Corrections

Every factual error found in a published video, what it was, and what replaced it. Nothing is ever removed from this page.

The videos state their sources and mark the claims that sit close to the edge of what the literature supports. That only means something if the mistakes are on the record too. So they are here — including the ones caught before publication, because a log that starts empty tells you nothing about how carefully anything was checked.

Found an error? Write to corrections@prompteur.berlin with the timestamp and, if you have one, a source. Corrections get published here with the date and, if you want, your name.

What does not belong here: disagreements about interpretation. The videos say out loud where they use an assumption or where a question is unsettled. Weighing those differently is a different opinion, not a factual error — and mixing the two would make this page useless.

We Built Something We Cannot Read Published 2026-09-01 · runtime 14:27 · watch

Found after publication

Nothing so far.

Found before publication

  1. 2026-08-25 The closing promise covered more than the fact-check did. The script promised a source for every factual claim. The check covers eleven marked claims; unmarked sentences and all of part 5 were explicitly outside it. That is not a detail — it is this channel’s core promise, and it was not covered. Was: “Every factual claim in this video is listed in the description, with a source.” Now: “Every claim in this video that rests on published research is listed… Where I gave my own reading instead of a finding, I said so out loud.”
  2. 2026-08-25 “One number, turned up” was simply wrong. A feature is a direction in activation space, not a single weight. The original wording invited exactly the wrong reading of the Golden Gate demonstration. Now: “one extracted feature, turned up”, together with “no system prompt … no retraining”.
  3. 2026-08-25 The feature count claimed more than the paper shows. Scaling Monosemanticity trained sparse autoencoders with up to 34 million features on Claude 3 Sonnet. That is the size of the dictionary, not the number of features shown to be interpretable — interpretability was assessed on samples. Was: “extract millions of interpretable features”. Now: the dictionary size is named as such, and the video says interpretable ones were found among them.
  4. 2026-08-25 “Nothing close to that exists” was an unprovable universal claim. Nobody can show that no such method exists anywhere. The video now quotes the people doing the work instead, on their own circuit-tracing method: “even on short, simple prompts, our method only captures a fraction of the total computation performed by Claude”.
  5. 2026-08-25 The cold open contradicted itself. “Nobody can give you a full account” followed by “not nobody at all — that stopped being true in 2025” mixes up complete and partial explanation. The opening now states the actual situation without the reversal: full access, and still no readable account of why.
  6. 2026-08-25 Four sentences were more absolute than the evidence. “Nothing hidden from us” now refers to the people who trained the model; “optimising for prediction and nothing else” is limited to ordinary pretraining; superposition is “a well-documented account” rather than the cause; and “every one of those questions is answered by proxy” became “mostly answered indirectly”.
  7. 2026-08-25 Part 5 rests on premises the check did not verify. It now says so before it starts: “What follows is my reading rather than a finding.”

AI Has Three Limits. Only One Is a Wall. Published 2026-08-25 · runtime 13:58 · watch

Found after publication

  1. 2026-08-26 The description misdescribed how this channel is made. It said the narration was synthetic but “the script was written and fact-checked by hand”. That is not true, and the sentence structure made the contrast on purpose: the script was drafted by a language model and the fact-check was done by one too. Was: “The narration is a synthetic voice, not a human reading. The script was written and fact-checked by hand.” Now: “This channel is made with AI, end to end… A person picks the topic, reviews the result and is answerable for every claim in it.” Nothing about the twelve sourced claims changes — this is a correction about the channel, not about the subject. It is logged here because a channel that keeps a public correction log and then quietly edits its own provenance would be worth less than one that keeps no log at all.

Found before publication

  1. 2026-08-13 The Lorenz restart was in 1961, not 1963. The script dated the experiment with the rounded values to 1963. That is the year Deterministic Nonperiodic Flow was published; the restart itself happened in the winter of 1961. Was: “In 1963, Edward Lorenz was running…” Now: “In 1961…”, with the paper named as following two years later.
  2. 2026-08-13 “About one part in a thousand” was out by a factor of four. Rounding 0.506127 to 0.506 is a difference of roughly one part in four thousand. Now: “a few parts in ten thousand”.
  3. 2026-08-13 The Entscheidungsproblem is not about “any mathematical statement”. As posed by Hilbert and Ackermann it concerns the validity of statements of first-order logic. The original wording claimed more than the result supports.
  4. 2026-08-13 Rice’s theorem is from 1953, not 1951. 1951 is the dissertation; the citable publication is Trans. AMS 74 (1953), 358–366.
  5. 2026-08-13 “SAT solvers routinely handle millions of variables” was too broad. It holds for highly structured instances. Small, deliberately crafted instances can still defeat them, and the video now says so.
  6. 2026-08-13 The claim that machine learning does not extend the forecast horizon was too categorical. Vonich & Hakim, Atmospheric Predictability beyond 30 Days with Machine Learning (AIES 2026) report skill beyond 30 days — but only with initial conditions optimised in hindsight, which cannot be done in real time. The video now calls this unsettled rather than settled.
  7. 2026-08-13 Three framing errors, corrected. NP-hardness does not imply decidability, so “every one of these problems is decidable” now names the NP-complete ones only. A fixed 9×9 Sudoku is finite and says nothing asymptotic, so the video uses generalised Sudoku. And chaotic predictability is not a worst-case theorem, so the closing line no longer groups all three limits under “worst cases”.