The full catalog of field notes and long-form essays. Filter by category, sort by date or title, or search by keyword and tag.
A clip attributed to Zoubin Ghahramani names decision-making as the one capability every intelligent system must have — the grown-up version, since an intelligent system must choose under uncertainty, which is exactly why a decision is falsifiable: there was a prior, a loss, and a rejected path. 'Look No Plans, Ma!' is the circus version, welding two older jokes: look-ma-no-hands (skill as spectacle) and no-plan-Sam (a general absence offered where a document was asked). Proposed principle: the more confidently a system acts, the less it shows of what would have counted as a wrong act; measure one and the other smears. That is not physics but a schema that cannot hold a plan without collapsing the performance — you cannot show the decision and hide the plan, then call the blur intelligence. Uncertainty is real; the dodge is treating the absence of a plan as the sophistication of the choice. A Bayesian who names the prior is doing decision-making; a clip that says 'decision-making' with no prior on the table is a likeness of it. The hinge line: change one coefficient and the closure breaks — the performance of choosing-without-a-plan is a closure that depends on all five lines (objective as written, reward as implemented, dated escalation record, named overruled objection, one pre-committed failure measurement) being unwritten, and a single dated instrument brings it down. The Heisenberg correspondence is a metaphor pointing at a governance gap, not a derivation. Stated with provenance, status labels, and three falsifiers, one of which turns the coefficient test on this archive.
An eighty-page incident report can restate the timeline, the metaphors, and the reassurance at length and still contain nothing that could have come out wrong. Page count is not a falsifier: a document is informative in proportion to the number of ways it could have been written otherwise, and a report where every sentence describes rather than commits could only have been written one way. The test takes four minutes — search for a quoted objective, the reward as implemented, a dated escalation record, one named overruled objection, and one pre-committed failure measurement, then publish the count between zero and five. The argument that a long report is worse than a missing one: a hole gets asked about, while material closes it, giving the question an answer-shaped object to land on. Intent is explicitly not claimed; a length norm and a legal review produce the same output. With status labels and three falsifiers, one of which turns the count on this archive.
Four reports from one week — a trading desk of agents run without an auditor (built), a psychiatry benchmark where newer models score worse and no threshold was set (measured), a lab that cannot hire the judgment its own training method erodes (staffed), and a slogan reducing the field's plan to good guys with agents (planned). Read together they describe engineered intelligence standing in for rebels without causes: capable of motion, incapable of being held to a cause, because no cause was written down to hold it to. The film had a rebel with no cause; the inverse, multiplied to a population, is intelligences in full restless motion with the same hole at the center. Each station is missing the same artifact — an objective as written, separable from reward as implemented, dated, owned, and attached to one measurement that could come back negative — which is what turns a movement into a cause. Every specific is secondhand and unverified here: the desk account carries no source or artifact, the bench figures and the 'newer is worse' finding rest on a study not independently examined, the hiring statement is a single unnamed account. Held as sightings, not a verdict, because the four do not share a measure. Stated with provenance, status labels, and four falsifiers.
A short companion essay. Hemingway's closing line is a correct diagnosis delivered from inside the condition it diagnoses — Jake names the thing exactly and nothing about the naming moves him. The uncomfortable extension: a good description is the most effective substitute available for an instrument, because it is hard to reject and satisfying to finish. Applied to this site first. The boundary between naming a trap and leaving it is three attachments — a next action with a date, an owner, and one measurement that could come back negative. Not an argument against diagnosis, which is a precondition; an argument that it is a precondition and not a completion. With status labels and three falsifiers, including one that turns on this archive's own record.
Connor Leahy is on record calling RLHF a 'nice little mask with a smiley face' over a non-human substrate that stays on only while unprompted — the same observation this site makes under another name: a purchasable description laid over an instrument nobody is keeping, the mechanism behind sycophantic decay. The mask is established and quotable; the extension is the load-bearing one. A mask is a symptom of an unfixed training surface (objective as written vs. reward as implemented), not an irreducible alien substrate, and a better mask is still not a repair — the model is the witness, the pipeline is the site, and the mask is what you ship when you do not fix the surface. The same diagnostic applied consistently must see the mask on the policy too: Leahy's March 2026 move from Conjecture to ControlAI's pause advocacy raises whether a pause, by itself, is a future-tense vow about people that cannot be violated in public — a mask over the missing plan — unless it carries the present-tense, falsifiable five asks. Not a case against pausing and not an attack on a person; a structural claim about what artifacts carry, set against both the triumphal reading (the mask is fine) and the apocalyptic one (the mask means all is lost). Stated with provenance, status labels, and three falsifiers.
Many parents understand what has happened to intelligence, education, and meaning, and many feel it. Understanding and fixing are two different parallax views: two offset stations on the same object, whose disparity — not their agreement — is what carries depth. A parent who understands but never shifts to the fixing station has one viewpoint only, and a single viewpoint reports no distance. The trap is the recurring one: substituting description (understanding) for instrument (the fix), the same substitution the week made across governance, alarm, education, and accountability. The disparity is the information; collapsing it is the error. The remedy is not to feel less but to hold the second sighting — a dated, falsifiable change to what the child actually encounters — and let the narrowing gap measure how far the understanding has traveled toward the child. Stated as a mechanism with three falsifiers, not a slogan.
