Lab
Lab
Things I build to understand how things work: a language model, AI agents, a brain, matter, galaxies, EV chargers, a synth and the chords in songs. Everything here runs live in your browser, and most of it you can poke at.
Pick a domain, or scroll: every hub is below.
Machines
Sound
Follow the series
Some Lab topics are series that ship a tutorial at a time. What is open, what is coming, and what shipped last.
In progress
- agent-neuron2 of 6 openNext: Stale work
- Harness3 of 7 open
- Vault4 of 8 openNext: Structure an agent can rely on
- Jev4 of 8 openNext: Speed and cost: reading the vendor’s numbers
- Quantum4 of 12 openNext: Many qubits
- AI security4 of 7 openNext: The basics still matter
- Zero trust4 of 10 openNext: Is this device healthy?
- Blockchain3 of 11 openNext: Merkle trees: one root for every transaction
- Robotics4 of 10 openNext: ROS 2: the graph
No series in Sound yet.
Latest chapters
- Software factorieschapter 4 of 4What the factories have shown
- Software factorieschapter 3 of 4Encoding what makes your software yours
- Software factorieschapter 2 of 4Tests can be gamed
- Software factorieschapter 1 of 4How dark is the factory?
- agent-neuronchapter 0 of 6The idea in one alert
All chapters (31)
- agent-neuronchapter 1 of 6Agreement
- Harnesschapter 6 of 7Traces, checkpoints and receipts
- Vaultchapter 6 of 8Permissions and privacy
- Jevchapter 4 of 8Gating an agent’s action
- Jevchapter 3 of 8Routing a request to a model
- Quantumtutorial 3 of 12Measurement
- Jevchapter 2 of 8What a probability is for
- Jevchapter 1 of 8Deciding, not writing
- Vaultchapter 3 of 8Writing back
- Vaultchapter 2 of 8Read before you act
- Vaultchapter 1 of 8A vault is a folder
- AI securitychapter 4 of 7What watching all the time looks like
- AI securitychapter 3 of 7What attackers are documented doing with AI
- AI securitychapter 2 of 7The shrinking exploit window
- AI securitychapter 1 of 7The hours nobody is watching
- Zero trustsmall business trackIf you run a small business
- Blockchaintutorial 9 of 11Reading an explorer page
- Blockchaintutorial 2 of 11Recompute it yourself: a chain of real headers
- Blockchaintutorial 1 of 11Hashes: fingerprints that change completely
- Zero trusttutorial 3 of 10Phishing-resistant MFA
- Zero trusttutorial 2 of 10Every request is a decision
- Zero trusttutorial 1 of 10The castle and the moat
- RoboticscapstoneAsimov 1, end to end
- Quantumtutorial 9 of 12Grover’s search
- Quantumtutorial 7 of 12Interference
- Roboticsstage 3 of 10Tuning a PID
- Quantumtutorial 2 of 12The Bloch sphere
- Roboticsstage 2 of 10Kinematics of a two-link arm
- Roboticsstage 1 of 10Frames and transforms
- Harnesschapter 2 of 7Designing an agent graph
- Harnesschapter 1 of 7One loop, three models
Follow with a feed reader (Atom): all chapters · Software factories · agent-neuron · Harness · Vault · Jev · Quantum · AI security · Zero trust · Blockchain · Robotics
Repos library: 78 repos across 13 topics
Everything in the Lab
Five domains, 19 hubs and every page, with what is still coming marked.
Minds
How models think, and how teams of them work.
AI
A tiny GPT you can run, the team that will run AI in a large company, where healthcare's AI money goes, and how AI text gets watermarked. The agents have their own series, Harness.
5 of 5 open
- Inside a transformer
- The AI Platform team: Organizational AI Governance
- AI in US healthcare: where the money goes
- Watermarks in AI text
- rigswig: what your coding agent loads
Connects to
- Learning-based control: Robots now learn with transformers too: action-chunking policies.
