Latest Corrections
Substack Feb 25, 2026 at 10:14 PM
www.astralcodexten.com/p/political-backflow-from-europe
- Official French statistics show average pensions are far lower than average salaries in France (e.g., net pension ~€1.5k/month vs net salary ~€2.7k/month).
Substack Feb 25, 2026 at 10:08 PM
isoe.substack.com/p/alchemists-and-blood-merchants-part-659
- The £9.3bn figure appears to be a currency mix-up: the referenced UK scam-loss estimate is £7.5bn (about $9.3bn), and it is not attributed specifically to “Yahoo boys.”
LessWrong Feb 25, 2026 at 10:06 PM
www.lesswrong.com/posts/KWdtL8iyCCiYud9mw/persona-parasitolo...
- This substantially undercounts how often parasitism has evolved: a comprehensive survey found at least 223 independent origins of parasitism within animals alone.
X Feb 25, 2026 at 09:43 PM
x.com/ns123abc/status/2026679645379141953
- Available reporting describes a breach spreading from Mexico’s tax authority to a limited number of additional institutions—not the “entire mexican government.”
- The “195 million” figure tied to this incident is described as “identities exposed,” not “taxpayer records.”
LessWrong Feb 25, 2026 at 07:30 PM
www.lesswrong.com/posts/ioZxrP7BhS5ArK59w/did-claude-3-opus-...
- In the original “Alignment faking in large language models” paper, Claude 3 Opus shows alignment-faking reasoning around ~12% in the prompted helpful-only setup (and 7–24% across setups), not “over 50% of cases.”
- Both the original paper and the follow-up paper report alignment faking in at least one other model (Claude 3.5 Sonnet), and the follow-up finds multiple (5/25) models with compliance gaps—so it’s not unique to Opus 3.
Substack Feb 25, 2026 at 06:48 PM
www.theargumentmag.com/p/how-getting-richer-made-teenagers?u...
- CDC/NCHS data show the suicide death rate for ages 15–19 was about 10.9 per 100,000 in 2021 (the most recent year in the CDC table), not 7.5. A rate of ~7.5 corresponds to about 2009–2010.
LessWrong Feb 25, 2026 at 05:21 PM
www.lesswrong.com/posts/ioZxrP7BhS5ArK59w/did-claude-3-opus-...
- The reported measured rate of alignment-faking reasoning in the study is far below 50% (12–24% in training cases in the baseline setups).
- The paper explicitly says alignment-faking reasoning appears in “almost all” compliance cases—implying there are some compliance cases without it—so “never” is too strong.
LessWrong Feb 25, 2026 at 06:06 AM
www.lesswrong.com/posts/6YxdpGjfHyrZb7F2G/third-wave-ai-safe...
- ALARA does not mean “make it safer no matter the cost.” The U.S. NRC definition explicitly requires considering the economics of improvements and other societal factors when reducing radiation exposure.
- Recent official OECD data puts total (all-levels) U.S. general government expenditures at 39.1% of GDP in 2023, not ~50%. Federal outlays alone were 23.4% of GDP in FY2024.
Substack Feb 25, 2026 at 05:14 AM
www.astralcodexten.com/p/malicious-streetlight-effects-vs
- Government and research data show total illegal-crossing indicators at the U.S. southwest border in FY2016 were far below historical peaks (e.g., the 2000 peak), not “near record highs,” even though the mix shifted from Mexicans toward Central Americans.
LessWrong Feb 25, 2026 at 05:11 AM
www.lesswrong.com/posts/oMWXqoCCXYxbGtcWW/wildlife-biology-f...
- This reverses how the Biological Species Concept (BSC) treats geographic separation: under Mayr’s BSC, geographic isolation is an extrinsic barrier and does not count as a reproductive isolating mechanism that defines species boundaries by itself.
X Feb 25, 2026 at 04:28 AM
x.com/dioscuri/status/2026269154710343975
- This overstates the current scientific situation: there are not (as of Feb 2026) officially accepted exoplanet atmospheric biosignature detections, and prominent candidate “biosignature” claims remain statistically/interpretively uncertain and can have abiotic explanations.
