Words of crazy,
from the workshop.
Our corner for thinking out loud. Half of what lands here is documentation, half is us working an idea in public, and some of it is a hunch that had no business working — until it did. We build in the open; this is where we say what we found. New entries as they happen.
Mage-Flow, First Light
Microsoft shipped a 4B image model that does generation and instruction editing under an MIT licence — the licence being the part that usually kills these for client work. We had it measured on our own GPU the same evening. It beat its own published speed at half the advertised memory, and it renders readable text on embroidery, blueprints and screenprints. Eighteen images, one seed, no retries — plus the attention-backend trap that cost us an hour, written down so it doesn't cost you one.
Mage-Flow — The Full Sample Set
Every frame we generated while evaluating Mage-Flow: embroidered patches, cyanotype blueprints, risograph prints, a neon-noir alley, stained glass, and a three-step edit chain. One seed, first output kept, nothing cherry-picked.
Daily Dive — What Shipped in Open Models This Week
One day's pass over the open-model firehose: what's genuinely new, what only looks new, which releases fit real hardware budgets, and which are too big to be useful to anyone without a datacentre.
Two Thinking Tools — /office-hours and /diverge
Two tools we built to fight premature convergence: one runs YC-style forcing questions at an idea before anyone writes code, the other spawns isolated agents under distorted frames so nothing anchors on the first answer. What they do, when each one earns its keep.
Giving an AI Eyes on Video
How we let an AI actually watch a video instead of guessing from the title — scene-aware frame extraction, deduplicated shots, transcript from captions or local speech-to-text. The how-to, and what it costs to run.
Watch Brief — Signal from the AI Video Firehose
We pointed our own video-watching tool at the week's AI channels and pulled out the few claims worth acting on — including a speed-up we decided to go measure ourselves rather than take on faith.
Is a Mac Mini Really a Local-AI Machine?
A viral video claimed Apple just killed AI subscriptions. We pulled the claim apart against what the hardware actually does for local inference — where it holds up, where it doesn't, and who should genuinely buy one.
Field Notes from the Model Firehose
Hundreds of open models drop every week and almost none of them matter to you. How we triage what's actually worth running — licence, size, and whether it survives a same-prompt test on hardware we own.
audhd — Divergent Ideation for Coding Agents
Open source. Spawns isolated parallel agents under distorted cognitive frames to fight premature convergence, then scores and clusters what comes back. A fork of UditAkhourii/adhd that names the systematizing-critic half. MIT.
watch — Give Claude the Ability to Watch Video
Open source. Scene-aware frame extraction plus local-first speech recognition, so an agent can actually answer questions about a video instead of guessing from the title. Cloud transcription is opt-in, not required.
We Gave the Robots a Babysitter
An autonomous agent with a terminal will happily run rm -rf if a poisoned web page tells it to. So before any of ours runs a command, a world model dry-runs it in an imagined world and flags DESTRUCTIVE before it happens. The predict-then-act gate that catches rm -rf, agentjacking, and exfil — and the honest reason “we keep it wrangled” is a mechanism, not a slogan.
SPECVOICE — Speculative Execution at the Dialog Layer
Borrowing a trick from CPU design and applying it to conversation: start computing the likely answer before the caller has finished asking. What speculative execution looks like when the pipeline is a voice agent rather than a processor.
Two Years to the Taste of the Stew
The honest arc — from fighting Windows to a fleet of local models that finally feel real. The MCP tool-use moment, a memory layer a frontier model graded A− (and told us exactly where we're still behind), the month local inference got real, and a take on the buggy-whip crowd.
Fleet Log — Thirty Days in the Open
A month of building in the open, logged as it happened.
Scrappy Tools — Teaching Small Models to Use CLI Tools
Small local models can drive real command-line tools if you teach them properly. What worked, what didn't, and why tool-use is the capability that makes a small model useful instead of merely fast.
Three-Tier Memory — Breaking the 200-Line Limit
An AI coding agent forgets everything between sessions, and the usual fix — one big instructions file — collapses under its own weight. This is the architecture we landed on instead: three tiers, what each one is allowed to hold, measured results, and the tier we tried that failed.
The GPU Warranty Fight
A dead card, three companies, and a two-year paper trail. What actually happens when a warranty claim goes wrong — documented in full, because everyone deserves to know before they buy.
QB Agent — An AI Bookkeeper for QuickBooks Online
Invoices, expenses, reports and purchase orders through plain English, wired directly into QuickBooks Online. No dashboards to learn and no forms to fill out — you talk to it like you'd talk to a bookkeeper.
Multilingual TTS — Live Demo
Our text-to-speech running live in the browser, across languages. Type something and hear it back.
- /watch v1.3.0 Rebuilt our video-understanding pipeline with scene-aware frame selection — one ffmpeg pass, no ML. It matched a trending 950★ open-source tool on frame quality while keeping the instant, timestamped, zero-cost transcripts it doesn't have. More in the essay →