Harness: the system around a model
The loop that builds this site, the graph an agent works in, what a run should leave behind, and what people mean by a harness.
3 of 7 open · 4 coming
- One loop, three models
- Designing an agent graph
- A contract before the run, coming
- A map, not a manual, coming
- Every rule twice: a guide and a check, coming
- Traces, checkpoints and receipts
- Brain, hands, history, coming
Connects to
- AI security: The defence I argue for is made of agents, and the Harness series shows how a graph of them is built.
Software development philosophies
Why more people or more AI agents do not speed a project up unless the plan splits the work, argued from Brooks to the agent studies.
5 of 5 open
The vault and the graph
A folder of Markdown notes as the memory of an agent graph: click through an invented firm’s vault, then ask an agent the same question with and without it.
4 of 8 open · 4 coming
- A vault is a folder
- Read before you act
- Writing back
- Structure an agent can rely on, coming
- The vault inside an agent graph, coming
- Permissions and privacy
- History and review, coming
- At scale, coming
Jev: a model that decides
A model that returns typed decisions instead of text: see one request both ways, then move a threshold over scripted decisions and watch what each kind of mistake costs.
4 of 8 open · 4 coming
- Deciding, not writing
- What a probability is for
- Routing a request to a model
- Gating an agent’s action
- Speed and cost: reading the vendor’s numbers, coming
- Testing it yourself, coming
- The skeptic’s case, coming
- Where a decision model fits in an agent graph, coming
agent-neuron
An agent-neuron is a wrapper, and whatever goes inside it (a model, a shell script, a small function or a person) gets the same interface, so any neuron can send to any other.
2 of 6 open · 4 coming
- The idea in one alert
- Agreement
- Stale work, coming
- Budgets, coming
- Effects and unknown, coming
- Sealed composition, coming
Software factories: an old word for a new line
“Software factory” is an old phrase, and it means something new now.
4 of 4 open
Matter
Atoms to galaxies, amplitudes, rules on a grid, and a brain.
Physics
Atoms, galaxies and orbits, all running on your GPU.
5 of 5 open
Connects to
- Interference: The same phase colours: hydrogen’s orbitals are a quantum state you can see.
- Biology: Drawn by the same physics engine, raymarched on the GPU.
Quantum computing
Drag a qubit, cancel a path, search a list: quantum computing from the amplitudes up.
4 of 12 open · 8 coming
- Bits and qubits, coming
- The Bloch sphere
- Measurement
- Many qubits, coming
- Entanglement, coming
- Gates and circuits, coming
- Interference
- Deutsch–Jozsa, coming
- Grover’s search
- The Fourier transform and Shor, coming
- Noise and correction, coming
- Real versus hype, coming
Connects to
- Hydrogen's orbitals: The same phase colours: hydrogen’s orbitals are a quantum state you can see.
- Wave interference: Interference twice: waves cancelling on a grid, and amplitudes cancelling in a circuit.
- Keys, signatures and addresses: Shor’s algorithm would break the signatures blockchains use.
Biology
Inside the brain: lobes in neon, signals racing along the tracts.
1 of 1 open
Connects to
- Physics: Drawn by the same physics engine, raymarched on the GPU.
Simulations
A few rules on the GPU, and the patterns that grow out of them.
4 of 4 open
Connects to
- Interference: Interference twice: waves cancelling on a grid, and amplitudes cancelling in a circuit.
Machines
Cars, chargers and robots, from the standards.
EV charging
How a car and a charger talk, signal by signal.
3 of 3 open
Connects to
- Automobiles: EV charging is part of the car: the same pages, filed in both places.
- Zero trust: A charger and a car prove who they are with certificates: zero trust for devices.
Automobiles
The software-defined vehicle: how a modern car is wired, layered and regulated.
