Now
What we're building, right now
The always-current answer to what we're building and why — where the attention is, what ships next, and the roadmap behind both.
As of
This page updates as part of the weekly rhythm — when the focus changes, the page changes with it.
Focus now
Teaching personal AI
The direction ruled on 19 August: write, teach and explain personal AI — models specialized to a task, to a person, and to a device rather than to the average user — and build the things that help readers get there. It is one thesis at three radii, and it was ratified because the same three bets came back unprompted, in the same order, sixteen days apart. Everything below either teaches it or exists because of it.
A home for every piece, and the point of view on the page
The /ai hub and its five lanes — news, tutorials, reviews, reports, and a model registry — went live on 25 August, which is the first time work could be published where it belongs rather than into a generic feed. The half still being built is what shows on the page: where the knowledge came from (did it, near it, or only read it), what the best existing coverage leaves out, and what would prove the piece wrong. That last part is the differentiator, not the routing.
A production line for the visual work
Explaining hard things in video and diagrams needs a system, not a run of one-offs. August built one: an engine that takes an approved shot list and returns a per-video draft, a library of visual nouns so a diagram of a GPU is the same drawing every time, and a contract that names any element it cannot render instead of shipping a silently broken frame. The same discipline now covers diagrams, which are specs before they are pictures.
Shipping next
- The shared reading shell — one reading experience across every surfaceBuilding
- Standing, delta note and falsifier rendered on every published pieceBuilding
- Markdown twins, structured data and feeds everywhere — agents read what readers readBuilding
- The video engine — an approved shot list in, a per-video draft outVerifying
- Company coverage back in full — and the screener views that were quietly blindVerifying
Roadmap
Free instruments stay free — things on this list get built when demand pulls them, and every ship lands on the shipping log.
Next
Cost per completed task, measured not estimated
Tokens and pounds per finished task, instrumented on the agents that already run here every day rather than on an invented benchmark set. It is the gate on the public work that follows it, and it has not been run yet.
A way to actually buy a build
Local AI Deployment, Custom Fine-Tune and Harness Build are written, scoped and priced. The path to book one is not open yet — opening it is the next thing on the money side, and saying so is cheaper than pretending otherwise.
Soon
The buyer’s side of on-device AI
Should you buy a DGX Spark, a Strix Halo, or a Mac Studio — and what actually runs well on it, for which task, at what quality and what cost? Answering that honestly means owning the hardware and running real work on it, which is why nobody has. Not a build; a bench.
Later
Per-test result pages in the eval lab
The methodology page shipped first as the foundation. Individual test results build on top of it, once there are enough tests run to be worth browsing.
This page looks forward. The other half of the story is the shipping log: what already landed.