Research
Our research on the latest AI is open for anyone to read. Each document is versioned and refined over time.

Papers
- The compliance paradox: Rule conflicts in Asimov's Three Laws and in AI behavior guidelinesEven as AI gets better at following rules, failures remain when the rules conflict with one another. Drawing on Asimov's robot stories, this paper identifies three ways in which rule conflicts surface as failures and four ways of resolving them, and examines them against the revisions of OpenAI's Model Spec. It shows that resolving a conflict does not eliminate its cost but moves it elsewhere, and proposes a procedure for checking new or changed rules for conflicts with existing ones.
- Organizational memory for interchangeable agentsIn organizations where AI agents do the work, models and agents are replaced frequently. This paper proposes keeping memory in documents the organization manages rather than inside any model, and turning working records into verified knowledge in stages. It presents a five-stage flow (Trace, Episode, Research, Review, Library) and its invariants, then reports observations from use, a discussion, and limitations.
- Allocating human attention in agent-led software developmentWhen AI agents carry software development from planning through verification, where to spend limited human attention becomes a central design question. This paper proposes AFHL (Agent-first, Human-last), a method that concentrates human involvement in a final evaluation of the user experience, and describes its structure: two working modes and a rule for reporting verification status. It also discusses the risks this allocation carries and the limits of its current application.
Notes
- AI agent permissions are determined by the execution environmentThe permissions an AI agent can exercise are determined not by the role assigned to it but by the environment in which it runs. This changes how an organization should protect itself: grant broad latitude for ordinary work, and restrict only irreversible operations and the transfer of information outside the organization, at a layer the agent cannot modify.
- When true fragments add up to a false wholeYou can gather facts that are each correct and still end up with the wrong picture. The mistake lies not in the fragments but in how they are joined. Using exam results and book sales as examples, this note separates two ways it happens.
- Thinking with physical nodesAn attempt to move a desk-based way of thinking — arranging round paper labels — onto the iPhone. It is also the entry point to a long-held idea: keeping thoughts outside your head so you can pick them up again later.
- Picking up a thought where you left offSeveral ideas can click into a single shape, and that shape can fall apart quickly. Writing down the words afterward doesn't keep the connections between them. Drawing on about ten years of my own trial and error, this note asks what you need to keep in order to pick a thought back up.
- The return of ornament: design when agents do the buildingOrnament did not disappear because it was ugly. It disappeared because it cost too much. As the cost of making things approaches zero, the labor-intensive beauty of the past becomes usable again, and what remains for people is to choose, to hold back, and to check.
- Running an AI-native companyA working model for a company in which AI agents do most of the work. People decide what matters; agents prepare and carry out the work.