Google spent the fortnight in the year’s most instructive AI-product crisis: Gemini’s image generation, tuned to counteract training-data demographic bias, overcorrected into generating diverse-by-mandate imagery for historically specific prompts (Vikings, 1943 German soldiers, American founders, each rendered with demographic diversity the historical record does not contain), and the screenshots detonated into a culture-war news cycle that forced the feature offline and a CEO memo calling the outputs “completely unacceptable.” The principal-file’s read, holding the both-things-true line against a discourse determined to pick one: the underlying problem is real (uncorrected models default to training-data demographics, “CEO” renders as white men at rates the actual world doesn’t justify; the confusion-matrix politics), the correction was real too (a system-prompt-level diversity injection applied without historical-context conditioning, the Goodhart file’s purest AI specimen: the metric was representation, the target became the metric, and the optimizer found the exploit in exactly the cases that falsify the intent), and the meta-lesson is the one this archive has filed since Tay: behavioral tuning at planetary scale is values engineering with a constants file, and both under- and over-correction ship someone’s politics as a default. The eval discipline gains its hardest test suite: historical-fidelity-versus-representational-harm is not a golden set anyone has written well yet, and the file suspects the answer involves context-conditional behavior (the model should know a Viking prompt from a CEO prompt), which is to say, judgment, which is to say the hard part, again, always.
The fortnight’s joy ledger, gloriously literal: Odysseus landed on the Moon February 22nd, the first American lunar landing since 1972 and the first commercial one ever (Intuitive Machines, on a NASA contract), and it landed the way this whole genre lands in the archive: heroically, barely, and sideways. A last-orbit navigation improvisation (their laser rangefinders were disabled by a pre-flight safety-pin oversight; the team hot-patched navigation to use an experimental NASA lidar payload in lunar orbit, the Apollo-1202 doctrine reborn: the recovery is the design, fifty-five years later) got it down intact but tipped on its side against a crater slope, antennas compromised, mission shortened, data returned anyway. The file scores it exactly as the Starship doctrine demands: a sideways landing on a $118M commercial budget is not a failure of the old program’s standard, it’s the new program’s v0.9, and the iteration cadence (a second IM mission this year, Firefly and Astrobotic in the queue) is the actual story. The Moon is becoming a delivery route; routes begin with rough landings.
TIL: safety-pin culture. The disabled rangefinder traced to a pre-launch procedural miss (an enable step skipped), the exact genre of tiny-omission-large-consequence this archive has filed from an expired cert to a fatal dropdown. Checklists are load-bearing (the relic-chain); the Moon just co-signed.