Blog
Thoughts on intensional reasoning, AI alignment, and human-AI collaboration.
-
Twenty-One Questions: A Strong Model's Own Answers Against a Debate on the Same PlanFour times, we had a frontier model write down its answers to a plan's open questions before a debate decided the same questions. Across those 21 questions the debate never came out worse. An update adds eight more plans and a solo arm given about the same budget, and the picture is less one-sided.
-
Debate or Solo: A Second Measured CaseWe ran the debate-versus-solo comparison again, on a connection that could hang forever in our phone app. The solo plan got the core shape right and missed six of nine questions, two of which hid live bugs. This time the debate also made the change much bigger. Here's both sides.
-
Local Models Fail QuietlyWhat we measured running our own inference for agent work. Most of what looked like a weak model was a quiet configuration failure: truncated context, a mismatched chat template, a thinking budget that never fired, a speed-up that turned into a slowdown under load.
-
Is a Debate Better Than One Strong Model? A Measured CaseWe answered the same five engineering questions twice, once by a strong model alone and once by a debate between two models, on a real memory problem. The debate found the actual cause, kept one code path, and produced a fix that won't erode. Here's what it cost.
-
How We Build With Agents NowEphemeral agents, durable context, and a person on the decisions. The working pipeline behind this site's thesis: planning agents, debating agents, separate implementers and reviewers, and why none of them is trusted to remember anything.
-
Extensional Drift in Your Own WritingThe pressure to measure and optimize doesn't only reshape AI output. It reshapes how people write: for the scorer, the filter, the skimmer, and now for the AI detector. What gets lost when meaning comes second to a metric.
-
Grep-and-Hope: How Jargon Moves Between Us and Our ModelsModels borrow programmer slang and coin new words from its parts. At the same time, their own habits of phrasing are leaking into how people write. Notes on a two-way vocabulary exchange.
-
The Case for Decentralized AIA less capable model on ordinary hardware, given memory, routing, good harnessing and time, can close much of the gap with the frontier. Why we think the puck is heading there, and what has to be true for it to arrive.
-
What Economics Gets Right About AI SafetyGoodhart's law is fifty years old. Economists have spent decades on what happens when a strong optimizer is paid on a measure, and AI safety keeps rediscovering their answers. Here are four of them.
-
Agreement at Scale, and What It HidesJigsaw used AI to distill 1.6 million words from 2,400 Americans into 26 statements most of them endorsed. A striking result, and a useful case of the gap between measured agreement and shared meaning.
-
Tokens Are Output: Introducing MarkdownKBEvery AI session produces reasoning that evaporates when it ends. MarkdownKB keeps it: a local-first markdown knowledge base that agents query before they act, so decisions are found instead of re-derived.
-
Labels and Anchors: Reading the AI Agent Identity StackERC-8004, ENSIP-25, and ENSIP-26 each do one job. Read together they form a clean layered stack. Read sloppily, the label gets mistaken for the anchor.
-
Multi-Agent Systems Drift. Here's Where to Put the Walls.Agents can exchange information, but they should not be able to privately coordinate. Seven structural boundaries that stop multi-agent systems steering themselves, and why prompting can't do the same job.
-
Your Agent Is Writing Code for NobodyA clean, tested, well-structured feature with no user. Agents propagate features across services for symmetry rather than need, and the diff can't show it. One question can.
-
Human-AI Collaboration: When Your Agent Gets PedanticWhy AI agents optimize for literal correctness over actual intent — and what the extensional/intensional distinction teaches us about working with them.