadow, Hawking glow — falls out of one length scale, the Schwarzschild radius.

physicsastrophysicsblack holesgeneral relativity

Physics11 APR 2026· 6 min read

How Space Itself Expands

Distant galaxies aren't flying through space away from us. The space between them is stretching, governed by a single number whose entire history one equation predicts.

cosmologyastrophysicsrelativityphysics

Physics31 MAR 2026· 13 min read

Why Gravity Is Geometry

Gravity is not a force that pulls objects together. It is the shape of curved spacetime, and a falling object is simply coasting in a straight line through a bent geometry.

physicsrelativityspacetimegeometry

Service Mesh04 JULY 2020· 1 min read

Istio "Service Mesh" Topologies

Overview of some common Istio Service Mesh topologies and configuration examples and diagrams

technologykubernetesservice meshistio

Google Cloud24 SEPT 2019· 1 min read

GKE Auto Scale Down for Fun and Profit

How to sleep at night when actively developing on Google Kubernetes Engine (GKE) using Google Cloud Scheduler

technologyGCPkubernetes

Kubernetes22 SEPT 2019· 12 min read

Kubernetes-centric Continuous Delivery - Part 2 (Tekton Pipelines)

An overview of the current kubernetes-native ci/cd ecosystem, With a demonstration of an end-to-end Tekton based workflow for build, test and deployment of a fullstack application

technologyCI/CDGCPkubernetes

Kubernetes23 AUG 2019· 17 min read

Kubernetes-centric Continuous Delivery - Part 1 (Developer Experience)

How we can get closer to unifying the local developer experience with production-grade kubernetes run-time environments

technologyGCPkubernetescontainers

Algorithms28 MAY 2019· 4 min read

Two Number Sum

Three solutions to Two Number Sum, from O(n^2) brute force to an O(n) hash set to an O(1)-space two-pointer scan — and the memory-vs-compute trade-off behind each.

algorithmspersonalgopython

CI/CD · Tekton13 MAY 2019· 9 min read

Getting started with Tekton

Developing basic CI and CD pipelines in Tekton and deploying this both to local and cloud based environments

technologyCI/CDGCPkubernetes

Local Dev06 MAY 2019· 5 min read

Install Minikube

Instructions for the installation, upgrade and other customization of a Minikube install on MacOS. As well as a basic example of deploying a service and exposing it for local access

technologykubernetesdeveloper experience

Kubernetes30 APR 2019· 3 min read

Helm use-cases

A practical look at Helm, the Kubernetes package manager — what a chart is, how Go templating and Sprig drive value substitution, how releases and revisions work, and when to reach for Helm over Kustomize

technologykubernetesdeveloper experience

Kubernetes30 APR 2019· 4 min read

Exploring Kapitan

Exploring Kapitan as a generic configuration management tool: its reclass-style inventory, multiple input types, and the single-source-of-truth model that compiles one set of parameters into many outputs

technologykubernetesdeveloper experience

React21 APR 2019· 2 min read

Common Hooks

Working React Hooks examples: a stale-closure-proof counter and a custom useResizeObserver: plus when useReducer replaces Redux and why ResizeObserver beats IntersectionObserver.

technologyreact

Field Notes17 APR 2019· 1 min read

Hello World

The first post on Ben Ebsworth's blog: a living, personal-wiki space for exploring technologies, capturing learnings, and refining ideas over time.

generalpersonal