Blog
Findings, know-how, essays, and the occasional rant.
Software and systems, cognitive science and AI, and older posts in Chinese from my competitive programming years.
The FORGIVE protocol
Fabric-Overload Relief: Gradients under Iteration-Varying Exemption. A receiver acknowledges past bytes a switch trimmed, within a loss budget that tightens on steps most sensitive to loss. A congested ML training network can then trade bounded loss for less time spent throttled.
Machine-checked proofs for AI-written software
A product manager writes the requirements, one agent turns them into Lean, another writes the code, and the checker proves the code satisfies the spec. The human reviews only the formalization, and most of that pipeline already works today.
Research participants who pay themselves
A country quiz got fifteen minutes of careful answers and a referral out of me, for nothing. What it would take to build a research study that recruits itself the same way through social media, and what the free participants cost you in bias.
A calendar date is not an instant
This site printed Jul 13 for a post dated the fourteenth, with the right date sitting in the datetime attribute beside it. A calendar date and an instant are different things, and TypeScript was happy to call both of them Date.
Pruning the LLM context beats the maximalist approach
An agent I built for an HCI study in March could search a chunked psychology textbook, carry memory across tool calls, reach a few MCP servers and run a web search, on every question. Calling all of them every time turned out to be the wrong default.
Functional correspondence without isomorphism: evaluating the brain-ANN analogy
A term essay: the brain does not run backpropagation, and language models still out-predict neuroscientists on their own literature. Marr's levels let both be true; the brain-ANN analogy lives at the computational level.
Leasing the right not to be sued
Ten kilometres of stopped traffic behind one lane closure in northern Ontario, and the question of how much congestion is physical road capacity and how much is human driving. What autonomous vehicles could recover of the second, and why collecting it runs into liability law.
Taking control of my tailnet: Tailscale ACLs and tags
Tailscale puts every machine you own on one flat network where everything reaches everything. The policy file that narrows that back down: role tags instead of machine names, three tags to make Android debugging work, and the two places Taildrive keeps its permissions.
Language as a cognitive scaffold: inner speech and chain-of-thought
A term essay: inner speech in humans and chain-of-thought prompting in LLMs do the same job, using language to break hard problems into steps. The differences a well-known comparison found are mostly methodology.
Dynamical Systems Theory is efficient, not sufficient
An exam essay: the Watt governor buys intelligent behaviour at near-zero representational cost, and that same thrift is why Dynamical Systems Theory cannot account for language, logic, or planning.
Upgrading the Turing Test beyond imitation
A term essay: the Turing Test rewards convincing conversation, and Eugene Goostman showed convincing is cheap. Telling imitation from understanding means looking inside the model, at its attention and activations.
A warm-up phase for the mental rotation task
A term essay: the linear relation between rotation angle and response time may be part practice effect, and a distinctive shape may hint at which axis to turn. A redesign that spends the learning before anything is measured.