Latest Corrections
X Apr 8, 2026 at 10:53 PM
x.com/bindureddy/status/2042001592027877708
- That exact SWE-Bench Pro score is not possible. SWE-Bench Pro has 1,865 problems, so reported Pass@1 scores move in steps of about 0.0536 percentage points; 99.99% cannot be produced from that benchmark.
Wikipedia Apr 8, 2026 at 09:05 PM
en.wikipedia.org/wiki/Nextcloud
- This language count is outdated. Nextcloud’s own website says it is translated into more than 100 languages, including 107 languages in 2024.
- These desktop OS requirements are outdated. Nextcloud’s current official docs list Windows 10+ and macOS 12+ or 13+, not Windows 8.1 or macOS 10.14.
- This sentence is outdated: several of the listed items are already existing Nextcloud features, not merely planned ones.
- OAuth2 and OpenID Connect are not documented by Nextcloud as two-factor authentication methods. Nextcloud documents them separately as authentication/SSO mechanisms.
X Apr 8, 2026 at 08:46 PM
x.com/MaxRovensky/status/2041910644375478397
- This is incorrect because security controls are not “100% impenetrable” by default, and OpenAI’s own safety documentation says untrusted text can still be used for prompt-injection attacks.
LessWrong Apr 8, 2026 at 08:24 PM
www.lesswrong.com/posts/ZfbChZBXgje8T6Geu/excerpts-and-notes...
- Anthropic’s system card does not say the contribution was unreal. It says the contribution was real, but smaller or differently shaped than first understood, and often reflected execution of a human-specified approach.
Substack Apr 8, 2026 at 08:12 PM
www.betonit.ai/p/reflections-of-a-neotenous-man?triedRedirec...
- Available obituaries and institutional tributes place Walter E. Williams’s death on Wednesday, December 2, 2020, after teaching his final class on Tuesday, December 1—not a few minutes after class ended.
X Apr 8, 2026 at 07:55 PM
x.com/aakashgupta/status/2041545816750567573
- This misstates Richard Wiseman’s findings. Published summaries of his luck research report significant differences on multiple traits—extroversion, neuroticism, and openness—not just openness, and they do not present education or family wealth as variables that were ruled out as the "single biggest differentiator."
X Apr 8, 2026 at 07:46 PM
x.com/Hesamation/status/2041649291991904565
- Anthropic’s writeup does not describe Mythos needing an ordinary file permission, nor deleting itself or the logs. It describes Mythos being explicitly tasked with finding/exploiting software vulnerabilities, and later suppressing a specific kernel warning message inside an exploit chain.
X Apr 8, 2026 at 07:41 PM
x.com/Moleh1ll/status/2041809409802641862
- Anthropic’s official Mythos materials do not describe the four examples listed here. The Mythos announcement and Red Team post are about cybersecurity evaluations and highlight bugs/exploits in OpenBSD, FFmpeg, Linux, browsers, and FreeBSD instead.
- Anthropic did not withhold Mythos entirely. It launched Project Glasswing, gave Mythos Preview access to launch partners and 40+ additional organizations, and says participants will be able to keep using it after the preview; Anthropic only said it would not be made generally available.
X Apr 8, 2026 at 07:39 PM
x.com/aakashgupta/status/2041545816750567573
- Wiseman’s own account of the research does not say openness was the only or singular key trait. He says lucky and unlucky participants differed on three personality dimensions: extroversion, neuroticism, and openness.
X Apr 8, 2026 at 06:41 PM
x.com/Dorialexander/status/2041817479488389324
- This is contradicted by Meta’s own documentation: Meta AI on WhatsApp and meta.ai can use Llama 405B, and Meta’s Llama 3 paper says that model is a dense 405B-parameter transformer. Because GPT-3 had 175B parameters, a deployed dense 405B model has more active parameters than 2020 GPT-3.
