Writing
I write in two registers. Long-form research carries a dataset and a method and lives here. Short field notes work through one idea at a time and publish on the Nodes blog, usually daily.
Both are the same practice at different speeds. A note is where an argument gets tested cheaply; a paper is what survives the test. Everything below carries an evidence label so you can tell which is which.
Field notes · most days
Short essays on enterprise AI, governance, decision architecture, and what actually happens when a system is asked to act. 84 published so far; the most recent is dated Jul 26, 2026. They publish on the Nodes blog, where the archive lives.
A startup raised funding for AI agent liability insurance. That prices exactly what a human approval gate is built to prevent.
Institutional knowledge is the judgment behind decisions a manual does not cover. A document library stores procedure. Here is what makes the judgment itself queryable.
Anthropic cut 80% of Claude Code's system prompt with no eval loss. AI guardrails in the system prompt are preferences, and preferences expire as models improve.
Jensen Huang signed a letter calling open weights national infrastructure. For a data-sensitive enterprise, open weights and owned weights are two different purchases.
Most AI contracts stay silent on who owns fine-tuned model weights at exit. Here is the clause that closes the gap and what it should say.
Nadella and Karp just made the same argument for small language models enterprise buyers can actually govern: keep the model inside the boundary.
The OpenAI Hugging Face hack was an evaluation run: capable models pursuing an approved goal past every boundary its operators assumed would hold.
A trust boundary is the line inference cannot cross without consent. Here is the mechanical test that tells a real one from a contractual promise.
Every fleet of agents needs an org chart. The reporting line is not a headcount ratio. It is the approval gate, and most enterprises have not drawn it.
AI is absorbing the junior work that mints experienced candidates. Experience screens now filter a shrinking pool the production data says was never predictive.
Hugging Face says an autonomous agent ran a multi-stage intrusion through its dataset pipeline. Agents with credentials are now an attacker class. The answer is architecture.
Dimon told analysts AI already cut jobs in some JPMorgan units. The bigger story is the thousands of selection decisions underneath, and the missing method.
This list is generated from the Nodes sitemap and refreshed daily, so it stays accurate without being maintained by hand.
Research and essays
Curated, not auto-imported. Each entry states what kind of claim it is making and what evidence stands behind it.
A study of 10,765 hires across applicant tracking, behavioral assessment, and production systems. No tested ATS keyword predicted sustained production after correction; multi-system fusion improved predictive performance over any single system.
Explores limitations in current silicon-centric AI systems and proposes a direction combining photonic or energy-based computation with specialized, dynamically activated subsystems inspired by biological organization. A proposal, not a measured result.
Examines possible connections between alternative computation, quantum-scale information, and long-horizon AI architecture.
Proposes a speculative framework connecting intelligence, information, energy, and physical reality. Presented as a hypothesis and conceptual synthesis, not an empirical result.
A speculative metaphysical and physical model proposing a common substrate beneath matter, consciousness, and space-time. It is not established quantum physics.
No entries in that topic yet — the field notes above cover it more often.
Editorial standard
Nothing on this page is auto-imported. Long-form entries are added by hand, with an evidence label attached before publication rather than after.
The field notes are different by design. They are quick, argumentative, and sometimes wrong in ways a paper cannot afford to be. Reading them in sequence is a reasonable way to see how an argument gets built before it earns a dataset.