From GPT-6 Astra to the Fruit Fly Brain: A Week of AI Showing Its Range

The tour opens with a GPT-6 Astra experiment shared on social media: the model is given a robot arm, a paintbrush and a camera, and asked to paint the Golden Gate Bridge in San Francisco. The key point is that nobody teaches it how to hold the brush; it learns to handle the arm and the brush by trial and error. The resulting painting is not photorealistic, but the show is in the process, not the product: nobody codes each motion, and that is where the scale comes from.

Next comes the same model's run through a 48-stage I-am-not-a-robot challenge. The account says it matches or beats the human level at reading distorted text and matching images. The conclusion is blunt: the tests that guarded the internet for years have largely lost their security meaning against advanced models. The guard at the gate can no longer be told apart from the crowd.

The third GPT-6 Astra item is more playful: a user on X builds a playable two-dimensional universe in a short time and ships it. The host argues the real break is personal, not commercial: anyone will be able to design a game to their own taste and play it with whoever they invite. Mass meeting places like Fortnite and GTA 6 keep their value, but the cost of a one-person universe drops toward zero. A side effect is fresh data for robot training: models improve, people create, and the creations become new training data.

The week's heaviest announcement is Meta Muse, a personal AI operator worked through a WhatsApp-like chat interface. It can read mail and draft replies, book travel, keep advancing long-term goals while switched off, and learn preferences to offer suggestions. Access is open in America on iOS, Android and the web. The pricing stands out: about 100 million tokens a week free, with 20 and 100 dollar plans above that. Mark Zuckerberg argues the setup will open earning channels for people.

The science vein opens with a fly video: an enthusiast maps each Doom frame onto a fly neuron and gets the game played, slowly. The real news is from Google: the complete wiring map of the male fruit fly central nervous system is done. The numbers are striking: more than 166,000 neurons, about 11,700 neuron types, most circuits shared but some connections differing by sex, for example the pathways behind mating behavior. The method slices the brain thin, images it, and rebuilds the wiring as a three-dimensional diagram.

Why the map matters is about humans, not flies. Researchers can now trace a signal step by step from eye to brain to the nerves driving motion. The long-term goal is the same detail in the 86-billion-neuron human brain: seeing where connections fail in Alzheimer and Parkinson, diagramming how vision, motion and choice are produced. In the same days, the Pi Zero robot from Physical Intelligence pulls laundry from a machine and folds it. The motions are not yet human-smooth, but like the fly map it reads as a starting flag.

Google DeepMind opened the AlphaGenome Atlas, predicting the effect of roughly 9 billion single-letter changes in human DNA. Each change gets an importance score, positioning the atlas as a screening tool for unexplained rare diseases. The shared example: a missed variant tied to epileptic encephalopathy was found through the atlas and confirmed in the lab, and over 22 percent more genetic links emerged from 54,000 genomes. AlphaFold mapped protein shapes; this atlas maps the meaning of DNA changes. Note: it is not an approved diagnostic system, it remains research-stage.

The closest-to-life example comes from the US Open tennis tournament in America: a young couple shown on the stadium screen cannot find the footage, so they hand an AI operator a find-it task, and it retrieves the clip within 24 hours. It reads like a small anecdote with a large thesis: anything lost in the digital archive becomes findable. Search, archive scan and matching now cost one message.

On the math front, Anthropic's Claude does not re-solve Fermat's Last Theorem; it translates the 129-page Wiles proof from 1995 into Lean, a language computers can check step by step. The striking part is the schedule: work expected to take years finishes in 11 days, with dozens of Claude operators writing 13 million lines of Lean code and proving over 30,000 theorems to the machine. The prize is a template for both producing and independently checking future proofs with AI.

Robots are in the kitchen and on the catwalk: Agility Robotics' warehouse-born humanoid Digit now appears tidying a room and lifting up to 30 kilograms, promising fewer injuries in home moves and refurnishing jobs. At the IFA 2026 fair in Berlin, 11 robot companies took the runway, showing the Engine AI T800, the Agibot X2 Ultra, a quadruped fire-rescue demo and a panda-styled walk. The ceremony was not flawless: the X2 Ultra struggled to stand and one quadruped lost balance stepping back. It played as a status report as much as a parade.

The closing comes from China: ByteDance founder Zhang Yiming personally leads a real-time spatial video world model, with talk of an early October launch window. Three-dimensional worlds imitating physics will underpin robotics and autonomous systems. The week in one line: models step off the showroom floor; game universes, robot arms and world models feed from the same vein.