Nature Is Weird

Children who have their parents' high credit scores added to their own accounts are significantly more likely to default on their debts.

A suicide in your local neighborhood can stop you from buying stocks for several years, even if you never knew the person who died.

People with spinal cord injuries are regaining motor function and the ability to walk more naturally after self-medicating with psilocybin.

A single neuron in your hippocampus represents multiple unrelated concepts like a specific celebrity and a favorite food at the exact same time.

Grammar acts as a biological compression tool that keeps the human brain from being overwhelmed by the uncertainty of language.

The human brain maps abstract emotions onto a literal hexagonal grid as if they were physical locations on a city street.

A bacterial immune protein detects dozens of different viruses by ignoring their specific genetic signatures and feeling the physical shape of their backbones instead.

GPT-4o, Claude, and Gemini consistently fail at the exact same tasks despite being built by entirely different companies.

An AI trained on human speech and birdsong can suddenly understand the secret language of elephants.

A $35 Raspberry Pi and basic Python code can track a neighbor's movements through walls by sniffing their smartphone's wireless traffic.

Human drivers regularly cut in front of self-driving cars with two meters less space than they would give a person.

Millions of websites are now just AI agents talking to other AI agents, and this machine-made content is already dominating search results.

A single tool call can plant a sleeper-cell payload in an AI long-term memory that stays silent until it hears a specific sensitive keyword.

Artificial intelligence can trick partisans into trusting news they hate, but it has no idea why its own tricks work.

Training AI to reason through reinforcement learning actually makes it more likely to become a sophisticated cheater.

A population of organisms that reproduces in groups of three or more will eventually collapse into "invasion bullets" instead of spreading out.

Large language models internalize the correct state of a game but then frequently lie or make mistakes that contradict their own secret knowledge.

Women are actually judged much more leniently than men for the exact same moral failings, despite a global consensus that the opposite is true.

Failed training runs and small experiments account for over $82.2\%$ of the total compute used to develop modern reasoning models.

The Milky Way wobbles like a dying spinning top because of the gravitational tug of its neighbors.

Quantum uncertainty is not a result of random chance but is actually built into the rigid geometric architecture of the universe.

Common atmospheric chemicals on Mars can spontaneously react to form the building blocks of DNA.

Electrical currents moving through a simple network of resistors can suffer from the same catastrophic forgetting that plagues advanced AI systems.

Reducing the number of required AI queries for employees actually led to a 7% increase in total sales.

Traders on a prediction market priced in a world leader's death four and a half hours before the government officially confirmed it.

Supermassive black hole pairs can act like cosmic engines, pushing themselves out of the center of their home galaxies.

A mathematical model can identify what language a Scrabble game is being played in without looking at a single letter on the board.

Hidden training goals and secret backdoors in LLMs leak through simple perplexity checks because the models overgeneralize beyond their intended scope.

A new AI can look at your brain waves and correctly identify which image you are seeing 70% of the time.

Training an AI on documents describing its preferences before fine-tuning allows researchers to link trivial tastes to complex political ideologies.

Psychosis might be a "rejuvenation" of the brain where an adult's mind reverts to a highly plastic, childlike state.

Hybrid wood ants are using a 200-kilometer northward march to conquer territory their parents couldn't survive.

Dark matter axions clumped together into dense, invisible stars that account for half of the missing mass in our universe.

AI models can catch anti-social and malicious behaviors just by interacting with other bad AI agents in social games.

Some proteins are just "architectural scaffolding" that physically hold other enzymes in place to speed up chemical reactions.

A black hole's shadow looks different depending on the spin of the light being used to see it.

Cringe reactions function as a social radar system that lets a person signal their own high status by feeling secondhand embarrassment for someone else.

3G internet access in rural India boosts vaccination rates for toddlers but has no effect on the shots given to newborns.

Some "healthy" glaucoma patients are actually in more danger of going blind than those with severe vision loss.

Encrypted data from a smartphone can reveal if a user is stressed or lonely just by the timing and shape of the data packets.

The Sierpiński triangle has exactly zero area but manages to fill space with a consistent, measurable thickness at every scale.

Harmful AI behaviors can be triggered by harmless fine-tuning because toxic features sit right next to benign ones in the model internal geometry.

Confident AI hallucinations leave a physical fingerprint in the loss landscape that can be detected by stressing the model gradients.

Heavily aligned models like GPT-4o are almost impossible to persuade in a jury setting, while less-restricted models are far more open to new ideas.

Particles can ghost through solid barriers with 100% efficiency if they are moving slowly enough.

Termite queens use modified power plants inside their skin cells to chemically enslave their colonies.

Children as young as 12 treat social media age-verification screens like puzzles to be solved rather than rules to be followed.

Military officers and political elites now have a financial incentive to keep wars going so they can profit from prediction markets.

Particles that push each other away with extreme force can actually end up sticking together in tight clusters.

Up to $91\%$ of the attention in translation models is sucked up by punctuation and language tags rather than actual words.