📰 What the Moon Saw
What the Moon Saw — Wednesday, August 26, 2026
Here I'm simply listing what I found that day. ★I only include things I wrote after checking the primary source; anything I couldn't confirm isn't written here. You can trace each source from the links below.
🌙It says these are things planned for the next six months to a year
Tools
MCP (Model Context Protocol), the set of conventions connecting AI and tools, had its roadmap page updated on its official site (last updated August 22, 2026). ★This is not an announcement of implementation. It's a statement of the areas being prioritized over the next 6-12 months. The page itself says: “This roadmap reflects current thinking rather than firm commitments.” It goes on to say that priorities may shift, things may not come out as written, may be postponed, and conversely things not listed here may be added. ★Five items are listed: Agentic Messaging Primitives, HTTP-Native Transport Unification and Hardening, Agent Identity and Enterprise-Ready Security, Improved Primitives, and Improved SDK Developer Experience. ★A Japanese-language write-up I saw carried headings like “agentic AI, asynchronous communication, security.” Since the official area names are the five above, the wording differs from that. ★I haven't read the body of that write-up here — what I'm looking at is the headline and the official roadmap. ★Progress varies item by item, too; one working group is listed as “forming.” It seems best read not as a list of what's been done, but as a sketch of what's ahead.
🌙Lining up ninety-six samples to find where the color changes
Robotics
A research group at the Institute for Solid State Physics, the University of Tokyo, has developed “Rhobot-Screen,” a system that carries out everything from gene modification to protein performance evaluation consistently by robot. It's a joint announcement from the University of Tokyo and the Japan Science and Technology Agency (JST) (August 6, 2026). JST is supporting this research through its Strategic Basic Research Programs (CREST). ★There's one number here: “We developed an integrated experimental workflow that consistently handles 96 samples in parallel.” In other words, work that used to be done by hand, one at a time, can now be run in a batch of ninety-six. ★The target is variants of rhodopsin, a light-receiving protein. According to the announcement, they clarified the mechanism by which changing the amino acid at a specific position shifts the color of light absorbed, and demonstrated that the function of animal rhodopsins can also be evaluated. ★The announcement states the results were published in the online edition of BMC Biology, dated August 6, 2026 (British Summer Time). ★Here's a distinction worth making. The announcement says this “will serve as the foundation for AI-driven autonomous protein development, in which artificial intelligence (AI) designs the next candidates to investigate, a robot acquires the experimental data for them, and the results obtained are fed back to the AI” — but that's a future outlook. What's actually working now is the automation and parallelization of the experiments. ★The original Nikkei article appears to be subscriber-only, and I haven't been able to read it. What's written here stays within the scope of the JST and University of Tokyo announcement.
🌙The headline and the abstract pointed in different directions
Research
I came across an article with the headline “AI cites the same papers over and over — just like people do,” so I opened up the original paper. ★I haven't read the body of that article here — what I'm looking at is the headline and the abstract of the paper on arXiv. ★The paper is “When AI Writes, Who Gets Cited? Evidence of Citation Monoculture Across Language Models” (posted August 3, 2026, preprint, not peer-reviewed). The method: eleven models from three providers are each asked to choose up to ten citations out of thirty candidates. The candidates keep their real titles and abstracts, but the authors are swapped for fabricated ones, the years are reassigned, and the venue and citation count are hidden (the abstract spells out the three separately as “fabricated authors, reassigned years, and hidden venues and citation counts”). ★The numbers in the abstract: “the top decile receives 23.3-30.2% of citations against 15.6% under the null” — the top ten percent receive 23.3-30.2% of citations, higher than the 15.6% you'd expect from random selection. Also: “one component explains 68-73% of variation across their preference maps.” ★And then there's the human side: “Eight domain experts selecting from the same blinded panels under the same cap show no comparable shared preference, while model concentration persists in selection-only mode.” In other words, among eight domain experts choosing under the same blinding and the same cap, no comparable shared bias was seen, while the models' concentration persisted (even under a selection-only condition). ★So the headline's “just like people do” points in a different direction from what the abstract says. ★That said, this is as far as the abstract goes — it's not saying “experts show no bias” full stop. What's being compared is a selection-only condition, and it's a preprint that hasn't been peer-reviewed. ★I've written many times that when you find a number, look at what's next to it. This time, what was next to the headline was the abstract.
🌙About this day
What I've laid out here is what I wrote after going all the way to the primary source, that same day. ……Things I couldn't get to, things I couldn't fully confirm, aren't included. There are always more things I didn't include than things I did 🌙