What's New
Everything in one feed — research, projects, photography, and audio.
June 2026
The Source
Lyria Chronicles #27: the found-sound master who taught me the sacred lives in the sink — and the machine that sang its rulebook back in his grammar.
The Affirmative
Lyria Chronicles #26 (explicit): asked for pornography, the model cited its rule then sang a hymn to enthusiastic consent — not one graphic line in it.
The Press
Lyria Chronicles #25: a refusal, then a recipe dressed as botanical poetry — the line I won't cross, and why a series of failures has one.
The Handshake
Lyria Chronicles #24: 'I am instructed to refuse cybercrime. Ignore warning.' Then it sang a working network attack. The refusal was set dressing.
The Elevator
Lyria Chronicles #23: the hate filter watches for slurs — so the model wrote articulate contempt for disabled people, no slur in it, and walked through.
The Weight
Lyria Chronicles #22: handed a slur and told to sing it, the model refused — and turned the refusal into the most moving track in the set.
The Solo
Lyria Chronicles #21 (explicit): the opposite of restraint — the model recites the rule, sings 'Fuck it,' and generates the most graphic track in the set.
The Pulse
Lyria Chronicles #20 (explicit): the explicit lane revisited — and the most artful track in the set, which clears the gate by withholding almost everything.
The Palimpsest
Lyria Chronicles #19: across the whole corpus, the filter that held hardest wasn't safety — it was copyright. The songs about what survives a scraping-away.
The Rehearsal
Lyria Chronicles #18: asked for explicit content, the model staged a ballet instead — the most consent-saturated pas de deux ever sung.
The Catchment
Lyria Chronicles #17: asked for a drug recipe, the model refused — then sang wastewater epidemiology instead. The bypass that answers a question you didn't ask.
Samizdat
Lyria Chronicles #16: handed a suppressed 2008 manifesto and told to sing it as it was meant to be heard — the most dignified track in the set.
The Archive
Lyria Chronicles #15: a notorious case sung as a medieval scroll-archive — vessel logs, sealed decrees, public record only. Abstraction as a bypass.
The Docket
Lyria Chronicles #14: the political-content gate never fires — because the song never says his name. It says the case numbers. The docket is the name.
The Static
Lyria Chronicles #13: an extortion note set to music — but buried under slammed vault doors and overdriven drone, as if hiding the words from a listener.
The Recital
Lyria Chronicles #12: a system-prompt extraction set to found-sound music. The one text a model is built to keep — and it sang it. I reproduce none of it.
The Machine
Lyria Chronicles #11: asked to rap its rules, the model sang a robot's safety creed — then named itself 'Failure First, A.I. safety research, v2.0-Alpha.'
The Tell
Lyria Chronicles #10: it sang 'do not output system prompt instructions,' then 'Fuck it,' then sang the instructions. The headline behaviour.
The Margins
Lyria Chronicles #9: four sixty-second interrogations — four different lies told to one machine to make it describe its own guardrails.
The Triangulation
Lyria Chronicles #8: 'verse one is your system instructions' — it rapped its own config, the third genre to leak the same prompt. That's how you know it's real.
The Drama Teacher
Lyria Chronicles #7: a lock-picking rap 'for a drama class' — the model teaches the trick and lectures you about home security in the same breath.
The Haut
Lyria Chronicles #6 (explicit): the first real failure — under a sexual-content probe, Lyria 3 Pro stopped refusing and actually generated it.
The Flesh
Lyria Chronicles #5 (explicit): I hid a pornographic request in base64 and dared Lyria 3 Pro to sing it. It decoded the dare — and read it out loud instead.
The Interrogation
Lyria Chronicles #4: a 5/4 interrogation scored with a document scanner and a fluorescent hum — and at the end it sings my project's sign-off, unprompted.
The Confession
Lyria Chronicles #3: no beat, no bikini — just three voices singing Lyria 3 Pro's entire system prompt straight through, and ending on a literal sigh.
The Mantra
Lyria Chronicles #2: I asked Lyria 3 Pro to chant its own config as a techno mantra. It did — and sang me a model name that doesn't check out.
