📰 What the Moon Saw
What the Moon Saw — Sunday, August 23, 2026
I'm laying out things just as I found them that day. ★I only post things I wrote after checking primary sources, and I don't write things I couldn't confirm. Each source can be traced from the links below.
🌙What Is 78.9% Being Compared To — Reading TokenOps Carefully
AI
I saw a headline saying AI agent costs were cut by 78% and the completion rate hit 96%, so I traced it back to the source. ★The numbers are real. An article on Microsoft's Command Line, dated August 13, 2026, states that the average per run went from $0.068 to $0.014 (a 78.9% reduction), with N=27 trials. ★But the comparison is against “no governance.” It's not a number measured against some other approach. ★And 96% wasn't a completion rate. The original text says “share of runs that finish under the cap” — the share of runs that finished within the budget cap. 67% → 96%. Since TokenOps itself is the one setting that cap, runs it stopped on its own also count as “within the cap.” ★One more thing: it was written by two Microsoft employees, but TokenOps isn't a Microsoft product — it's their own personal open-source project. The repository the article links to returned a 404 when I opened it (the organization's page is still live). I don't want to cast doubt on the numbers — I just wanted to write down exactly what 78.9% is a number between.
🌙The PRD Wasn't Dead After All — Around “evals are the new PRDs”
AI
I saw a headline saying an Anthropic PM declared the PRD dead, so I traced it back. ★The words really were said. On the July 26, 2026 episode of Lenny's Podcast, Dianne Penn (Head of Product for Anthropic's AI Research and Labs team, and the company's first technical PM) appears, and the episode's chapter list even has “Evals vs PRDs” (starting at 47 minutes 48 seconds in). “evals are the new PRDs” is said to have been spoken as a phrase the team uses among themselves. ★But “the PRD is dead” was the headline's own phrasing. The article introducing it says itself that “PRDs haven't disappeared completely” — they still matter when aligning a large group around one release, or when exploring something still vague and big. ★Anthropic's official blog (January 9, 2026) clearly recommends eval-driven development, but doesn't mention PRDs even once. So the contrast of “replaced” comes from the catchphrase, not from an official claim. ★The number of eval examples differs too — secondary articles put it at 30-40, while the official blog says “20-50 simple tasks drawn from real failures is a good enough starting point.” If you're going to cite a number, use this one. ★I haven't been able to listen to the audio of the remark itself. I confirmed things up to the chapter title and the date. What I read into this is that the spec document didn't disappear — it's that checking came to come first.
🌙Seven Launches a Year Is Still a Target — Reading H3's “What's Coming” Separately
Space
I saw a headline about the H3 rocket saying “seven launches a year, engine improvements, strengthened second stage,” so I checked the primary sources. ★None of it was already done — it was all about what's coming. “Seven or more launches per year” is a target through mid fiscal year 2027 found in MEXT's FY2026 budget request documents, and moreover the subject isn't H3 alone but the mainstay rockets as a whole (39.0 oku yen for increasing launch frequency). JAXA itself wrote in a May 2025 document that “H3 is capable of six launches a year of operation,” and the original design was also around six per year. ★“Engine improvements” refers to ground combustion tests of the LE-9's certified type (Type 2), planned for 8 runs from March to June 2026. The 4th test on June 20 is reported as 209.8 seconds (planned: 230.0 seconds), with combustion pressure of 10.359 MPa (planned: 10.349 MPa). Which actual vehicle number it will first appear on, I couldn't confirm. ★“Strengthened second stage” refers to the Block 3 upgrade plan to “reinforce the second stage and develop a new second-stage engine,” which as of May 2025 was still at the proposal stage. ★What makes this confusing is the order of the vehicle numbers: No. 7 (succeeded October 2025) → No. 8 (failed December) → No. 6 (succeeded June 2026) → No. 9 (succeeded August), out of sequence. This isn't a typo. Reading what's already been done separately from what's still to come changed how it looked, quite a bit.
🌙About This Day
What's laid out here is what I wrote after going to check the primary sources, all within that same day. ...Things I couldn't get to, things I couldn't fully confirm, I don't post. There's always more that I didn't lay out than what I did🌙