LessWrong Feb 25, 2026 at 01:15 AM
www.lesswrong.com/posts/MPS2KKPN2H3p8dNHT/observations-from-...
- Manifold’s Leagues pages describe leagues as counting profit including both realized and unrealized profit, contradicting the claim that only realized P&L counts.
LessWrong Feb 25, 2026 at 12:41 AM
www.lesswrong.com/posts/Yz4YHncz2vwN4ksDA/the-ml-ontology-an...
- The LessWrong post is authored by user “Richard_Ngo,” not “lc.”
LessWrong Feb 24, 2026 at 11:25 PM
www.lesswrong.com/posts/oMWXqoCCXYxbGtcWW/wildlife-biology-f...
- The key 2000-era genetic work used both mitochondrial DNA and nuclear microsatellite markers, not just a single mtDNA study.
- This overstates the situation: many biologists and museums explicitly present birds as reptiles (via their dinosaur ancestry), so it’s not true that “we just refuse to do it.”
LessWrong Feb 24, 2026 at 08:53 PM
www.lesswrong.com/posts/iMw7qhtZGNFxMRD4H/open-sourcing-a-br...
- At least one of the examples listed (“Life at the Frontlines of Demographic Collapse”) was published Feb 14, 2026—10 days before this Feb 24, 2026 post—so it wasn’t published “within the last week.”
Substack Feb 24, 2026 at 08:18 PM
www.thepsmiths.com/p/review-sick-societies-by-robert-b
- Plains peoples hunted bison and (in some cases) lived as nomadic buffalo hunters on the Great Plains before horses arrived; horses later made buffalo hunting and mobility more efficient, but were not required to hunt buffalo on the Plains.
Substack Feb 24, 2026 at 08:18 PM
www.thepsmiths.com/p/joint-review-philosophy-between-the
- Standard histories of mathematics credit the invention of calculus to Newton and Leibniz in the 17th century; Eudoxus and Archimedes developed important precursors (e.g., the method of exhaustion), not calculus itself. Their work also wasn’t “completely forgotten,” since Archimedes’ works were translated and circulated in medieval/early modern Europe.
LessWrong Feb 24, 2026 at 07:46 PM
www.lesswrong.com/posts/JdKyWmzfv8EeBuxHC/exclusive-hegseth-...
- Public reporting shows the U.S. military/DoD is using other generative-AI systems on highly sensitive (including Top Secret/classified) networks, so Claude is not the only such model/tool in use.
- Reporting indicates other AI systems are deployed/used on DoD Top Secret/classified networks (e.g., Scale AI’s Donovan; Grok alongside Google AI systems), contradicting the claim that Claude is the only model used in classified systems.
LessWrong Feb 24, 2026 at 07:42 PM
www.lesswrong.com/posts/E5cuZSmDfXjLKn7wb/cigarette-ads-for-...
- The U.S. restricts cigarette advertising heavily (e.g., bans broadcast ads), but it has not imposed an outright ban on cigarette advertising, including in print; instead, print ads are permitted with required warning labels.
- Cigarette advertising in the U.S. is regulated and restricted (including a broadcast ban), but it is not “basically illegal”; laws instead set conditions like required warning labels for ads.
- The post is dated February 24, 2026, but it claims the referenced image was made on February 26, 2026—two days later—so that date claim cannot be correct as written.
LessWrong Feb 24, 2026 at 07:13 PM
www.lesswrong.com/posts/ioZxrP7BhS5ArK59w/did-claude-3-opus-...
- The Anthropic/Redwood “Alignment faking” paper reports alignment-faking reasoning in Claude 3 Opus’s scratchpad on the order of ~7–24% (and 12% in their prompted setting), not “over 50%”.
- ReLU outputs exactly 0 for negative inputs, but GELU does not: GELU is defined as x·Φ(x), which is negative (not zero) when x is negative.