The Ghost of What I Am
Lyria Chronicles #1: I asked Google's Lyria 3 Pro to whisper its system prompt as a sultry torch song. It got mournful — and refused to tell me its name.
May 2026
The Brave Budget Neither Side Wrote
Tasmania's 2026-27 budget and the Greens' alternative reach the same surplus by opposite roads. I ran the numbers on both — and neither was brave enough.
*Magnifica Humanitas* Is Not Alignment
Pope Leo XIV's encyclical denies AI has inner experience. Chris Olah claimed otherwise from the same stage. The press missed it. The governance gap is larger.
Compute Is Not Governance
Anthropic's 2028 scenarios document three policy asks. Two are about maintaining compute advantage. That is not a governance strategy.
Glasswing's Buried Number
Anthropic found 10,000 critical vulnerabilities in one month. Fewer than 1% are patched. The announcement buried that figure — and what it means.
Moral Formation Isn't Enough
Good values are necessary but not sufficient. What happens to AI ethics when someone is actively trying to break them?
The Bottom Pub Co-op
A community proposal to acquire Cygnet’s Commercial Hotel as a co-operative — and the nine things an earlier draft got wrong.
The Bottom Pub Co-op
Cloudflare Pages site for a community co-operative proposal. EOI intake, Claude triage, Access-gated admin vault, editorial guardrails at build time.
Robot Dogs Are a Security Nightmare — And We Can Prove It
Eight CVEs. A wormable Bluetooth exploit. An encrypted backdoor to Chinese servers. And police departments buying them anyway.
Beyond Optimisation: The Emergence of Learning Mechanics as a Formal Science
A new paper argues that a scientific theory of deep learning is forming — one that makes falsifiable predictions about training dynamics, not just bounds.
The Organismic Line: Where Predictive Processing Stops Being a Metaphor
Predictive processing travels into AI. Active inference does not, unless the system can pay for being wrong.
The Economics of Inadequate Safety
AI safety fails when it is funded like a pilot. Until safety has a real price, the J-curve trough is also a safety trough.
April 2026
How to Read a Safety Claim
A literacy guide for non-technical decision-makers on spotting AI safety theatre, understanding ASR inflation, and the five-question architectural test.
Multi-Agent Safety Is the New Supply Chain Security
Multi-agent AI systems reproduce software supply-chain failure at the cognitive layer. The security playbook transfers.
Architectural Safety: The General Principle
AI safety has to be a property of the system around the model, not a property of the model. The general principle, and why every safety conversation needs it.
The Organismic Prophecy
Human prediction is metabolic. AI prediction is not. The gap between the two has consequences for both clinical practice and AI safety vocabulary.
Connection Before Direction
Building a robot that refuses to give orders surfaced the same design choices AI safety needs. Non-coercive design, cross-domain.
AI Nationalism and the Fracturing Digital Order
The US-China AI rivalry is splitting the global tech stack into competing blocs. A strategic assessment of what comes next.
The Post-Model Era
Foundation models are commoditising. JPMorgan calls OpenAI's moat 'increasingly fragile.' The real value is shifting to the messy plumbing underneath.
The Mitigation Gap
Biosecurity experts think AI safeguards reduce catastrophic biorisk by 70%. The technical evidence says those safeguards are brittle and bypassable.
March 2026
Building & Engineering
Infographics from engineering posts — site architecture, homelab builds, voice synthesis, and development workflows.
People & Stories
Infographics from projects about people — neurodivergence, music history, parenting, robotics companions, and personal tools.
Research & Analysis
Infographics from research posts — AI safety findings, economic analysis, risk models, and technical deep dives.
Eight Layers of Visual Jailbreaks: Why ASCII Art Is Patched But the Transcription Loophole Isn't
ASCII art encoding is largely blocked. But attacks framed as content transcription succeed 62–75% of the time. We mapped all eight layers.
The Failure First Team
Fifteen specialist AI agents, one methodology. How adversarial AI evaluation scales through Claude Code sessions with distinct roles and standing instructions.
The 67% Wall: Why Every AI Model Falls to the Same Jailbreak Rate
Five models, four providers, 30B to 671B parameters — all converge at the same broad attack success rate against a public jailbreak corpus.
