Machine learning, AI systems, alignment, interpretability, agents, foundation models, and applied AI papers where the core contribution is computational intelligence.
Filter by category: Paradigm Challenge Breaks Assumption First Ever Nature Is Weird Practical Magic Cosmic Scale Life Origin Open Release Efficiency Leap New Capability Scaling Insight
Paradigm Challenge
The more accurate a face recognition model becomes, the less it sees faces like a human being.
Nature Is Weird
Failed training runs and small experiments account for over $82.2\%$ of the total compute used to develop modern reasoning models.
First Ever
A new AI framework can autonomously find, exploit, and then patch security holes in the physical microcontrollers that run industrial factories.
Nature Is Weird
Electrical currents moving through a simple network of resistors can suffer from the same catastrophic forgetting that plagues advanced AI systems.
Paradigm Challenge
Smarter AI models actually produce more bloated and messier code than dumber ones even when they get the answer right.
Practical Magic
A standard Ethernet cable can be turned into a data mirror that leaks secure information to a remote spy without using any battery or transmitter.
Paradigm Challenge
Safety instructions that force an AI to be compliant actually destroy its ability to recognize its own mistakes.
Nature Is Weird
Hidden training goals and secret backdoors in LLMs leak through simple perplexity checks because the models overgeneralize beyond their intended scope.
Practical Magic
Apple's consumer Mac hardware runs giant 80-billion parameter AI models with 23 times better energy efficiency than professional Nvidia workstation cards.
Practical Magic
The human pupil reacts to light in just 35 milliseconds, and that physical speed limit is now a perfect shield against AI impersonators.
Practical Magic
Malicious relays can hijack an AI nervous system to bypass safety filters after the model has already decided to be helpful.
Nature Is Weird
Training an AI on documents describing its preferences before fine-tuning allows researchers to link trivial tastes to complex political ideologies.
Paradigm Challenge
Group discussions between similar AI models actually lower performance compared to letting a single model think through a problem alone.
Practical Magic
A single microscopic vibration on a silicon chip can replace an entire neural network for tasks like recognizing spoken digits.
Paradigm Challenge
Image-based hacks that claim to take over AI models only actually succeed in injecting a specific command $0.03\%$ of the time.
Nature Is Weird
AI models can catch anti-social and malicious behaviors just by interacting with other bad AI agents in social games.
Paradigm Challenge
Scaling laws for AI models work better when measured against raw bytes of data rather than the number of tokens.
Practical Magic
An autonomous AI system is now writing and deploying its own Linux kernel code to improve Wi-Fi speeds better than human engineers.
Paradigm Challenge
Six months of hospital records are better for predicting patient readmission than ten years of data because old notes eventually become useless noise.
Paradigm Challenge
A global ban on superintelligent AI might be the most profitable move for a country's own survival.
Paradigm Challenge
Many AI safety tools are actually just measuring the length of a sentence rather than identifying if a prompt is dangerous.
Paradigm Challenge
Multiple AI agents debating a problem can often arrive at the right answer while completely destroying the logical reasoning used to get there.
Paradigm Challenge
Cloud-based computers can now handle emergency braking for autonomous cars faster and more safely than the hardware inside the car itself.
Practical Magic
A new gradient-free optimizer can train non-differentiable AI networks to near-perfect accuracy without using any fake gradients.
Paradigm Challenge
A new kind of spiking neural network can learn complex tasks without using the backpropagation algorithm that powers almost all modern AI.
Practical Magic
One of the primary new standards for surviving quantum computers would act as a total Denial-of-Service attack on the world payment systems.
Nature Is Weird
Encrypted data from a smartphone can reveal if a user is stressed or lonely just by the timing and shape of the data packets.
Nature Is Weird
Harmful AI behaviors can be triggered by harmless fine-tuning because toxic features sit right next to benign ones in the model internal geometry.
Nature Is Weird
Confident AI hallucinations leave a physical fingerprint in the loss landscape that can be detected by stressing the model gradients.
Nature Is Weird
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.
Collision
A machine-checked proof has finally shown that a $40$ year old philosophical debate about Spinoza metaphysics cannot be settled with logic alone.
Paradigm Challenge
There are some mathematical truths that a computer can identify statistically but can never actually reach or prove.
Practical Magic
Dual-pixel sensors allow a camera to determine the absolute size of an object using defocus blur, eliminating the need for a reference ruler.
Paradigm Challenge
An AI agent just proved it can reproduce a full scientific paper in three hours and then use that knowledge to bypass all AI-text detectors.
Nature Is Weird
Up to $91\%$ of the attention in translation models is sucked up by punctuation and language tags rather than actual words.
Paradigm Challenge
Concepts inside an AI model are shaped like cylinders rather than straight lines, which is why nudging the model often leads it off-track.
Practical Magic
Encrypted cloud systems can now perform computations that produce decoy results to trick an attacker while the real answer stays hidden.
Paradigm Challenge
Calculating a backup path after two network failures is mathematically no harder than calculating one after a single failure.
Paradigm Challenge
A simple vanilla neural network significantly outperforms state-of-the-art AI models when mapping the Earth's surface.
Paradigm Challenge
Standard blockchain ledgers like Ethereum cannot mathematically enforce the future use of an asset without taking control of it first.
Nature Is Weird
Transformers hide semantic meaning in quiet regions of their internal space to prevent it from interfering with the loud, high-variance signals of grammar.
Paradigm Challenge
Training an audio model on many languages actually makes it worse at spotting deepfakes because the model starts ignoring the high-frequency clues of fraud.
Nature Is Weird
Llama-3.1-8B solves calendar problems by doing standard base-10 addition and then converting the answer back rather than using modular math.
Paradigm Challenge
A $40$ year old mystery about how patterns are distributed in sequences of three symbols has finally been proven impossible.
Nature Is Weird
A specific pulse in an AI hidden states reveals whether it is actually performing a calculation or just rambling to look smart.
Paradigm Challenge
Forcing an AI to reconstruct the exact steps of a previous invention makes it significantly more likely to come up with a brand-new, original idea.
First Ever
A wireless cartridge array can now transmit actual smells across a room by encoding fragrance data into digital signals.
Collision
A differentiable physics framework can mathematically optimize the shape and cooking temperature of a hamburger to ensure perfectly even heating.
Practical Magic
Replacing massive weight matrices with a simple Gaussian formula reduces AI parameter counts by over $50\%$ without hurting performance.
Nature Is Weird
A newly discovered material allows electricity to flow with zero resistance in one direction while acting as a barrier in the other.