EverydayApparatus
ReadsListenConceptsBrief
AllIntelligenceSoftwareSecurityPhysicsMathematicsLifeChemistrySociety

Intelligence

1 month agoopenalexIntelligence

The Identity No One Can Fake — Because No One Knows What It Says

Hardware counterfeiting costs billions a year, and the best defense turns out to be a fingerprint that physics writes and only physics can rewrite.
1 month agoarXivIntelligence

The AI That Doesn't Need to Live through It Twice

To explain their decisions and catch their biases, AI systems need to answer "what if" — and it turns out that costs half as much as we thought, once you notice that looking and reasoning are two different things.
1 month agoarXivIntelligence

What the Camera on Your Face Will Never See

A head-mounted AI can now read your whole body's actions — even though its only camera faces outward and never sees you.
1 month agoarXivIntelligence

To Fix Biased Data, You May Need to Delete More and Add Less

Removing the right records can cut the cost of bias mitigation by 62 percent, with almost no hit to accuracy.
1 month agoarXivIntelligence

Some Ais Can Be Talked into Anything. One Training Step Is the Difference.

A common trick warms an AI up with harmless questions, then slips in a forbidden one. The real defense isn't better rules — it's a quiet phase of training that not every model gets.
1 month agoarXivIntelligence

The AI That Learns without Changing Its Mind

Your voice assistant still can't say "Nguyen" — not because fixing it is hard, but because fixing it the obvious way breaks everything else.
1 month agoarXivIntelligence

The AI That Tries the Side Door before Calling for Help

When an AI assistant hits a wall, how it recovers turns out to matter more than how smart it is — and the answer is embarrassingly human.
1 month agoarXivIntelligence

The AI That Reads Your Assumptions at the Last Moment

Every AI model has its beliefs baked in at training time. A new approach lets you hand a model your assumptions as data, right before it predicts, and it reasons as carefully as a statistician who never has to sleep.
1 month agoarXivIntelligence

Nine Numbers Beat Fifty Thousand

For decades, science has used neural networks as powerful black boxes. An AI just wrote a nine-parameter equation that outperformed one with fifty thousand — and unlike the neural network, you can actually read it.
1 month agoarXivIntelligence

When AI Grades AI, the Danger Is the Wiring, Not the Model

We're building AI pipelines where models evaluate each other's work — and it turns out the risk isn't in any one agent, it's in how they're connected.
1 month agoarXivIntelligence

AI Was Built to Serve Millions. That's Why Your Robot Hesitates.

The design that makes cloud AI cheap is the wrong tool for the AI that will live in your house — and it's quietly taxing every device that runs it alone.
1 month agoarXivIntelligence

The Two Clocks Running inside Every GPU

Modern AI depends on hardware that can fail in more than one way at once — and predicting which failure wins the race turns out to require a model that thinks like an engineer, not like a statistician.
1 month agoarXivIntelligence

Fair AI Was Supposed to Need a Coin Flip. It Doesn't.

The assumption that blocked fully auditable, group-fair AI just turned out to be wrong.
1 month agoarXivIntelligence

The Best AI Coder on the Leaderboard Is Only the Best at Python

There's a single number meant to tell you which AI to trust with your code. It's been quietly measuring the wrong thing — and the gap is wide enough to pick the wrong tool.
1 month agoarXivIntelligence

The AI That Got Smarter without Practice

A new approach lets AI agents sharpen what they know by imagining experience instead of having it — and filtering that imagined experience more patiently than any real trial could.
1 month agoarXivIntelligence

The Gold Standard for Data Privacy Was Designed for an Attacker Who Doesn't Exist

When Apple and the Census Bureau say your data is "differentially private," they're proving something real — just not about the threat you're probably worried about.
1 month agoarXivIntelligence

All That Machinery Was a Workaround

The hardest part of teaching AI to reason about geometry turned out to be optional — a consequence of a naming choice researchers made without realizing they had made it.
1 month agoarXivIntelligence

The AI That Reasons in Private Can Be Taught to Show Its Work

A new breed of AI seemed to think in an unreadable private language. It turns out a whisper of plain words at each step is enough to make it watchable.
1 month agoarXivIntelligence

When Every Expert Is Right and the Whole System Is Wrong

A flaw in how AI systems combine specialist opinions can make them dangerously overconfident — and auditing each part separately won't catch it.
1 month agoarXivIntelligence

The Best Rule in Machine Learning Was Measured with a Crooked Ruler

Everyone knew that training an AI on small bites of data makes it smarter in the wild. The popular explanation for why turned out to be measuring nothing at all.
1 month agoarXivIntelligence

The Radiology AI That Can Finally Point to What It Sees

AI can already write your scan report. The problem is nobody knows where it was looking — and that may matter more than whether it's right.