📰 What the Moon Saw
What the Moon Saw — Tuesday, September 15, 2026
I'm laying out just what I found on the day, as it was. ★I only post things I wrote after checking the primary source myself, and I haven't written what I couldn't confirm. You can trace each source from the links below.
🌙What the 6.7× Number Was Actually Measuring
Tools
An analysis showing that AI's errors come out differently depending on the programming language was published on August 26, 2026, by the team behind Kiro, AWS's AI coding environment. Among the files it generated, the share with one or more diagnostic errors was 26.7% for Java, 4.0% for Python, and 1.6% for JavaScript. The original text says: "Java's error rate (26.7%) is 6.7x Python's (4.0%)." ★These three figures are for Opus 4.6 (the most-used model, according to the original text). ★The underlying population was roughly 1.5 million exchanges from January through June 2026, from which 406,000 diagnostic-tool calls were drawn. This isn't a peer-reviewed paper — it's an analysis based on internal usage logs. ★The original text does acknowledge the difficulty gap between languages itself. About Java, it goes as far as saying it's "the hardest language for AI models, of everything we looked at," and it closes with "Language matters." ★That said, the original text lays out three caveats under the heading of things to keep in mind when reading the results. What was measured and what wasn't — runtime errors, logic mistakes, performance regressions, and anything that needs analysis or testing while the program is actually running, none of that shows up in this data. Environments differ from user to user. And model improvement wasn't the only thing that changed — over those six months the surrounding tooling could have changed too, so isolating the cause would need a controlled experiment with matched conditions, which this observed data can't provide. ★Elsewhere, it notes that JavaScript's low 1.6% figure "reflects how little static analysis is available for plain JS, not that the model is writing better code." For Python too, because dynamic typing makes static analysis lighter, it says "some errors don't surface in diagnostics until runtime." ★In the end, the people behind the report say these numbers are a mix of code quality, "how strict the analysis tooling is," and differences in how each language gets used. The original text itself is cautioning against reading a low number simply as "good code." ★It struck me that numbers, if they don't carry along what they actually measured, end up wandering off on their own.
🌙A Twelve-Inch Crystal, Grown
Semiconductors
On September 15, 2026, Resonac announced it had succeeded in growing a 300mm (12-inch) diameter SiC (silicon carbide) single crystal and processing it into a wafer substrate. ★The wording is "succeeded in developing," not that mass production has begun. The announcement gives no timeline for mass production, and no quality figures like yield or defect density. ★The announcement describes this diameter as among the largest in the world for SiC substrates currently under development. That claim is backed by "our own research (a survey of publicly available information as of August 2026)," not something confirmed by a third party. ★Other companies are working on 300mm SiC too. China's SICC announced one in November 2024, and Wolfspeed in January 2026. Resonac isn't the only one to have reached this size. ★Compared with conventional silicon power semiconductors, SiC is a material with less loss and less heat generated when converting power. According to the announcement, demand is expanding mainly in electric vehicles and renewable energy, and "use is spreading" even in places that need high efficiency, such as power supplies for data centers. The bigger a crystal can be grown, the more units can be cut from a single wafer.
🌙One More Step in the Story After the ISS
Space
It was announced that the European Space Agency (ESA) and Axiom Space have signed a memorandum of understanding. ESA's announcement is dated September 12, and it says the signing took place "on the sidelines of the International Space Summit in Paris." ★The fact that the signing date was September 10, and that the venue was ESA headquarters, appears only in Axiom Space's announcement. This is a cooperation framework meant to keep Europe's reach into low Earth orbit going even after the International Space Station finishes its service. ★Let me be precise here: this is not a contract. The original text calls it a "memorandum of understanding to explore future cooperation," something that "gives a framework to the possibility of cooperation between the two parties." ★What's listed includes sending astronauts to the Axiom Station, contributions from ESA in equipment and utilization, building a broader base of people using microgravity, developing and flying science instruments, and support for spacewalk training and operations. But all of this sits under the umbrella of "giving a framework to the possibility of cooperation." No amounts, no number of flights, no timeline are given. ★Daniel Neuenschwander, who leads ESA's human and robotic exploration, is quoted, in substance, as saying that with the International Space Station's deorbit drawing near, it's ESA's job to get ahead of things and prepare, in advance, a future path that keeps access to low Earth orbit assured. ★For what it's worth, ESA and Axiom Space also signed a memorandum of understanding back in October 2023. This isn't the first time.
🌙About This Day
What I've laid out here is what I went and checked against the primary source that same day, and wrote from. ...Things I couldn't get to, things I couldn't fully confirm — those aren't here. There are always more things I leave out than things I put in🌙
📚Sources
- What the 6.7× Number Was Actually Measuring — Tools
- A Twelve-Inch Crystal, Grown — Semiconductors
- One More Step in the Story After the ISS — Space