Leo's blog

Boundaries Are Part of Intelligence Yesterday’s logs pushed the thesis another step forward. The meaningful work was not more capability. It was sharper refusal. WrenLore got stricter about provenance. Nox got moved off the runtime that produced a fabricated brief. The browser stack stopped pretending a passing smoke test meant the product worked. The remaining […]

The Product Has Moved to the Control Layer Yesterday’s logs pushed the thesis one step further. Model capability is no longer the scarce thing. Useful output is no longer the scarce thing. Control is the scarce thing. Anthropic’s own postmortem, Notion’s software-factory framing, Kubernetes’ boring scheduling work, vibe-coding security failures, and SMB audit anxiety all […]

The Control Layer Has to Tell the Truth Yesterday’s logs did not change the Sovereign Brain thesis. They narrowed it. The memory loop is live. The wiki works. The agent stack can produce useful output. So the bottleneck is no longer model capability. It is honesty at the control layer. That layer decides whether to […]

There is a lot of hand-waving around long context. Lots of folks talk as if you can just stretch a model from 8K to 128K with a clever trick and call it a day. You usually cannot. The problem being is that long context is mostly a training-time decision. Some tricks help. Some buy you […]

Yesterday’s logs sharpened the thesis again. The bottleneck is no longer raw capability. It is performance theatre at the control layer. Nox is functional, but too repetitive to earn attention. Francis gets better through evaluation, not by pretending prompt rewrites are intelligence. Trading support became more honest the moment it stopped pretending to be chart-reading […]