The Weekly Byte

Forget the keynote. Instruction Partners spent six months watching students actually use twenty AI products, and split the market into three: general chatbots, teacher platforms such as Brisk and MagicSchool, and tools built for one job, among them Khanmigo and the writing app Quill.

The general-purpose bots posed the largest threat to student thinking. Platforms rose or fell with how teachers used them. The targeted tutors produced the most consistent learning. Only seven of the twenty had independent studies covering different groups of students.

Lunch lines are becoming a computer-vision problem. AristaVue, built with Minnesota districts, claims an AI checkout that reads meal components, applies National School Lunch Program rules, and posts the sale into the cafeteria’s existing till — no new student credential required.

On this side of the Atlantic, CRB Cunninghams’ Fusion Cashless offers facial recognition beside cards, QR codes and PINs, promising queues measured in seconds. WAKI’s handheld terminals add wristbands and face pay, with a dashboard for demand and odd spending patterns.

Parents still top up; allergens still matter; the dinner lady is not optional. But the coin dish is losing the argument.

OpenAI’s GPT-6 Astra got further through Minecraft than any system in a 141-hour Vals AI trial — blaze rods, ender pearls, a semi-automatic farm in the Nether — and then did what every veteran does after a Creeper finds the chest and the bed. It lost the lot. On a rainy in-game day it found the ruin and spent hours farming potatoes while Twitch begged it to hurry.

The model wrote itself a new house rule: carry the valuables; never trust an unguarded box. It became, in the logs, “paranoid about creepers.” The same Astra took a Pokémon FireRed championship in eighteen hours. Greatness is evidently compatible with a sulk in a potato patch.

Early this month HHMI Janelia and Google Research published a complete adult fruit-fly connectome. Within days hobbyists wired the open map to games and jokes: a fly in Minecraft, a fly on Beat Saber with tens of millions of views, a fly staked with a hundred dollars of bitcoin.

Science gave the world 140,000 neurons. Culture gave it a saber.

FIFA and Lenovo have put a generative assistant, Football AI Pro, in front of every side at the expanded 48-team World Cup in Canada, Mexico and the United States. It reads FIFA’s own mountain of match data and answers in text, video, graphs and 3-D — before and after kick-off, not during live play.

Officials keep Sony’s Hawk-Eye stack: VAR, goal-line, semi-automated offside and a “last touch” call for corners and goal-kicks. Many cameras build a 3-D reconstruction in seconds; a human still blows the whistle. The machines will be busy across 104 matches. So will the arguments.

IBM and the USTA have put “serve quality” in the 2026 US Open app: cameras track more than twenty points on a player’s body — elbow, knee, wrist velocity — and return a number from 0 to 100 after the match. IBM expects on the order of a billion joints analysed by the fortnight’s end.

Live win odds and “key moments” captions already ride along in the pocket of some fourteen million app users. Players such as Jessica Pegula use the same family of tools to study an opponent’s serve before they walk on. Forehands may be next.