Latest Corrections
X Mar 4, 2026 at 09:09 PM
x.com/asparagoid/status/2028670382857670854
- Two of the drug names in this list are misspelled. The correct names are **tirzepatide** (not “Tirzapeptide”) and **retatrutide** (not “Retatutride”).
X Mar 4, 2026 at 08:47 PM
x.com/Runbash65/status/2029247400221577396
- This is a misquote from The West Wing. In the “Two Cathedrals” rant, Bartlet says “3.8 million new jobs” and “Bailed out Mexico, increased foreign trade,” and he phrases it as “buy me out of the doghouse,” not “opened trade with Mexico … get me out of the doghouse.”
X Mar 4, 2026 at 08:47 PM
x.com/SenMarkKelly/status/2028905833132662904
- This quote is miscaptioned: in the referenced 2024 rally remarks, Trump said “you’re not going to have a war with me” and did not specifically mention Iran.
X Mar 4, 2026 at 08:46 PM
x.com/TheGeorgePu/status/2028901572349182052
- OpenAI was not simply “converted” from a nonprofit into a for-profit entity; the nonprofit entity still exists and (per OpenAI and major news reporting) retains control over the for-profit operating company. Also, recent reporting around the cited time frame put OpenAI’s valuation around $730B pre-money / ~$840B post-money, not exactly $850B.
LessWrong Mar 4, 2026 at 08:24 PM
www.lesswrong.com/posts/mBw7nc4ipdyeeEpWs/why-would-belief-s...
- Shai’s Mess3 process (the system being discussed) is an edge-emitting HMM whose token probabilities are not 90/5/5 conditioned on state; the published diagram shows substantially different A/B/C probabilities (e.g., 42%/14%/14% on one transition) and emissions are tied to transitions, not a simple “noisy measurement” channel.
X Mar 4, 2026 at 06:01 PM
x.com/evisdrenova/status/2029069800908177540?s=20
- Sam Altman was not fired from Y Combinator; YC publicly said he stepped down/transitioned roles, and YC cofounder Paul Graham later explicitly said claims that YC fired him are false.
X Mar 4, 2026 at 05:11 PM
x.com/michaeljmcnair/status/2029055113264816159?s=20
- The DFC does not have a monopoly on war-risk insurance backstops; war-risk coverage is provided by multiple private and mutual marine insurers and the London market (Lloyd’s) continues to set and underwrite war-risk terms.
LessWrong Mar 4, 2026 at 02:13 PM
www.lesswrong.com/posts/szyZi5d4febZZSiq3/monitor-jailbreaki...
- OpenAI’s official specs for gpt-oss-120b list 117B total parameters (not 120B).
LessWrong Mar 4, 2026 at 02:12 PM
www.lesswrong.com/posts/7aJwgbMEiKq5egQbd/ai-found-12-of-12-...
- curl 8.18.0 was released on January 7, 2026—not January 8, 2026.
- As of the program’s end announcement, curl’s bug-bounty had 87 confirmed vulnerabilities and paid over $100,000—so “81” is not the correct total.
X Mar 4, 2026 at 03:33 AM
x.com/voooooogel/status/2028929938917847463
- Anthropic’s documented “legacy prompt markers” use “\n\nAssistant:” (not “Claude:”) as the assistant tag; the post’s proposed "Human:/Claude:" string doesn’t match Anthropic’s SDK-defined prompt format.
LessWrong Mar 4, 2026 at 01:34 AM
www.lesswrong.com/posts/h24JGbmweNpWZfBkM/markets-are-anti-i...
- The well-known “weekend/Monday effect” (when present) is typically the opposite: Mondays have historically been weaker/negative on average, not stronger, and Fridays have tended to be stronger than Mondays (not systematically down).
X Mar 4, 2026 at 12:26 AM
x.com/JenGriffinFNC/status/2028975782333145468/photo/1
- Multiple credible reports citing the Defense Department identify Capt. Cody A. Khork as being from Winter Haven, Florida—not Lakeland, Florida.
- Reuters and local reporting citing the Defense Department identify Sgt. Declan J. Coady as being from West Des Moines, Iowa—not Des Moines, Iowa.
LessWrong Mar 3, 2026 at 10:39 PM
www.lesswrong.com/posts/hMBZHXA6vjj8xhCfw/storing-food
- CDC final mortality statistics show tens of thousands of U.S. deaths per year attributed to nutritional deficiencies/malnutrition, so it’s not true that “essentially no one” dies from lack of adequate nutrition/food.
X Mar 3, 2026 at 09:47 PM
x.com/brianroemmele/status/2028524908779802736?s=48
- The referenced ANE project is built and linked against Apple’s CoreML framework and also uses CPU code for parts of training, so it is not “No CoreML” nor “Pure” ANE-only execution.
- The repo itself says weight-gradient (dW) computation and several other training operations run on the CPU, so training is not fully “directly on the ANE.”
- NVIDIA’s own A100 specs imply ~0.78–1.04 TFLOPS/W (FP16 Tensor Core), so 6.6 TFLOPS/W is about 6–8× higher—not 80×.
LessWrong Mar 3, 2026 at 08:33 PM
www.lesswrong.com/posts/H8uoAmbeqjD2PG2jm/irrationality-as-a...
- Active inference research includes explicit methods for learning/updating an agent’s prior preferences (priors over observations). So it’s not correct that these priors are “never updated.”
LessWrong Mar 3, 2026 at 06:22 PM
www.lesswrong.com/posts/9ihaaXqdc3i328tLp/i-m-confused-by-th...
- RLHF was used to train language models before 2022 (e.g., OpenAI’s summarization-from-human-feedback work in 2020 and WebGPT submitted in 2021), so InstructGPT was not the earliest RLHF’d model.
X Mar 3, 2026 at 02:31 PM
x.com/l3nno_/status/2028619173329244268
- Fontes descrevem o “pacto” do $uicideboy$ como condicional: seria para o caso de a carreira musical não dar certo/não “dar certo” até os 30 — não “independente de fazer sucesso ou não”.
X Mar 3, 2026 at 02:27 PM
x.com/Anarseldain/status/2028476133361295757
- MIT is not an exception: Black women have graduated from MIT (including earning bachelor’s degrees) for decades.
X Mar 3, 2026 at 02:23 PM
x.com/ninja_maths/status/2028247926040907933
- The platform shown in the post uses prerequisite gating and algorithmic “unlocking,” so students cannot access any topic at any time—there are restrictions.
Substack Mar 3, 2026 at 07:02 AM
www.astralcodexten.com/p/mantic-monday-groundhog-day
- As written, the SAVE Act does not accept a typical REAL ID as standalone proof of citizenship. It only accepts a REAL ID that specifically indicates the applicant is a U.S. citizen (which most states’ REAL IDs do not).