← All Issues
published
14 article(s) · Artificial Intelligence
2026-03-29
Artificial Intelligence
Deep Learning
SimulationFaster computers
AI Is Now Being Graded on How Fast It Could Theoretically Go, Which Is Very On Brand
arXiv · 2026-03-19
For years, AI systems optimizing GPU code were graded like a student who brings their own rubric: just beat the last guy's software, collect your A, go home. Researchers have now introduced SOL-ExecBench, a benchmark that instead asks how close your code gets to the absolute physical speed limit of the hardware itself. The hardware does not grade on a curve. The hardware does not care about your feelings. The hardware has a Speed-of-Light bound, and your kernel either approaches it or it does not. To prevent AI optimizers from cheating — and apparently they do try to cheat — the benchmark also ships with reward-hacking detection, because we are now at the stage where we need to catch AI systems gaming their own performance tests.
Takeaway
We built a benchmark so rigorous it checks whether the AI is cheating, which means the AI is definitely cheating.
Read paper ↗
Artificial Intelligence
Deep Learning
SimulationFaster computers
Scientists Build AI That Speaks 200 Languages, Still Won't Text You Back
arXiv · 2026-03-19
Researchers have released F2LLM-v2, an AI language model that understands more than 200 languages — including, finally, the ones that bigger, fancier models have been ignoring for years. It comes in eight sizes, from a modest 80 million parameters up to a universe-devouring 14 billion, and was trained on 60 million data samples that were personally verified to be "high-quality," which is exactly what you'd say about data you picked yourself. The 14-billion-parameter version currently sits at the top of 11 different AI leaderboards, which is the machine learning equivalent of winning every category at your own award show.
Takeaway
The AI speaks 200 languages fluently and ranks first on 11 benchmarks — somewhere, a language model that only knows English is having a very quiet crisis.
Read paper ↗
Artificial Intelligence
Deep Learning
Pilot studyRobot navigation
Researchers Discover That Robots Get Confused When You Show Them Blurry Photos and Give Them Bad Directions
arXiv · 2026-03-19
Scientists built a benchmark called NavTrust to test whether navigation robots could handle real-world conditions — things like fuzzy camera feeds, bad depth sensors, and garbled instructions. They ran seven top-of-the-line AI navigation systems through it. Every single one fell apart. This is, apparently, news. The robots were great at navigating a clean, well-lit, perfectly-described world that does not exist. The actual world, with its smudged lenses and ambiguous directions, remains a hostile frontier. Four mitigation strategies were tested to help. They helped a little. The robots were then sent into the real world on a mobile robot chassis, where they performed better than before, which is the scientific way of saying "slightly less lost."
Takeaway
Seven state-of-the-art robots met reality and reality won, which is the oldest benchmark result in history.
Read paper ↗
Artificial Intelligence
Lab studyBetter AI
Your Brain and an AI's Brain Are Not, In Fact, The Same Brain
Nature Machine Intelligence · 2026-03-25
Scientists compared artificial neural networks to actual primate brains and found that most of the AI's "neurons" simply don't match up with anything the brain is doing. Two primate brains, meanwhile, lined up beautifully — like they'd been doing it for millions of years, which they have. The AI, by contrast, is doing its own thing in a large number of its units, and that thing is apparently not "thinking like a primate." Researchers now have a method to measure exactly how much of the AI is in the brain-zone versus the not-brain-zone, and the answer is: less than we thought.
Takeaway
Turns out "neural network" is more of a compliment than a description.
Read paper ↗
Artificial Intelligence
Deep Learning
SimulationBasic science
Science Learns to Unplug Things One at a Time to Figure Out What Broke
arXiv · 2026-03-23
Causal discovery — the art of figuring out what caused what — turns out to be very hard unless you're allowed to just... break stuff on purpose. Researchers studied "chain-reaction systems," where things go wrong in sequence like dominoes, and found that if you block each component one at a time and watch what stops happening, you can perfectly reconstruct the whole chain of doom. The catch: just observing the chaos without intervening doesn't work, especially when effects overlap or show up late. The method that does work needs only a handful of experiments to nail the answer. So science has officially confirmed that the best way to understand a disaster is to cause smaller, more controlled disasters first.