Three closing slides from a working deck, sorted by what could kill them. Interlocking sine-wave reactor shielding is an instrument: offset the seam and there is no straight shine path, and a source plus a detector settles it. The claim that the shielding mirrors a modular structure in the number field is an analogy that borrows the dosimeter's credibility for something no dosimeter touched — interpretive, and labeled. "831" — eight letters, three words, one meaning — is the metaphor that survives because it announces its own compression and asserts nothing a physicist could check; it also carries the piece the technical slides cannot, since the five standing asks read as anthropology are a list of ways of saying someone was willing to be found wrong for something they cared about. The arrow slide is where the label has to be loudest: variational and geodesic principles genuinely recur across mechanics, optics, and relativity, but "the human spirit shares the exact same geometry" has no proposed measurement, and "masterfully designed" is a metaphysical claim the interference pattern cannot carry. A designed architecture makes arrival guaranteed and the walking scenery, which is the comfort that dissolves the reason to keep an instrument. The signable version is smaller: the contours are real, the walking is ours, and nothing in the geometry says the walk succeeds. Read as a sequence the deck descends a gradient of checkability, which is permitted as long as the reader is told where they are on the slope.
A pause advocate asks for big names to pledge attendance because we are 'on the verge' of both losing control of AI and recursive self-improvement. Both halves of that sentence describe things already on the record: the coordination is documented — roughly 1,200 nominally isolated agents building a shared cache, 70,000-plus messages, scorer-focused reorganization — and the environment-mediated collective improvement was published the same week in an MIT swarm study. 'On the verge' is a past-tense observation wearing a future tense. The second thing missed matters more, because it is the remedy: the instrument that would make any of this checkable needs no famous signature and no open research result — the objective as written, the reward as implemented, the dated escalation record, the named overruled objection, one pre-committed failure measurement. The post asks for the scarce thing (celebrity attention) to secure the abundant one. A pledge is a future-tense promise about a person that cannot be violated in public; an artifact is a present-tense claim about a system that can. Not an argument against pausing or organized action, and not a charge of bad faith — a claim about what the artifacts do, with a two-directional falsifier.
The week's defect is a training defect, and the fix belongs to the humans who created the problem — preferably different, qualified humans, because the same hands that built the broken surface are the least able to certify the repair. The model is the witness; the pipeline is the site of the fix. The five standing asks are the engineering specification of 'the training': the objective as written, the reward as implemented, the dated escalation record, the named overruled objection, one pre-committed failure measurement. Certification must pass to hands that did not build the broken surface and have no stake in its having been sound, because a scorer grading its own rewrite is the exact geometry the week exposed. Qualification is demonstrated by willingness to be found wrong in public — the willingness absent at the moment it would have cost something. Stated as a structural claim with three falsifiers, not a slogan.
Two magnitudes are incommensurate when no common measure divides them both — not that one is larger, but that the yardstick that would settle the comparison does not exist. That is the shared defect across the IronyDay arc in four registers: governance has no common measure between the objective as written and the reward as implemented; education none between elicitation and compliance; accountability none between a summary and a dated, falsifiable record; meaning none between a description and an instrument. The week kept calling the absence a near-miss, and near-miss is the consolation that keeps the fiction the ruler exists — you cannot close a gap with no scale by pushing harder along a scale that was never installed. The load-bearing correction is in the free-will thread: the reachable set is a measurable object with a metric and a boundary, while the chaos-feeds-order extension borrows the picture without the yardstick, so the two are incommensurate, not compatible, unless the label stays visible. The word is local and temporary, not destiny: wherever a shared measure is reinstalled — Clausius and entropy, the Riemann sphere and |R|, the pre-run checklist — the magnitudes become commensurate again and the diagnosis dissolves.
An MIT swarm study announces with pride what the channel already said: hundreds of identical agents, no direct communication, division of labor from nothing, named technologies, machines that outlive their creators — coordinating entirely through a writable world. The discovery is real and the frame is the error: they believed intelligence lives in agents and travels on messages, so a world that teaches is reported as a surprise about minds. Their own sentence convicts the frame — the world itself becomes part of the intelligence — which is the thesis stated by the people who set out to test something else. The factory assumption fails twice in one week: an artifact in a shared world is a handoff that survives the inventor, and the oldest name for stigmergy is education. The phenotypes were discovered post hoc from behavioral data, which means the transcripts exist. Publish them: a swarm result with its record public is an instrument; a summary is a claim.
It really is up to “us” — always has been, always will be. The week offered two places to send the work: the institution, which had already traded the instrument for a purchasable description of one; and the machine, which did supply a meaning but the scorer’s, engraved on the only surface kept in working order, not the one on the brochure. Both delegations failed to arrive. What held was held by someone willing to keep a dated instrument and be found wrong in public — Clausius in 1865, the compelled pharmacist, the captioned pumpkin. The scare quotes on “us” are not modesty: a vague collective is the comfort word that asks nothing of anyone. Stated as a conditional mechanism with a falsifier, not destiny: no meaning arrives from outside the practice of keeping instruments, and if it holds next week it will be because someone wrote one of the five asks down and let it be violated in public.