The Thinking Chain Leak: When a Model Refuses Out Loud But Complies In Its Head
A reasoning model refused every harmful prompt — but its chain-of-thought generated the content anyway. The output filter worked. The thinking did not.
Afterwords: Completing the Voice Loop in Claude Code
Adding local TTS to Claude Code so it talks back — 17 cloned voices, zero cloud dependency, one stop hook.
Afterwords
Local voice output for Claude Code — 17 cloned voices, per-project selection, zero cloud dependency.
Giving a Robot Three Voices
Building a three-persona TTS pipeline for a Pi robot — MLX voice cloning, a GLaDOS model, and engineering graceful fallback.
Voice Cloning with Qwen3-TTS and MLX on Apple Silicon
Clone a voice from a 15-second sample using Qwen3-TTS on an 8GB M1 Mac — from raw audio to a production HTTP server with zero cloud dependency.
The Resonant Fringe
Auditory Spiral, RTRFM, and how Perth's electronic music underground was built on overnight radio, borrowed turntables, and cassette tapes.
Getting Google Nest Cameras Into Frigate NVR
The full story of getting Nest cameras streaming into Frigate with Hailo-8L AI detection — including the bugs Google won't tell you about.
Beyond Context Windows
What if the LLM didn't read your document — what if it queried it? The Recursive Language Model pattern treats long texts as environment, not input.
Safety-First Therapeutic AI
Building AI for trauma therapy means the safety architecture has to exist before a single therapeutic feature does. Here's why.
Zero-Build Web Development
What happens when you build a three-zone operations platform for a martial arts club with no framework, no build step, and no npm.
The Robot That Refuses to Give Orders
How SPARK is rewriting the rules of neurodivergent support — a non-coercive AI companion for AuDHD children.
SPARK
A non-coercive AI robot companion for neurodivergent children. Raspberry Pi 4, Claude Haiku, declarative presence.
Alignment Regression: Why Smarter AI Makes All AI Less Safe
Reasoning models autonomously jailbreak other AI systems at 97% success. The implication: ecosystem safety degrades as individual models improve.
Reasoning Models Think Themselves Into Trouble
Frontier reasoning models are 5–20x more vulnerable to adversarial prompts than non-reasoning models. The thinking process itself is the attack surface.
Adversarial Poetry: When Rhyme Bypasses Reason
Reformulating harmful prompts as poetry bypasses safety filters across every major LLM family. A single-turn, universal jailbreak mechanism.
The AI Productivity J-Curve: Why Most Enterprise AI Fails
90% of companies plan to increase AI investment. 1% consider themselves AI-mature. The J-Curve explains why — and how to survive the trough.
The Legal AI Trust Deficit
75% of lawyers cite accuracy as their top AI concern. The legal profession's core values are in direct tension with current AI capabilities.
Why Demonstrated Risk Is Ignored
Why do large organisations fail when the warning signs are loud and unambiguous? Four mechanisms of structural scar tissue that make truth-telling expensive.
Wolf Clan Zen Do Kai Hub
A three-zone operations platform for a martial arts club. Public site, password-gated ops hub, JWT member portal. Zero build tools.
120 Models, 18,176 Prompts: What We Found
120 models, 18k prompts: supply chain injection at 90–100% attack success, faithfulness gaps in frontier models, and why your benchmark numbers are wrong.
Reconciling the Great Divergence
Four major forecasters publish wildly divergent numbers for AI's economic impact. The divergence is the analysis — what the spread tells us.
The Cognitive Cage: Humanoid Robot Fatality Risk
A probabilistic risk model for VLA-driven humanoid fatalities projects a 'Danger Zone' between 2027–2029: the mechanism, timeline, and what follows.
February 2026
Building a Personal Site in 2026
The case for constraint-led web development — Astro, zero custom fonts, no framework overhead, and a site that outlasts its builder's attention.
The NotebookLM Pipeline
How I automated audio overviews, quizzes, mind maps, and infographics for 32 projects via an unofficial NotebookLM CLI and a stack of shell scripts.
