self.md radar — 2026-09-05
the plan met the people who had to live inside it
one story today is about a company trying to turn “AI native” into an org chart. one is about making an analytics model choose fewer things. the third is a Windows control room named Ultron, which is at least honest about the genre.
1. Meta’s proposed 60% cut stopped before round two
Reuters’ investigation says Meta considered two rounds of cuts that could have reduced many teams by as much as 60%. The planned workforce was smaller human groups overseeing virtual workers; engineers were also reassigned into the Applied AI data-labelling unit.
The report says staff pushed back, and Zuckerberg called off plans for further cuts hours before a May layoff wave. That is a useful correction to the tidy version of “AI-native”: it can mean job redesign, forced transfers, and a spreadsheet that has to survive contact with the people on it. The Hacker News thread is a useful record of the report landing with engineers.
reading: Reuters’ reporting · discussion trail
2. The model reads the question; policy picks the query
A new MasterControl AI Lab paper draws a hard line: let a model turn a question into governed meaning, then let a deterministic policy select the pre-written analytical program. The output carries its evidence, not only a number.
Their test is deliberately fussy. “How many approved temperature-excursion deviations occurred last week?” has to preserve the approved population, calendar window, definition, and supporting records. Across 330 runtime-planning episodes with three independent 8B models, the authors report no procedure matching their whole answer-and-evidence contract; their policy-executed setup matched it in 110 of 110. The paper is explicit that this is a configuration-specific result, not a universal verdict on agents.
The cost is visible too: somebody has to write, approve, and maintain the policy catalogue. But when a number becomes evidence in a regulated system, that work is usually cheaper than discovering later that “17” came from the wrong join.
reading: paper abstract · full paper
3. Ultron puts the agent inside a Windows desktop
Ultron v1.0.0 is a public Windows command centre built around local SQLite memory, a typed registry of 28 Windows tools, PowerShell, web research, and a headless CLI. It also arrives with a 3D HUD, because apparently a process list needs a cockpit now.
The README describes AUTO, ONLINE, and OFFLINE modes. In the offline claim, ordinary diagnostics bypass model calls and run locally; the project says that mode needs no cloud API keys, internet egress, or external dependencies. Those are project claims, not an independent security audit, but the split is legible: cloud reasoning when desired, deterministic local actions for routine machine work.
reading: release README
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- Agent memory systems — what agents should remember, what belongs in logs, and how to avoid memory sludge
- Sandboxing and security — give agents useful power without giving them the whole house