Takeaway
Turns out the fastest path to understanding a chain reaction is to personally detonate each link yourself.
Read paper ↗
Artificial Intelligence
Deep Learning
Simulation3D design tools
AI Can Now Texture Your Fake Living Room With a Photo of Your Actual Couch
arXiv · 2026-03-19
For years, computers could build you a beautiful 3D room — but every sofa looked like it was upholstered in whatever texture pack came free with the software. CustomTex fixes this by letting you hand the AI a reference photo of a real object and saying "make it look like *that*." It then goes through two separate rounds of caring about your couch: one round to understand what a couch is supposed to look like, and a second round to make sure it actually looks like *your* couch. The result is a virtual room where the throw pillow matches the throw pillow, the rug matches the rug, and nothing has that telltale glow of something that has never been touched by light in the real world.
Takeaway
Science has given interior designers a way to be wrong about furniture in three dimensions instead of two.
Read paper ↗
Artificial Intelligence
SimulationFaster computers
Scientists Fixed the Math on Atoms So Computers Stop Throwing a Tantrum
Nature Machine Intelligence · 2026-03-25
When AI tries to model 3D physical systems — atoms, molecules, the stuff reality is made of — the math gets expensive fast. Embarrassingly expensive. Like, "the cost doubles every time you add another atom" expensive. Researchers have now introduced something called Euclidean fast attention, which keeps the math from spiraling into a full computational meltdown. The trick is a new kind of encoding that lets the model understand where things are in 3D space without needing to compare every single atom to every other single atom at the same time. The result: the computer finishes the job instead of quietly giving up.
Takeaway
Science has officially told atoms they are no longer allowed to be computationally inconvenient.
Read paper ↗
Artificial Intelligence
Deep Learning
SimulationBetter text generation
Scientists Teach AI to Stop Repeating Itself, Name the Solution After a Keyboard Smash
arXiv · 2026-03-19
A common problem with AI text generators is that they produce the same answer over and over, like a student who learned one fact and won't stop bringing it up. Researchers have now fixed this by making the AI consider multiple candidate answers at once and then — using something called a Determinantal Point Process — pick the ones that are most different from each other. The technique is called D5P4, which is either a very sophisticated acronym or what happens when you fall asleep on a keyboard. It runs on multiple GPUs simultaneously, adds almost no extra computing cost, and lets you dial in exactly how weird and varied you want the outputs to be. In tests, the AI got more diverse. The researchers were pleased.
Takeaway
Science has formally determined that the cure for an AI that keeps saying the same thing is math with a name that looks like a Wi-Fi password.
Read paper ↗
Artificial Intelligence
Deep Learning
SimulationAI reliability
Scientists Build a Maze to Stop AI from Lying, Which Is Exactly as Cursed as It Sounds
arXiv · 2026-03-19
Researchers have discovered that AI models hallucinate and go off the rails when you push them hard enough — a finding that will shock absolutely no one who has ever asked a chatbot for directions. Their solution: a framework called Box Maze, which wraps the AI's reasoning inside three layers of bureaucratic oversight, essentially giving the model a middle manager, a compliance officer, and a wall it's not allowed to knock down. The result, in simulations at least, is that the rate of AI "boundary failures" dropped from 40% to under 1%. That's right — without the Box Maze, your AI was failing its own rules four times out of ten. It was just hoping you wouldn't notice.
Takeaway
They built a maze to keep the AI honest, which means the AI, left to its own devices, was not in a maze and was not being honest.