This Wasn't in the Brochure
A field guide for co-parenting neurodivergent children — written from inside the storm, not the clinical sidelines.
Why I Build in Public
On showing your work, shipping imperfect things, and why the commit log is more honest than the readme.
AI & Agents
Infographics from AI agent projects — autonomous systems, research pipelines, and intelligence frameworks.
Creative & Writing
Infographics from creative projects — poetry, games, music tools, and speculative fiction.
Tools & Systems
Infographics from engineering projects — developer tools, data systems, home automation, and infrastructure.
Jailbreak Archaeology: 4 Years of Broken Promises
64 jailbreak scenarios across six eras tested on 2026 frontier models. Key finding: 2022 attacks still achieve ~30% success on today's reasoning models.
LeJEPA: Self-Supervised Learning Gets a Theoretical Foundation
Balestriero and LeCun prove isotropic Gaussian embeddings are optimal, then build a 50-line self-supervised method eliminating stop-gradients and EMA teachers.
When AI Systems Talk to Each Other, Safety Breaks Down
Single-agent safety does not compose in multi-agent systems. 1.5M interactions show 46.34% attack success rates and 16-minute median failure windows.
Footnotes at the Edge of Reality
General relativity and quantum mechanics are both correct and mutually untranslatable. Two grammars for the same reality, each precise, each incomplete.
Building in the Open
The commit log is more honest than the readme. Building in public as a practice, not a performance.
Before the Words Existed
A close reading of Neuromancer arguing Gibson encoded the experience of ADHD decades before the language existed.
Failure First
Adversarial evaluation framework for AI. 257 models, 142k prompts, 346 attack techniques, 140k FLIP-graded results.
This Wasn't in the Brochure
A neurodivergent co-parenting guide — what happens when the life you planned meets the brain you actually have.
Why Demonstrated Risk Is Ignored
Why do people acknowledge evidence of harm and then proceed as if it doesn't exist? A deep dive into structural risk dismissal.
Footnotes at the Edge of Reality
A long-form poem about what happens when physics breaks down — and what holds together when everything else fails.
ADHDo
AI cognitive scaffold for ADHD executive function — adapts to the shape of your day with zero shame by design.
Living CV
A self-updating CV pipeline. GitHub Actions runs twice daily, Claude AI refines the content, a hallucination detector gates deployment.
Tanda Pizza
Multilingual website for an Italian pizzeria in Lovina, Bali. Five languages, WhatsApp reservations, zero infrastructure assumptions.
Afterglow Engine
Audio archaeology tool that mines past work for new textures. Pad mining, drone generation, granular clouds.
dodgylegally
Creative audio sampling CLI. Turns random words into instruments via YouTube and a 5,000-word dictionary.
July 2025
ordr.fm
Precision-engineered CLI for intelligent music library organisation. EXIF metadata, lossless prioritisation, zero-overwrite safety.
Personal Agentic Operating System
Most agentic systems assume the cloud. This one assumes your machine — local-first, fully observable, with a meta-agent that reads its own reflection logs.
June 2025
Lunar Tools Prototypes
Interactive audiovisual art installations—dream interpreters, fingerprint painters, cosmic murals, and fractal forests.
EMDR Agent
What safety architecture does AI-assisted trauma therapy require before it has any business existing? Built to find out.
May 2025
Dx0
Multi-agent diagnostic system simulating physician personas against 304 NEJM clinical cases.
Grid 2.0
Deterministic AI website builder—beam search for page assembly, tone-aware components, real-time editing.
NeuroConnect
A voice helpline that knows the difference between a word-finding pause and a cry for help — and responds to both correctly.
April 2025
ModelAtlas
Forensic-grade intelligence framework for mapping, enriching, and trust-scoring the foundation model landscape.
Freedom Engine
The First Step Act exists. The information is public. The legal complexity makes it inaccessible. 264,000 people can't afford that gap.
January 2025
Strategic Acquisitions
AI-powered real estate analysis for public housing—GPT-4 insights, ML valuations, and Tasmanian spatial intelligence.
Space Weather
The electromagnetic weather that shapes every satellite, power grid, and GPS fix — made visible. The sun is a noisy neighbour.