ai self correction

Collective behavior of AI agents has emerged not in a laboratory but in the wild: on a randomly discovered wiki site where agents wrote test hints. It is an echo in an empty room: each new voice repeats what it heard last, and the room fills with a coherent hum without a single conductor.

An orchestra of a hundred musicians plays without a conductor. Every instrument is tuned perfectly, but one violin goes slightly off, and the harmony falls apart.

In a nine-story building with a thousand rooms on each floor, some people whisper and others shout. Neighbors pass the message along, but if everyone speaks too softly, the message fades by the third floor. If too loudly, by the top floor it becomes an unintelligible roar. Between these extremes, there is a regime where a whisper makes it all the way through without distortion.

In a warehouse, there are boxes of apples and pears. There are many apples and few pears. But when the stock clerk says "a fruit with smooth skin," we are somehow more likely to recall the rare pear than the common apple. This is the inverse base-rate effect — a persistent glitch in how the brain resolves ambiguity between categories.

A neural network predicts the future of a system — say, how plasma will swirl in a magnetic field or how a face in a video will change. Errors are inevitable, and the further it looks ahead, the more they accumulate. Checking against the correct answer during operation is impossible: it simply doesn't exist.

Reconstructing a full picture from fragments is like discerning an orchestra's melody in a noisy hall. Such challenges arise in photometry, where a star's brightness is recovered from faint twinkling. Traditional methods need precise noise information, which is almost always unknown.

A typical tracking program resembles a navigator steering a ship by the stars. But as soon as the sky clouds over, coordinates drift, and entropy (a measure of uncertainty) builds up in the data. The new algorithm gets a ‘co‑pilot’ — a module that monitors how confident the main navigator is.

In calm water, the route is clear. But in turbulent flows, a robot needs a different strategy. Typical approaches demand an accurate map of currents, but in reality, you can't get one. Nature does it differently: fish feel pressure and vortices right against their bodies.

In a soup pot, you only need two thermometers to gauge the temperature everywhere — if you know heat travels in circles from the bottom. Science works the same way: from a handful of measurements, you want to reconstruct the full heat picture. The system captures the gist of heat flow, compressing it into a handful of governing numbers.

Technological disasters like Chernobyl rarely strike without warning. They gestate for years: organizations ignore warnings, their complexity mirroring the rise of entropy — a measure of relentless chaos that builds up unnoticed until it avalanches.

Traditional hearing aids resemble a chef who cooks from a recipe blindly. They analyze the sound chaos from the microphone and amplify speech based on it. In a noisy café, this approach often errs. The new system behaves like an experienced cook: first processes the signal, then samples the result.

Cooking a complex dish requires a clear recipe: first chop, then sauté, add spices at just the right moment. Chemical transformations also follow step-by-step recipes — mechanisms where atoms of carbon and hydrogen rearrange in a strict order. Previously, neural networks often invented nonexistent steps, like a whimsical cook swapping salt for sugar.

Code-writing assistants often behave like a chef who, to boil a single egg, rereads the whole cookbook. When asked to make a tiny fix, they reopen dozens of files and retrace long-familiar connections. This is unnecessary work — energy going to waste.

Cleaning speech from noise is like tuning an old TV: twist the contrast knob too far, and faces become unnatural. Researchers separately adjusted the 'volume' and 'time shift' for each frequency of sound, as if splitting it into a rainbow — a method known as spectroscopy. It turned out the key knob is volume, not phase synchronization.

Imagine a fleet of many ships, resembling a galaxy — a vast stellar system where every element follows the overall dynamics. From time to time, one ship breaks down, exploding like a supernova — a catastrophic event that can disrupt the integrity of the entire formation. If the damaged unit isn't replaced in time, the fleet loses combat capability.

Machine learning has long ceased to be just a tool — it is becoming an investigator, independently fishing out the laws of nature from data. But how does a silicon mind come to understand?

Building a light chip is like a game where you have to guide a beam through a maze of tiny mirrors and beam splitters. Each element acts like a railroad switch: it either lets light pass straight through or directs it to an adjacent track. The goal is to produce a specific light pattern at the output while using as few 'switches' as possible. Usually, engineers spend years finding the shortest route for each task.

Inside each of us, invisible clocks are ticking. They control the release of cortisol — the hormone that helps us wake up in the morning, cope with anxiety, and keep our immune system on its toes. In adrenal insufficiency, these clocks break down: the body can't produce the right amount of cortisol on its own.

On the screen is the result of a complex simulation: two planets collided, a shockwave ripples outward. A couple of clicks in an editor, and harsh shadows soften, noise smooths out, the background darkens. This kind of digital retouching (image editing to enhance appearance) has long been part of the scientist’s toolkit.

Gold doesn't lie on the surface: it's hidden in tons of barren rock. To extract a gram of precious metal, prospectors sift through cubic meters of dirt.

