Latest Corrections
LessWrong Mar 5, 2026 at 12:21 PM
www.lesswrong.com/posts/TjcvjwaDsuea8bmbR/maybe-there-s-a-pa...
- Cordite was invented by Sir Frederick Abel and Sir James Dewar (1889), not Alfred Nobel.
- Szilárd fled Nazi Germany in 1933 (after Hitler’s rise to power), not 1937; 1937 is associated with his move to the United States.
LessWrong Mar 5, 2026 at 07:50 AM
www.lesswrong.com/posts/CJw2tNHaEimx6nwNy/ai-forecasting-one...
- The post reports MMLU SOTA as 67.5%, but the PaLM technical report (submitted April 2022) reports a higher MMLU score of 69.3% for PaLM 540B, meaning SOTA by June 30, 2022 was at least 69.3%.
X Mar 5, 2026 at 07:21 AM
x.com/TetraspaceWest/status/2029103306543390976
- Credible tallies put the number of nuclear test explosions alone at 2,056—already higher than 2,000—so the total number of nuclear detonations in history is greater than 2,000.
X Mar 5, 2026 at 06:41 AM
x.com/queens_parents/status/2028837472570867866
- NYC Public Schools had not eliminated Kindergarten Gifted & Talented (G&T) admissions as of early 2026; NYCPS enrollment guidance still described Kindergarten G&T as part of the Kindergarten application process.
LessWrong Mar 5, 2026 at 05:34 AM
www.lesswrong.com/posts/GcbkprYPCjXdysLq4/text-compression-c...
- This is a math/unit-conversion error: 1 trillion tokens/day × 2 bytes/token ≈ 2 TB/day (≈ 2000 GB), not ~500 GB.
LessWrong Mar 5, 2026 at 05:09 AM
www.lesswrong.com/posts/xCWiFGezwMPswZ6Ea/is-gdp-a-kind-of-f...
- Davis & Palumbo (2006) did not estimate land’s share for all U.S. residential real estate; they estimated it for 46 large U.S. metropolitan areas (an average across their city sample).
LessWrong Mar 5, 2026 at 05:08 AM
www.lesswrong.com/posts/sjeqDKhDHgu3sxrSq/sacred-values-of-f...
- Randomized trials of more vs. less generous medical coverage found measurable health improvements (e.g., reduced hypertension and improved vision) for some groups, contradicting the claim that those receiving more care were not healthier on the margin.
LessWrong Mar 5, 2026 at 05:07 AM
www.lesswrong.com/posts/LXQBcztrWKhtcgQfJ/current-activation...
- Karvonen et al.’s paper does report ~1M training examples, but the breakdown in this post is incorrect: the paper describes 600k context-prediction examples, 336k classification examples, and 64k SPQA (system-prompt QA) examples—not “300k system prompt extraction tasks.”
LessWrong Mar 5, 2026 at 12:04 AM
www.lesswrong.com/posts/3ZBmKDpAJJahRM248/proof-idea-slt-to-...
- For Solomonoff/Bayes-mixture prediction, the standard bound on cumulative expected KL divergence is an O(K(μ)) term (Kolmogorov complexity / -log prior weight), not an additional term growing linearly with the number of data points like D·H(h).
LessWrong Mar 5, 2026 at 12:03 AM
www.lesswrong.com/posts/SzjfsWe9bfHSX8fRg/petapixel-cameras-...
- A ground sample distance of 10 cm per pixel is far too low-resolution to enable facial identification/recognition; standards and industry guidance typically require on the order of tens of pixels across a face (hundreds of pixels per meter).
- 0.1 m (10 cm) per-pixel resolution is far too coarse for facial identification; guidance commonly requires about 40+ pixels across a face (hundreds of pixels per meter), not ~1–2 pixels.
Substack Mar 4, 2026 at 11:47 PM
www.hyperdimensional.co/p/on-recursive-self-improvement-part
- Public MLPerf Training v4.1 results included training benchmark submissions using NVIDIA’s Blackwell (B200) accelerators, including the GPT‑3 pre-training benchmark—so models have been trained on Blackwell hardware publicly.
LessWrong Mar 4, 2026 at 11:41 PM
www.lesswrong.com/posts/PBrggrw4mhgbksoYY/a-tale-of-three-co...
- Pete Hegseth’s last name is misspelled here: credible sources spell it “Hegseth,” not “Hegsted.”
- Microsoft reports GPT models are available in U.S. Government Secret/Top Secret cloud environments, which are classified networks—so Claude Gov is not the only LLM deployed there.
- Microsoft states GPT models are available in classified Secret/Top Secret cloud environments, contradicting the claim that Claude Gov is the only model deployed on classified networks.
Substack Mar 4, 2026 at 11:41 PM
www.hyperdimensional.co/p/clawed?hide_intro_popup=true
- U.S. “supply chain risk” procurement authorities (under FASCSA/FAR) are not reserved exclusively for foreign-adversary-controlled firms; they can apply to any non-federal supplier, and the statute explicitly says actions can’t be based solely on foreign ownership.
LessWrong Mar 4, 2026 at 10:57 PM
www.lesswrong.com/posts/QzpKq92nXqp8NHM34/neural-tangent-ker...
- This gives the wrong normalization constant for an n-dimensional (multivariate) Gaussian; it’s missing the factor of (2π)^{n/2} and should use √det(H) in the denominator for the multivariate density.
- Infinite-width NTK-style analyses don’t imply “always” zero training loss: convergence to zero depends on the NTK (kernel/Gram) matrix being strictly positive definite (no zero eigenvalues). The NTK is only guaranteed to be positive semidefinite, so zero-loss convergence is not unconditional.
LessWrong Mar 4, 2026 at 10:56 PM
www.lesswrong.com/posts/ybLaw7EuMBeM2dLn9/broad-basins-and-d...
- Computing an eigendecomposition of an n×n (dense) Hessian generally scales cubically (≈O(n^3)), not O(n^2).
LessWrong Mar 4, 2026 at 09:54 PM
www.lesswrong.com/posts/qFzWTTxW37mqnE6CA/iabied-book-review...
- Nuclear reactors can be (slightly) supercritical (k>1) during normal, controlled power increases; they do not “explode” merely because k is slightly above 1. What must be avoided is *prompt* supercriticality / prompt criticality, a specific threshold beyond mere k>1.
X Mar 4, 2026 at 09:32 PM
x.com/MDinCanada/status/2029235235913691194
- The journal action described in the underlying reporting was a mass *correction* notice (a “Society Note” stating it is a correction), not a retraction.
X Mar 4, 2026 at 09:16 PM
x.com/LundukeJournal/status/2028892521116635283
- Canonical/Ubuntu has explicitly said it has no concrete plan yet (and may not change at all), and that the discussion cited was only an informal community thread—not an adopted Ubuntu plan.
- The proposal shown in the cited thread is for a D-Bus interface named `org.freedesktop.AgeVerification1` (with a “1” suffix), not `org.freedesktop.AgeVerification`.
X Mar 4, 2026 at 09:16 PM
x.com/HellenicVibes/status/2028717381925888066
- LLMs can function as useful tools without any built-in “persona”: at their core they are next-token predictors, and conversation/persona behaviors are added via fine-tuning (e.g., instruction tuning/RLHF).
X Mar 4, 2026 at 09:11 PM
x.com/BoWang87/status/2028935492977475623
- In Knuth’s PDF, this quote appears in the opening paragraph, not as the paper’s closing line.