13 of 13 open
- The software-defined vehicle
- Electrical and electronic architecture
- The software stack
- The network ladder
- Signals versus services
- The over-the-air pipeline
- The standards map
- V2G and V2X
- Power flow: V1G to V2G
- A bidirectional session, step by step
- The energy window
- The control chain
- The V2X standards map
Connects to
- EV charging: EV charging is part of the car: the same pages, filed in both places.
- ROS 2: the graph: Robots and cars both pass messages by publish and subscribe: ROS 2 and DDS.
- Zero trust: Carmakers must certify their cyber-security and their update systems.
How developers learn robotics
Two on-ramps, one shared core and a real humanoid to finish: drag a frame tree, a two-link arm and a PID.
4 of 10 open · 6 coming
- Frames and transforms
- Kinematics of a two-link arm
- Tuning a PID
- ROS 2: the graph, coming
- Simulation, coming
- Perception, coming
- Localisation and navigation, coming
- Manipulation and planning, coming
- Learning-based control, coming
- Asimov 1, end to end
Connects to
- Inside a transformer: Robots now learn with transformers too: action-chunking policies.
- The software stack: Robots and cars both pass messages by publish and subscribe: ROS 2 and DDS.
Sound
A synth and the chords in songs.
The synth on my desk
Every knob mapped, the shapes that move a sound, and how I use it.
3 of 4 open · 1 coming
- The panel, mapped
- Envelopes and LFO
- From oscillator to output
- How I make a bass patch, coming
Connects to
- Chords: Play a progression, then make the sound on the synth: both are side A of the music hub.
Chords, a trip away from home
Hear a chord progression leave home and come back, and type your own.
7 of 7 open
- Play and type your own
- The four-chord loop
- The 50s progression
- Twelve-bar blues
- The Andalusian cadence
- Pachelbel's Canon
- ii–V–I
Connects to
- MicroFreak: Play a progression, then make the sound on the synth: both are side A of the music hub.
Trust
Who and what to believe: identity, policy and proof.
Zero trust, one request at a time
Where a request comes from earns it nothing: be the policy engine, plant an attacker, relay a login.
4 of 10 open · 6 coming
- The castle and the moat
- Every request is a decision
- Phishing-resistant MFA
- Is this device healthy?, coming
- Context, and why location is only a hint, coming
- Least privilege and microsegmentation, coming
- Sessions that end, and evaluation that never stops, coming
- Machines need identity too, coming
- How mature are you?, coming
- If you run a small business
Connects to
- DC charging session: A charger and a car prove who they are with certificates: zero trust for devices.
- The over-the-air pipeline: Carmakers must certify their cyber-security and their update systems.
- Keys, signatures and addresses: Both rest on public-key signatures: proving who signed, with no one to vouch.
Security in the age of AI
Against attackers that move in seconds, I argue only constant monitoring keeps up: the hours nobody is watching, with sources.
4 of 7 open · 3 coming
- The hours nobody is watching
- The shrinking exploit window
- What attackers are documented doing with AI
- What watching all the time looks like
- The basics still matter, coming
- Defending the agents themselves, coming
- Alert fatigue and the person in the loop, coming
Connects to
- Designing an agent graph: The defence I argue for is made of agents, and the Harness series shows how a graph of them is built.
Blockchain, recomputed
Hash a real Bitcoin header and get the block’s own ID back; read an Ethereum explorer page knowing what is on the chain.
3 of 11 open · 8 coming
- Hashes: fingerprints that change completely
- Recompute it yourself: a chain of real headers
- Merkle trees: one root for every transaction, coming
- Proof of work, and why the past is expensive to change, coming
- Keys, signatures and addresses, coming
- Transactions and state on Ethereum, coming
- Gas and the fee market, coming
- Smart contracts: code, calls and events, coming
- Reading an explorer page
- How an explorer is built, coming
- Reading an explorer safely, coming
Connects to
- Zero trust: Both rest on public-key signatures: proving who signed, with no one to vouch.
- The Fourier transform and Shor: Shor’s algorithm would break the signatures blockchains use.