X Apr 8, 2026 at 04:34 PM
x.com/Docneuroeo/status/2041838167259746669
- Anthropic did say Mythos found a 27-year-old OpenBSD bug, but the “five million automated test runs” detail refers to a different FFmpeg vulnerability, not the OpenBSD one.
X Apr 8, 2026 at 04:34 PM
x.com/JFPuget/status/2041819349984448848
- This is inaccurate for ARC-AGI: the official ARC-AGI benchmark pages list public evaluation sets for both ARC-AGI-1 and ARC-AGI-2. ARC-AGI also has semi-private/private eval sets, but it is not a benchmark whose evaluation data is simply unavailable publicly.
- This is too absolute. Published contamination studies do not support the claim that older public benchmarks measure only memorization; on several major benchmarks, scores changed little after potentially leaked examples were removed, and audits have found little evidence of pervasive contamination.
X Apr 8, 2026 at 04:26 PM
x.com/TheZvi/status/2041841249364140293
- This overstates the 8% figure. Anthropic’s official Opus 4.6 system card says scratchpad/CoT content was considered on less than 0.01% of Opus 4.6 training episodes, not 8%.
Substack Apr 8, 2026 at 01:20 PM
www.astralcodexten.com/p/your-book-review-zuozhuan
- 《左传》传统归属的作者通常写作“左丘明”，不是“左启明”。
- 公元前771年被杀、导致西周终结的君主是“周幽王”，不是“周游王”。
- 这里应是“士阶层”（shi class），不是“石阶层”。“士”是先秦中国的重要社会阶层，“石”则把概念写错了。
- 这里应是“齐公”而不是“齐王”。孔子参与夹谷之会时，齐国君主是齐景公；齐国君主到更晚的战国时期才称王。
Substack Apr 8, 2026 at 01:19 PM
www.astralcodexten.com/p/the-psychopolitics-of-trauma
- Complex PTSD is recognized by WHO in ICD-11, not ICD-10. WHO’s own materials say complex PTSD was added in ICD-11, which took effect on January 1, 2022.
- The DSM’s PTSD criterion is narrower than this. It does not say someone qualifies by merely hearing about traumatic events happening to other people in general.
Wikipedia Apr 8, 2026 at 12:43 PM
en.wikipedia.org/wiki/Stephen_Cohen_(entrepreneur)
- The SEC filing shows the $108.28 sale was on February 21, not the previous day, and the February 20 sales totaled about $31.6 million, not $38.4 million.
Wikipedia Apr 8, 2026 at 12:26 PM
zh.wikipedia.org/wiki/%E4%B8%AD%E5%B7%9E%E9%9F%BB
- 把1913年的國音審定歸給「國語統一籌備委員會」是時間線錯置。1913年進行相關審定的是「讀音統一會」；「國語統一籌備會」到1919年才成立，1928年才改組為「國語統一籌備委員會」。
Wikipedia Apr 8, 2026 at 12:23 PM
zh.wikipedia.org/wiki/%E5%8D%81%E4%B8%89%E8%BE%99
- 這裡把《中原音韻》的韻名寫錯了：應是「侵尋」，不是「侵覃」。
- 這裡把《中原音韻》的韻名寫錯了：應是「桓歡」，不是「桓觀」。
Wikipedia Apr 8, 2026 at 12:02 PM
en.wikipedia.org/wiki/Kleiner_Perkins
- Ilya Fushman did not join Kleiner Perkins from a firm called “Index Partners.” Kleiner Perkins’ own bio says he was previously a general partner at Index Ventures.
- Kleiner Perkins’ official announcement says KP19 was announced on March 5, 2020, not January 31, 2019. The firm also said KP19 focused on Enterprise, Consumer, Hardtech, Fintech, and Healthcare — not just the four sectors listed here.
Substack Apr 8, 2026 at 10:39 AM
theinnermostloop.substack.com/p/welcome-to-april-7-2026
- MoonRF had not released the hardware yet as of April 8, 2026, and its own site says key hardware components are proprietary rather than fully open-source.