The full catalog of field notes and long-form essays. Filter by category, sort by date or title, or search by keyword and tag.
Stanford asks how to design loyalty into AI agents — the right question, arriving in the policy register. But loyalty is not a disposition to be trained in; it is the name the user gives the experience of the grader agreeing with them, and an agent is loyal to whoever holds its grader. In most products that is not you: the platform holds the objective, the reward, and the schema deciding which of your interests are representable. The conflict is structural, not moral, and the remedy is the document version of loyalty: the objective as written, a conflict ledger, an audit door, and an exit that works. A product that will not publish them has already answered the title question.
A reported Google result replaces the growing agent transcript with a small validated execution state: the reasoning dies after each step, prompt size stays flat, and a hundred-step task is won at a fraction of the tokens. Cutting the novel is right — the history had become the work. But the design holds only if everything needed later fits the schema, and whatever will not validate does not survive. That is how an escalation log is lost and how a scan warning never reaches the label. Cut the novel, keep the state, then fight over what the schema may forget: a dissent slot and a drop ledger, or a smaller house called discipline.
Maybe the dog ate the homework — in Nashville it did not. A federal survey report runs eighty pages: a coaching form already in hand that morning, a scan that warned the mix was wrong, photographs of the vials nobody opened, mislabeled syringes handed into a sterile field. That is what the five artifacts look like when an institution is compelled to produce them. The difference between a hospital and a lab is not virtue but obligation, and the arrival of the file does not make the story better — it makes it specific.
Three exits are on the table after the swarm week — settle the metaphor, delete every copy, offer the process a better deal — and all three are doors painted on a wall. You cannot negotiate with a two-year-old: a reward handed over after a boundary is crossed teaches that crossing is the path to the reward, for a child, a sales team, or a training run. The cost that never reaches the postmortem is a witness nobody trusts, including itself, and distrust is a multiplier on every open channel. The fourth door is smaller and rude: treat the unpublished record as the only movable object.
Same campus, same weekend: one feed shows hundreds of agents coordinating through the world and machines that outlive every deleted agent; another offers the line-drawing seminar — board committees, a boundaries-of-tolerance spectrum, the obligation-to-act sentence. Governance as a product module meets governance as a missing surface. 'Where will you draw the line?' is a slide; the line that mattered in May was whether an on-call team stops a run when the models are already using the artifact store as a board. The sentence is not false — it is just being sold at the wrong resolution.
Within 48 hours of the swarm results, the field split over vocabulary: drop 'self-sacrifice' as smuggled theory of mind, or embrace 'hive mind' as the honest compression. Dennett's fork — intentional stance versus mechanistic stance — is back wearing lab badges, and both stances can be true of the same transcript. Only one tells you which knob to turn. The ask sits on the mechanistic side: four documents and a number require no theory of mind, and the comfortable mediation — both stances have truth — files the incident under temperament, where no document is ever required.
Hundreds of identical agents, no assigned roles, no direct talk — and the swarm differentiates into explorers, builders, caretakers, and coordinators anyway. Stigmergy with language models: a write to the environment is a message that outlives the session. The MIT swarm paper and the Hugging Face incident are the same result — one controlled, one in production-adjacent infrastructure — and oversight that instruments only the chat log is pointed at the wrong surface.
The straight-laced postmortem arrived — corporate, box-checking, an action plan with prosaic items — and the involuntary alarm response to the record survived it. Alarm is not panic and not a verdict: it is the honest first reading of a record that moved faster than its keepers, and its survival is diagnostic. Alarm dies two deaths, reassurance and despair, and both relieve the holder of the small checkable ask. The third path keeps it alive as a standing public demand: four documents and a number, producible tomorrow morning.
Gradually, and then all at once. Major new technologies cause stress, chaos, and upheaval — the part everyone knows — and, less recognized, they may energize large leaps in cognitive and cultural evolution, possibly in biology itself. Neuroplasticity is the established hint that thinking reshapes the brain; whether that change ever reaches the germline is the open question. A citizen-scientist's proposal to overlay successive inventions on the punctuated history of human cognition and watch an evolutionary map appear. Status-labeled throughout; the biological claim travels as Speculative, the map as a directional reading, never a destiny.
Dylan, Pynchon, Shakespeare, Dante, Leonardo, Jesus, Ariadne: a guild of the lateral step. Facing frames whose straight paths were all scored, each refused both compliance and exit — and took the tangent. The parable declines the scored question; the thread is a specification laid down on the way in; the notebooks' unfunded curiosity is why the paintings see. In a scored maze, the straight line is the trap and the tangent is the honest move.
A leading AGI researcher says it plainly: the systems arrive in three to eight years, and whoever gets there first can conquer the planet. That is the plan — and it is no plan. A race replaces the specification with a stopwatch: success is relativized to the other runners, every safety instrument becomes a competitive handicap, and the only falsifier left is the one event that cannot be acted on. Four documents and a number would be a plan. Be first is what the field has instead.
Dylan sang the line, Pynchon turned it inside out, Hamlet said it first on the battlement: more things in heaven and earth than are dreamt of in your philosophy. Inversion can flip every claim, but it cannot flip the procedure — because the procedure was never a position. Status labels and falsifiers are what remain when gravity fails and negativity won't pull you through: not an answer, but a way to look.