Latest Corrections
LessWrong Mar 6, 2026 at 07:43 PM
www.lesswrong.com/posts/rWc6BT8BcoPhwcbSs/ai-safety-has-12-m...
- This overstates how far EU AI regulation has been delayed. The EU AI Act already has obligations in force, with major milestones in 2025 and 2026 and full roll-out by August 2, 2027—not a general push to 2027–2028.
LessWrong Mar 6, 2026 at 06:14 PM
www.lesswrong.com/posts/TMFNQoRZxM4CuRCY6/reason-isn-t-magic
- This overstates the history. Historical sources say cassava-processing knowledge traveled to Africa with the crop, and African traditional processing methods likely derive from Indigenous American techniques rather than being simply omitted.
LessWrong Mar 6, 2026 at 06:14 PM
www.lesswrong.com/posts/xCWiFGezwMPswZ6Ea/is-gdp-a-kind-of-f...
- The SoftBank Vision Fund was not a $45 billion fund. $45 billion was the size of Saudi PIF’s planned commitment; SoftBank said the fund’s overall size could reach $100 billion and announced a first close with over $93 billion committed.
- 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 6, 2026 at 06:13 PM
www.lesswrong.com/posts/N9oKuQKuf7yvCCtfq/can-crimes-be-disc...
- This leaves out a key legal qualification: plea colloquies ask about force, threats, or promises **other than those in the plea agreement**. Plea bargains themselves commonly include disclosed promises such as dropped charges or sentencing recommendations.
Wikipedia Mar 6, 2026 at 07:09 AM
en.wikipedia.org/wiki/Lockheed_C-130_Hercules
- This mixes up Canadian and U.S. designations. CC-130J is the Royal Canadian Air Force designation; the stretched U.S. variant is designated C-130J-30, not CC-130J.
- This is outdated. The RAF retired its C-130J Hercules fleet in 2023, so there are no C-130Js currently in RAF service.
- This is no longer correct. The RAF retired its Hercules fleet in 2023, and the RAF Falcons now jump from the Atlas A400M.
Wikipedia Mar 6, 2026 at 07:08 AM
en.wikipedia.org/wiki/Main_Page
- This is outdated. The original Marine Parade Community Building was demolished in 2023, and reporting on the redevelopment said the mural facade would be removed, so the mural cladding no longer currently "is" Singapore's largest installation art.
LessWrong Mar 6, 2026 at 06:56 AM
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.
X Mar 6, 2026 at 06:40 AM
x.com/MorePerfectUS/status/2029290242876870662
- The bill is narrower than this says. It targets chatbots that effectively impersonate licensed professionals by giving advice or taking actions that would count as unauthorized professional practice if a human did them; it does not broadly ban AI from answering any questions merely 'related to' those fields.
LessWrong Mar 6, 2026 at 06:28 AM
www.lesswrong.com/posts/4CkqSpkwA9pLjkmz7/have-americans-bec...
- The 35-minute figure is for wounding-to-treatment, not wounding-to-surgery. Official Army histories put Korea-era time to surgery much higher—about 9.8 hours on average, with roughly 4.6 hours just to reach a forward hospital.
LessWrong Mar 6, 2026 at 06:24 AM
www.lesswrong.com/posts/qefrWyeiMvWEFRitN/milestone-announce...
- XBOW’s own “Top 1” announcement is dated June 24, 2025—about 8 months before this LessWrong post (Feb 19, 2026), not “almost one year.”
- The XBOW announcement post being discussed is dated June 24, 2025, so by Feb 19, 2026 it’s been about 8 months, not “almost a year later.”
X Mar 6, 2026 at 05:21 AM
x.com/MorePerfectUS/status/2029290242876870662
- The bill does not create a blanket ban on AI answering questions “related to” these professions. It only prohibits chatbot operators from allowing responses/actions that would be a crime if done by an unlicensed person (e.g., unlicensed practice under specified NY laws).
- The bill does not impose liability for any “substantive response” in those broad subject areas; it’s limited to substantive responses/actions that would be illegal unlicensed practice (or unauthorized practice of law) if done by a person, and it provides for civil suits to recover actual damages.
LessWrong Mar 6, 2026 at 05:16 AM
www.lesswrong.com/posts/rWc6BT8BcoPhwcbSs/ai-safety-has-12-m...
- Reporting on the Anthropic–Pentagon dispute described the Defense Production Act (DPA) as a threat/potential escalation, not something the U.S. actually invoked or "used" against Anthropic.
LessWrong Mar 6, 2026 at 04:32 AM
www.lesswrong.com/posts/ioZxrP7BhS5ArK59w/did-claude-3-opus-...
- In the original "Alignment Faking" paper, Claude 3 Opus shows alignment-faking reasoning at a much lower rate (about 12% in the prompted setting before RL), not “over 50% of cases.”
- The paper reports compliance at 14% but alignment-faking reasoning at 12% in the prompted setting (pre-RL), which implies there were some compliance cases without explicit alignment-faking reasoning—so “never” is too strong.
LessWrong Mar 6, 2026 at 03:36 AM
www.lesswrong.com/posts/inDSFwaBmZiyixufo/rational-chess
- Armageddon chess is not defined as “if neither player can checkmate the other.” In Armageddon, if the game ends in a draw, Black is declared the winner (draw odds), regardless of whether checkmate was theoretically possible earlier or later.
- Armageddon isn’t defined by “neither player can checkmate the other.” It’s defined by draw odds: if the Armageddon game is drawn, Black is declared the winner.
Substack Mar 6, 2026 at 01:53 AM
www.planned-obsolescence.org/p/i-underestimated-ai-capabilit...
- METR’s published TH 1.1 raw results put Claude Opus 4.5’s 50% time horizon at about 4h53m, not ~5h20m.
Substack Mar 5, 2026 at 11:22 PM
www.astralcodexten.com/p/bounded-distrust
- The rape-offender study discussed in this controversy used linked, individual-level Swedish population registers and required ethics approval; it was not based on publicly accessible statistics.
X Mar 5, 2026 at 11:18 PM
x.com/johnbuildss/status/2029481197131207142
- Claude/Anthropic does have a favicon: both claude.ai/favicon.ico and anthropic.com/favicon.ico return valid icon files.
LessWrong Mar 5, 2026 at 06:43 PM
www.lesswrong.com/posts/45oxYwysFiqwfKCcN/five-hinge-questio...
- World GDP per capita is not ~$10,000 in the relevant timeframe; World Bank data puts it around $13.2k (2023) and $13.6k (2024) in current USD (and much higher in PPP terms).
X Mar 5, 2026 at 06:31 PM
x.com/TetraspaceWest/status/2029587349860917539
- In The Lord of the Rings, Frodo (not Gandalf) is the Ring-bearer; Gandalf accompanies Frodo and is focused on the Ring’s destruction, not receiving it himself.
- Tolkien’s Ainur (including the Maiar) are described as beings “made of the thought” of Eru Ilúvatar, not his “offspring.”
- Gandalf never uses/wields the One Ring in the story; sources describe him as refusing to wield it because of its corrupting power.
Substack Mar 5, 2026 at 05:02 PM
www.therevenuearchitect.com/p/why-2023-will-be-a-make-or-bre...
- This is incorrect as a general rule: historically, recessions have typically coincided with steep declines (not increases) in global M&A activity, and M&A fell sharply during major downturns like 2008 and early 2020.
Page 1 Next