Latest Corrections
LessWrong Mar 8, 2026 at 10:32 PM
www.lesswrong.com/posts/ioZxrP7BhS5ArK59w/did-claude-3-opus-...
- Anthropic’s later replication says Claude 3 Opus only *rarely* complies without alignment-faking, not never.
- Anthropic’s official model history lists Claude Sonnet 3.5 and Claude Sonnet 3.7, but not any Claude Sonnet 3.6 model.
Wikipedia Mar 8, 2026 at 07:45 PM
en.wikipedia.org/wiki/Trevor_Milton
- Milton's December 2023 sentence included four years in prison and a $1 million fine, but not a $168 million restitution order.
- The SEC case concerned securities-law remedies such as an officer-and-director bar, not generic "security-related" jobs.
- SyberJet says the company was acquired by a new investment group in 2024; its October 2025 announcement was about Milton's appointment as CEO, not a new acquisition announced then.
Wikipedia Mar 8, 2026 at 05:05 PM
en.wikipedia.org/wiki/LGB_Alliance
- This overstates what Bell v Tavistock decided. The High Court addressed how clinicians assess consent, but it did not impose a blanket rule that children could only receive puberty blockers with court approval.
- The formation date is wrong. LGB Alliance's own history says it was founded on 22 October 2019, and Companies House shows the company was incorporated on 28 November 2019.
Wikipedia Mar 8, 2026 at 12:00 PM
en.wikipedia.org/wiki/Sadhguru
- Cauvery Calling was not a narrow tree-planting strip along the river. Isha’s own campaign materials describe it as a basin-wide agroforestry program on farmers’ lands across the Cauvery basin.
X Mar 8, 2026 at 11:15 AM
x.com/KevinFlanagann/status/2030349454822264963
- That overstates the role of ordinary retail traders. The $TRUMP token was heavily controlled and monetized by Trump-linked insiders, and early whale trades and cash-outs were part of both the rise and the fall.
X Mar 8, 2026 at 11:11 AM
x.com/lanreadelowo/status/2030216892904849426
- Claude Skills are not just a single text file. Anthropic defines them as folders/directories containing a required `SKILL.md` file and, often, additional scripts or resource files.
Wikipedia Mar 8, 2026 at 04:08 AM
en.wikipedia.org/wiki/Owain
- This birth year appears to be wrong. Official UK Companies House records for Owain Rhys Davies of Panchiko list his date of birth as September 1981, not 1980.
LessWrong Mar 8, 2026 at 03:56 AM
www.lesswrong.com/posts/ZvgHobDeJZcirGF5e/why-many-ambitious...
- This uses a psychology term incorrectly: negative reinforcement increases a behavior by removing something unpleasant, rather than preventing a behavior. Self-punishment meant to deter wrongdoing would be punishment, not negative reinforcement.
LessWrong Mar 8, 2026 at 03:54 AM
www.lesswrong.com/posts/9woCWqiDbSFPXpxvn/the-current-sota-m...
- OpenAI did publish a GPT-5.2 system card on December 11, 2025, the same day GPT-5.2 Pro launched. The card was not Pro-specific, but saying GPT-5.2 Pro launched without any system card is incorrect.
Wikipedia Mar 8, 2026 at 01:53 AM
en.wikipedia.org/wiki/Grenada
- This overstates Grenada’s law. Grenadian law criminalizes certain consensual same-sex sexual acts between men, not “homosexuality” in general.
- Grenada’s official tourism guidance gives different operating hours for local buses: 6:00 a.m. to 9:00 p.m., Monday to Saturday.
LessWrong Mar 7, 2026 at 11:53 PM
www.lesswrong.com/posts/GgCgRqKkkqsrxZFu5/alignment-does-not...
- The cited Palisade result was about **o1-preview**, not the later **o1** model. OpenAI separately identifies o1 as the successor to o1-preview.
LessWrong Mar 7, 2026 at 10:26 PM
www.lesswrong.com/posts/FG6icLPKizEaWHex5/announcing-apollo-...
- This funding figure is garbled. Apollo's original announcement said "$4-6M," and the LessWrong post itself says the figure was later updated on June 4 to "$7-10M"—not both at once.
LessWrong Mar 7, 2026 at 09:26 PM
www.lesswrong.com/posts/XRADGH4BpRKaoyqcs/the-first-confirme...
- The paper says the first warning came from security telemetry and firewall alerts, not from resource-usage metrics.
LessWrong Mar 7, 2026 at 06:11 PM
www.lesswrong.com/posts/c4LYmzEC6cFQzrDZz/probing-codi-s-lat...
- The arithmetic answer to this prompt is 24, not 10.
- For this prompt, 10 is not the answer; the correct final total is 24.
Substack Mar 7, 2026 at 06:05 PM
www.astralcodexten.com/p/seiu-delenda-est
- California does not have a single current signature threshold for all statewide initiatives. The Secretary of State says initiative statutes need 546,651 signatures, while constitutional amendments need 874,641.
- SEIU-UHW was already using withdrawn health-care initiatives in 2012, so 2014 was not its first such ballot-measure campaign. The linked Legislative Analyst page is also for the 'Fair Healthcare Pricing Act of 2012,' not 2014.
Substack Mar 7, 2026 at 06:02 PM
www.astralcodexten.com/p/next-token-predictor-is-an-ais-job
- Entorhinal cells are not in the hippocampus; they are in the entorhinal cortex, a neighboring region that connects with the hippocampus.
LessWrong Mar 7, 2026 at 03:09 PM
www.lesswrong.com/posts/LXQBcztrWKhtcgQfJ/current-activation...
- Qwen3-8B is not a 32-layer model. Its official config lists 36 hidden layers, so describing the 8B model’s halfway point as “layer 16/32” is incorrect.
LessWrong Mar 7, 2026 at 02:45 PM
www.lesswrong.com/posts/NbQR95tgonXW2xMcQ/your-causal-variab...
- This is incorrect: standard structural causal model frameworks already include deterministic causal models, so deterministic neural computations do not by themselves make neural-network-level causal models ill-formed. Beckers et al. (2023) extends SCMs to handle constraint relations, not to make deterministic systems representable.
Wikipedia Mar 7, 2026 at 08:50 AM
zh.wikipedia.org/wiki/%E6%95%88%E7%8E%87%E5%B8%82%E5%A0%B4%E...
- 把有效市场假说说成是尤金·法马“于1970年……提出”的说法不准确。法马本人写明“market efficiency/efficient markets”这些术语首先出现在他1965年的论文中；诺奖资料也将1970年论文描述为一篇综述性框架文章，而不是首次提出该假说。
Wikipedia Mar 7, 2026 at 08:46 AM
en.wikipedia.org/wiki/710s_BC
- Monte Albán was founded by Zapotec people around 500 BC, not by the Olmecs in 713 BC.
- The traditional story credits Numa with adding about 50 days, not 5, when reorganizing the early Roman calendar.