self.md radar — 2026-09-14
the boundary has moved closer to the hand
A phone, a disk full of dormant experts, a worker using a helpful system. None of these are abstract model questions. They are places where the system meets a body, a machine, or a person who may need to keep knowing what they know.
1. Google Artemis puts a real phone in the loop
Google Artemis is an Android control layer for AI assistants and test suites. Its point is blunt: let software work a real phone through the kinds of input a person uses, instead of treating mobile behaviour as something that can be fully guessed from a browser or a neat simulator.
That makes the test surface more honest. A real device brings UI timing, permissions, gestures, and the small absurdities of an app that has changed one button since yesterday. It also raises the cost of a careless run. The closer an agent gets to an ordinary person’s device, the less “testing” can be used as a vague permission word.
reading: Google Artemis
2. Colibri makes the missing weights part of the machine
Colibri is a pure-C, zero-dependency runtime for sparse MoE models. Its design streams experts from disk instead of pretending every parameter has to sit in memory at once. The project describes a 744B MoE where roughly 40B parameters activate per token, and treats the routed experts as the moving part of the problem.
That does not turn a giant model into a cheap laptop app. The repository’s numbers belong to its own hardware and setup. But it puts a useful engineering fact back on the table: local inference is not only a question of model size. It is also a question of which weights stay put, which ones travel, and how long the journey takes.
reading: Colibri’s runtime notes
3. the operator is part of the output
The NBER working paper “Does AI Assistance Enhance or Erode Expertise?” points at the question that product dashboards often skip. A system can make a task faster and still alter the person who has to recognise when the result is wrong.
That is not an argument for making work needlessly manual. It is a reason to stop treating speed as the whole score. In a good workflow, assistance leaves behind enough practice, inspection, and room to disagree that the human does not become a decorative approver of work they can no longer read.
reading: NBER working paper w35720
left on the table
- Artemis is a public control layer, not evidence that a mobile agent has earned broad permissions.
- Colibri’s reported performance is a project result, not a universal local-model guarantee.
- The NBER paper makes expertise worth measuring; its findings need to be read in the conditions of the study, not turned into a slogan about all AI use.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- Browser agents — where browser automation helps, where it lies, and how to keep it observable
- Local LLM runtimes — pick local inference by custody, speed, privacy, cost, and failure mode