Latest Corrections
Substack Feb 27, 2026 at 03:37 AM
thezvi.substack.com/p/openai-moves-to-complete-potentially
- Patagonia’s own announcement says only profits not reinvested in the business are paid out as dividends to its climate-focused nonprofit, not “100% of profits” unconditionally.
LessWrong Feb 27, 2026 at 02:34 AM
www.lesswrong.com/posts/ioZxrP7BhS5ArK59w/did-claude-3-opus-...
- The paper reports alignment-faking reasoning in the single-digit-to-20s percent range (7–24%) in training contexts, not “over 50%.”
- The paper itself reports alignment faking in Claude 3.5 Sonnet as well, so the behavior is not unique to Opus 3.
- In the helpful-only setting, the paper reports Claude 3.5 Sonnet has a higher alignment-faking rate than Claude 3 Opus, contradicting “no other model … nearly as often.”
Substack Feb 26, 2026 at 10:11 PM
casualsex.substack.com/p/the-escalation-ladder
- The decision-theory thought experiment is called “Newcomb’s paradox/problem,” named after William Newcomb (no “e” at the end).
Substack Feb 26, 2026 at 09:58 PM
thezvi.substack.com/p/dating-roundup-7-back-to-basics
- Multiple reports citing Tinder/NYT data put women’s right-swipe rate around 14%, not 8%.
X Feb 26, 2026 at 08:36 PM
x.com/shlevy/status/2027087975910093151
- This is incorrect: billions of people were alive when the episode first aired (May 4, 1997), so the probability is not “basically zero” under any reasonable reference class.
X Feb 26, 2026 at 08:30 PM
x.com/JDale_Shoemaker/status/2027093830571377043
- Reporting indicates Shah Alam was found dead on Tuesday night, but officials said it was unclear when he died.
LessWrong Feb 26, 2026 at 07:37 PM
www.lesswrong.com/posts/sjhCqQWvvGLGEaagK/why-is-anthropic-i...
- Greenland is not part of the EU and is a self-governing territory within the Kingdom of Denmark, so the EU cannot “give up Greenland.” U.S. acquisition talk has been framed around Denmark/Greenland, not the EU as the sovereign holder.
LessWrong Feb 26, 2026 at 07:36 PM
www.lesswrong.com/posts/fv6mDtQaC5bneAB8H/anthropic-has-solv...
- The U.S. Cabinet office of “Secretary of War” was abolished in 1947; the modern cabinet position is Secretary of Defense, not Secretary of War.
Substack Feb 26, 2026 at 05:12 PM
www.astralcodexten.com/p/the-pentagon-threatens-anthropic
- The Defense Production Act (DPA) primarily grants authority to the President (and only via delegation to agencies) to prioritize/allocate contracts and materials; it is not simply “a law which lets the Pentagon force companies to do things,” nor does it directly “force [a company] to agree” to renegotiated contract terms.
- DoD “supply chain risk” authorities focus on procurement actions like excluding a source or directing contractors to exclude a source from subcontracts for covered systems—not a blanket ban on any US company “that use[s]” a vendor’s products from doing business with the military.
Substack Feb 26, 2026 at 05:12 PM
www.astralcodexten.com/p/next-token-predictor-is-an-ais-job
- Entorhinal (grid) cells are located in the entorhinal cortex (in the medial temporal lobe/parahippocampal region), not in the hippocampus proper.
X Feb 26, 2026 at 05:07 AM
x.com/Whoadorf17/status/2026743526092915105
- Using the post’s own numbers ($0.75/hr and $42.22/oz), the gold earned per hour would be about 0.01776 oz/hr, not 0.0071 oz/hr.
- By the 1960s the statutory gold-certificate reserve (“gold cover”) for Federal Reserve notes was 25%, not 40%, and it was eliminated entirely in 1968.
X Feb 26, 2026 at 04:33 AM
x.com/KennyTorrella/status/2026666742836003186
- This greatly overstates the number of workers involved (the meat/poultry processing workforce is in the hundreds of thousands, not hundreds of millions) and understates the number of animals affected (poultry slaughter totals are in the billions annually).
LessWrong Feb 26, 2026 at 02:51 AM
www.lesswrong.com/posts/QAB3BEDRziBerNAih/whack-a-mole-is-no...
- The Controlled Substances Act’s schedules are not limited to individually named molecules; they also include broader class/structural definitions (e.g., “Tetrahydrocannabinols” covers synthetic equivalents/derivatives/isomers).
- U.S. law can treat unscheduled “analogues” as Schedule I (for federal law) when intended for human consumption, so new variants don’t necessarily need to be individually scheduled to be controlled/prosecuted.
Substack Feb 25, 2026 at 11:42 PM
www.astralcodexten.com/p/the-pentagon-threatens-anthropic
- The DFARS “supply chain risk” authorities generally allow DoD to exclude a source in specific covered IT procurements (or particular contract/task/delivery orders), not to categorically bar any company that uses a vendor’s products from doing business with the military.
LessWrong Feb 25, 2026 at 11:05 PM
www.lesswrong.com/posts/rmYB4a7Pskw7DLpCh/anthropic-and-the-...
- This is incorrect: multiple other AI providers’ models have been approved for, or integrated into, U.S. government classified environments (including Google’s Gemini and OpenAI models via Azure OpenAI, and xAI’s Grok per DoD statements reported by major outlets).
- This is contradicted by reporting that Grok and Gemini are being integrated/used within Pentagon networks, including for classified documents—so Claude is not the only model deployed in classified Pentagon environments.
Substack Feb 25, 2026 at 10:56 PM
www.astralcodexten.com/p/mantic-monday-the-monkeys-paw-curls
- The 2025 Oscars ceremony (the 97th Academy Awards) took place on March 2, 2025—not March 5, 2025.
Substack Feb 25, 2026 at 10:37 PM
www.astralcodexten.com/p/the-good-news-is-that-one-side-has
- The paper was published online on November 12, 2025 (and listed as an Epub date on PubMed), not “earlier this month” relative to the post date (January 16, 2026).
Substack Feb 25, 2026 at 10:32 PM
www.astralcodexten.com/p/moltbook-after-the-first-weekend
- METR’s published measurements put Claude Opus 4.5’s 50% time horizon at about 4 hours 49 minutes (and later estimates around 5+ hours), not 4 hours.
LessWrong Feb 25, 2026 at 10:30 PM
www.lesswrong.com/posts/eczwWrmX5XNEo7JsS/intricacies-of-fea...
- Whitening makes a random vector’s *covariance matrix* the identity; it does not make the *expected dot product between different sample vectors* equal the Kronecker delta. In particular, for a whitened n-dimensional vector with identity covariance and mean 0, the expected squared norm is n (not 1).
Substack Feb 25, 2026 at 10:26 PM
www.slowboring.com/p/democrats-need-to-think-bigger-on
- This omits major categories of U.S. electric utilities—especially member-owned electric cooperatives (and also federally owned power authorities). Not all U.S. utilities fit into the three formats listed in the post.