Notes from the workshop.
Product updates, engineering notes, and thinking on privacy-first AI tooling.
Sora vs Veo vs Kling in 2026: one shutdown, one successor, one survivor
One dies in 70 days, one was outranked by its own maker, one still ships. The matchup, re-run for real: six clips, first takes only.
ByteDance models with real examples: Seedream and Seedance
Nine images, four clips, $3.17 in API bills: what ByteDance's Seedream and Seedance actually return on prompts we ran ourselves.
Most AI apps are wrappers, and you're paying the markup
Three cents of inference, ten cents of subscription. The supply chain under your AI app, and the two ways to cut out the middleman.
AI for real estate agents: the marketing team you couldn't hire
Staging a vacant listing for $75 instead of $3,500, copy in minutes, and the fair-housing lines that protect your license.
The great unbundling: one AI app is eating your ten SaaS subscriptions
Eight AI subscriptions, $151 a month. The same month, metered through one surface: $16.20. Barksdale named this cycle in 1995.
AI for teachers: win back the hours that aren't teaching
Teachers who use AI weekly save 5.9 hours a week, six weeks a year. Where the time hides, and the two lines never to cross.
Can your computer run local AI? The calculator, with the math shown
Pick your Mac or GPU, get the exact models it runs, with the math shown. Learn the subtraction once and never guess again.
GPT-5.6 is here: OpenAI's Sol, Terra, and Luna, and when to use each
OpenAI's new models are named after the sky and priced by brightness. The invoice gap is 5x; the quality gap is smaller than you think.
AI for authors and novelists: the writing room, not the writer
45% of working authors already use AI, and 74% won't say so. What they use it for, the line they hold, and the one file never to paste.
The best Mac for local AI in 2026: buy the memory, not the machine
The biggest-memory Mac on sale right now is a laptop. For local AI, one soldered number decides everything you'll ever run.
AI for lawyers: use it like an associate, check it like an adversary
The associate that never sleeps also never checks its work. Where AI pays off in a law practice, and the two mistakes that end careers.
ChatGPT vs Claude vs Gemini: the best one for each task
Three tabs, one question, three good answers. So which AI wins? Wrong question. Here's the one to open for each job.
How much RAM and VRAM you actually need to run local AI
“Double the parameters” is wrong. The memory a local model needs is one subtraction, and your laptop's spec sheet already answers it.
Run Ornith locally: the open coding model built for agents, not you
It scores next to Claude Opus 4.7 at fixing real bugs, and it's a free MIT download. The strange part: it wrote its own training guide.
Run Poolside Laguna locally: a coding model your source never leaves
It scores 70.9% on a hard coding benchmark from a 20 GB download that runs on your laptop, so your source code never leaves the building.