AI Is Starting to Study Itself, and It's Already Discovering New Materials on Its Own

AI is starting to study itself. The big idea moving through AI research right now is that models should start learning to understand themselves. The thinking goes that no matter what an AI is asked to do, the one thing that's always in the room is the AI itself, so teaching it to model its own behavior is the next real unlock. Jürgen Schmidhuber, one of the founding figures in the field, laid this out this week as a serious direction for where things go next.

While the big labs work on that, hobbyists are getting surprisingly far with a much simpler idea. A developer this week reverse-engineered a routing tool from the AI lab Sakana into a tiny system called "tinyrouter". It's basically a traffic cop that reads each question, decides which specialist AI is best suited to answer it, and hands it off. Despite being incredibly small compared to the big models, it beat every one of them on a major reasoning test. The takeaway is that a lot of AI performance right now can come from just picking the right tool for the job rather than making the tool bigger.

Raw size still matters, though. Google's upcoming Gemini 3.5 Pro is rumored to be able to handle 2 million tokens of context at once, which is roughly double what Anthropic's current models can hold. In practical terms, that means it can read and reason over entire books or huge codebases in a single conversation, without losing track.

Anthropic's newest model, Claude Fable 5, has been getting attention for a different reason. Google researcher Andy Coenen credited three things for what makes it work: sheer size, a lot of training in long multi-step tasks (not just single questions), and something called a "constitution," which is basically a written set of principles the model learns to follow in ambiguous situations. Anthropic's own prompting guide backs this up, saying the model runs longer on its own without needing hand-holding, and that hard requests around hacking or dangerous biology get quietly routed to a more careful, older model. Developer Simon Willison noted that letting the model use its own judgment tends to work better than giving it strict rules, and it also uses fewer resources, which matters ahead of a coming price increase.

That kind of judgment is now being turned toward actual scientific discovery. Alibaba's research team just used AI to screen 2.4 million potential crystal structures in only 28 hours of GPU time, and they came away with four verified new superconductors. Superconductors are materials that carry electricity with zero energy loss, and they're one of the holy grails in physics because they'd transform everything from power grids to computing. This is the kind of thing that would have taken years the old way.

The developer world is starting to feel the shift too. One builder said this week that letting Claude Fable 5 manage its own team of smaller AI assistants has made hand-written code feel worthless, and that the only thing left for human developers is accountability, being the one who takes responsibility for what the code does. In a fun example of the same trend, someone turned an e-ink tablet into Tom Riddle's diary from Harry Potter, where your written words fade and an AI "spirit" writes back. Fittingly, the era this all replaces is closing, because Amazon is shutting down Mechanical Turk, the marketplace it started in 2005 to pay people small amounts to label the training data that made a lot of today's AI possible in the first place.

Making all this AI cheap enough to run at scale also means making the underlying chips cheaper. Sam Zeloof, a young engineer who became famous for making his own computer chips in his parents' garage, just renamed his company Fab2 and is now running three chip-manufacturing sites across Texas and California, all controlled from a browser. Meanwhile, researchers in Switzerland figured out how to float a single trapped atom above a chip to map its electrical fields with incredible precision, which is a new tool for debugging next-generation processors. Countries are actively competing to host the chip buildout too. France just landed 75 billion euros from SoftBank, and India got a record $48 billion commitment from Amazon.

The AI is also starting to leave the screen. A research group called Nous put their AI model into a robot dog that can see, hear, and hold a conversation. An F-35 fighter jet stole the show on the National Mall for America's 250th anniversary by hovering in place. Tesla is reportedly considering requiring a facial recognition check before letting drivers turn on Full Self-Driving, and framing it like a fitness streak to encourage safer driving habits. And EV batteries are quietly proving they last far longer than critics expected, into hundreds of thousands of miles, which has slowly killed the concern that used to keep a lot of buyers away.

There's also a growing conversation about how much we should be building in space. A European observatory study warned that the 1.7 million satellites currently on the drawing board, which includes proposed data centers in orbit and giant mirrors that would reflect sunlight down to Earth, could effectively blind ground-based astronomy. They're urging a cap at around 100,000 satellites total.

On the health side, a new report counts 158 different Alzheimer's drugs currently in 192 clinical trials, and it shows a real shift over the last decade. Researchers used to focus almost entirely on removing a specific brain plaque called amyloid, but the field is now moving toward treating brain inflammation and immune system problems as the real cause. A separate study from Loma Linda University found that eating five eggs a week is linked to a 27% lower risk of developing Alzheimer's.

History is getting rewritten with better tools too. Researchers found that the "hobbits" of Flores, a species of ancient tiny humans, apparently survived by eating what the local Komodo dragons left behind and never learned to make fire. And using a new three-DNA identification technique, scientists put a name to an American Continental soldier who had been buried in a shallow grave in Camden in 1780. His name was John Pumphrey, and he had enlisted at just thirteen years old. He got his identity back right in time for America's 250th. Related, the President told his new intelligence chief this week to "declassify almost everything", his UAP advisors publicly accused defense contractors of secretly running programs to recover crashed craft, and whistleblowers launched a new platform called Vanguard to protect people who come forward.

The rules of ownership are being rewritten last. The US Treasury launched something called 530A accounts, which give every eligible child a $1,000 starter investment in the stock market, tracked through a Robinhood app. The idea is that instead of handing out cash, the government seeds every kid with a stake in the actual economy. Marc Andreessen predicted this week that America is on track to reach 90% of global market value, and wealthy families are increasingly pulling their kids toward new schools like Forge Prep, which are built to prepare students for the world of 2040 rather than the one we grew up in.