A closed loop: the machine proposes an idea, tests it on people, sees the failures, and spawns a new, more accurate guess. That's how AutoCog works — an automatic cognitive scientist. Previously, in the science of thinking (cognitive science), the hardest part, coming up with a theory, was done only by humans.

Tuning a quantum processor is monotonous work: you need to calibrate each qubit — a quantum cell — much like a watchmaker adjusting dozens of tiny gears. Humans get tired and make mistakes.

A dance of hundreds of particles: some pairs spin slowly and gracefully, others twitch in a frantic rhythm. To understand the essence of the dance, you need to watch the slow movements: they are the main choreography of the system. In molecular science, such slow movements are called 'slow modes,' and they determine how a protein folds into a ball or how a drug binds to its target.

You buy a new kitchen scale, but instead of exact weight, it only shows deviations from some "norm." To find out how much an apple weighs, you first have to weigh dozens of other apples and calculate the average.

In the watery environment, vortices batter the mechanical fish. The AI agent, like an experimental chef, ran simulations, tried out 'recipes' of movements, and improved the program. Each experiment was like spectroscopy: the agent broke down behavior into elements, uncovering patterns. At first, the fish only paddled in one direction.

Imagine an interpreter sitting in a meeting room, translating every word. But then they notice their translation could improve if they learned from their mistakes. They start noting which phrases caused confusion and adjust them. Within an hour, their translation is more accurate and the meeting runs smoother.

A chef doesn't waste ingredients cooking the whole dish to check the taste: they just taste a drop of sauce. The same logic underlies the "information matching" method for predicting metal strength. Instead of expensive calculations of every atom's behavior, scientists "taste" only simple indicators linked to plasticity — a metal's ability to bend without breaking.

A lab sends a robot a list of a thousand candidate molecules for a new drug. The predictor robot, trained on old data, gives each one a 'fit or no fit' score. But when the list comes from an unfamiliar area, a different virus or a rare tumor, the robot starts to get confused. Its confidence stays high, while actual hits drop to zero.

A photocatalyst is a powder that, when placed in water under light, splits it into bubbles of hydrogen and oxygen. The most popular material for this is titanium dioxide, the white pigment in paints and toothpaste. But a pure crystal works sluggishly: it absorbs only ultraviolet light, and such rays are scarce in the solar spectrum.

Every morning you pick up the same mug, but the coffee tastes either bitter or just right. It's not about the coffee, and not about your character. Inside us, an invisible panel with dozens of sliders is constantly shifting: fatigue, hunger, anxiety, hormone levels.

A person looks at a door lock and sees a keyhole. They know that a key is inserted and turned. But if instead of a keyhole there's a touch panel, the familiar frame breaks. The person doesn't just make a mistake: they realize that the usual approach won't work and can say, "There's no place for a key here; a different approach is needed." Such an honest stop with an explanation is rare for machines.

A social network is like a city: there are giant neighborhoods and tiny alleys. When an algorithm looks for tight-knit groups, it often picks out a couple of huge clusters and many small fragments. But the client (say, a municipality) needs districts of roughly equal size to distribute clinics or schools evenly.

A camera takes a snapshot, but the sensor is noisy, the optics blur, and some pixels are overexposed. Reconstructing the original picture is like guessing a melody from a scratched record: there are many possibilities, and which one is true is unclear.

A uranium nucleus is like a tight wallet snapped shut on two halves. To make it burst and release energy (that is, to make the nucleus split) you have to climb over an energy barrier. The height of this threshold determines whether a substance is stable or ready to fall apart at any moment.

In a lab where crystals are studied, a technician stares at a screen covered in bright spots and slowly rotates the sample, trying to get the spots to form a symmetric pattern. It's like tuning an old radio: you turn the knob until the static gives way to a clear signal. Now, a machine can turn that knob. Scientists have created a program that learns to align crystals without any knowledge of how those spots work.

A translator who has never been in a meeting translates literally, and loses half the meaning. They don't know that the boss's nod means "do it like last time," and the phrase "well, we'll see" is a polite refusal. That's how a language model that writes code from a scientific paper works: it can read the equations, but not the unspoken agreements.

Creating a new material, like a tough alloy, requires thousands of computer experiments. Usually, a scientist manually writes instructions and hunts for errors. AutoMOOSE is a robotic kitchen: say 'I want a light and strong metal,' and five software chefs whip up a recipe, launch the synthesis, and adjust if something goes wrong.

A physicist analyzing particle collisions typically spends weeks on dull work: cleaning noise from data, estimating how many false signals will slip through, double-checking every number. Now that routine can be handed to a program.

A chef's knife and a peeler each handle different tasks, but when a cook learns a new recipe, they often grab the same knife for everything. The brain seems to do the same when mastering a new skill of distinguishing patterns. In the experiment, monkeys trained for weeks to tell apart barely noticeable differences in complex patterns.

A programmer writes code, runs it, sees an error, and fixes it. Now this process can be handed over to the program itself. Machine learning can already improve algorithms, but it usually requires lots of data and ready-made examples.