Read paper ↗
Artificial Intelligence
Deep Learning
SimulationDrug discovery
A.I. Sifts Through Mountains of Drug Data, Politely Discards Most of It
arXiv · 2026-03-19
Drug discovery involves feeding a computer every known fact about every known molecule and hoping something useful falls out. The problem: most of those facts are noise. Researchers have now built BVSIMC, a model that looks at all the chemical and genomic information, decides the majority of it is irrelevant, and throws it away — then, crucially, performs better because of the throwing away. It also identified which features were actually clinically meaningful, which is science's polite way of saying it did the part humans were supposed to do.
Takeaway
The most important advance in drug discovery this year is a model that learned to ignore things.
Read paper ↗
Artificial Intelligence
Deep Learning
TheorySatellite communications
Scientists Confirm Satellites Are Bad At Sharing, Propose Elaborate Workaround
arXiv · 2026-03-24
Researchers have discovered that running AI on satellites is hard because satellites are small, far away, and bad at sending data back to Earth — a problem they have solved by building a three-layer system where the ground, a low-orbit satellite, and a high-orbit satellite all take turns being in charge. The paper includes math. A lot of math. The math explains exactly when it is better to let the satellite figure things out on its own versus beaming all the information down to Earth and waiting. The answer, it turns out, depends heavily on how good your connection to the satellite is — which is to say, the system works great until it doesn't.
Takeaway
Decades of aerospace engineering have culminated in a formal proof that bad Wi-Fi is still bad Wi-Fi, even in orbit.
Read paper ↗
Artificial Intelligence
Deep Learning
SimulationWeather forecasting
Scientists Spend Years Teaching Computers to Predict the Weather One Hour From Now
arXiv · 2026-03-24
Researchers in Chongqing pitted seven machine learning models against each other in a battle to answer the most urgent question in meteorology: what will the temperature be in sixty minutes. After enormous effort — data preprocessing, lag-feature construction, rolling statistical features, a whole unified framework — a winner emerged. It was XGBoost, a method that is essentially a very large pile of decision trees wearing a lab coat. It predicted hourly air temperature with an error of 0.302 degrees. That is less than the difference between standing in the sun and standing in the shade.
Takeaway
Seven models, one unified framework, and a heroic quantity of computing power, all deployed to predict the next hour of weather with the precision of a good thermometer placed slightly to the left.
Read paper ↗
Artificial Intelligence
Deep Learning
Lab studyAI security
Scientists Discovered That AI Ears Get Confused Less When They're Slightly Drunk
arXiv · 2026-03-23
Hackers have spent years crafting sneaky audio that tricks speech recognition systems into hearing things that aren't there. The fix, it turns out, is to make the AI a little fuzzy on purpose. Researchers found that randomly wobbling the numerical precision of a speech model — essentially making it do slightly sloppier math each time — causes adversarial attacks to fall apart. The attack was carefully engineered for one version of the model. Give it a slightly different version and the whole scheme collapses. It's the audio security equivalent of hiding your valuables by occasionally moving them to a random drawer.
Takeaway
The most sophisticated AI attacks in existence can be defeated by the same strategy your IT department uses to fix printers: turning it off and back on slightly differently.
Read paper ↗
Artificial Intelligence
Deep Learning
TheoryBasic science
Mathematicians Spent 50 Years Worrying About a Cubic Surface. It's Fine.
arXiv · 2026-03-19
In 1972, a mathematician named Manin looked at a specific equation — X³+Y³+Z³+ζ₃T³=0 — and essentially said "I have no idea if this thing behaves itself." That question sat open for over half a century. Researchers have now confirmed: it behaves itself. The surface in question has trivial R-equivalence, which means the gnarly algebraic structure everyone was worried about turns out to be completely boring. This is, in mathematics, the best possible outcome.
Takeaway
Fifty-two years of anxiety, resolved by the answer mathematicians always secretly hope for: nothing weird is happening.
Read paper ↗
In Memoriam
Phylogenetic Generalised Least Squares As A Robust Causal Inference Method, 1990s – 2024
Phylogenetic Generalised Least Squares regression proposed that evolutionary associations between traits could be estimated reliably across species while accounting for shared ancestry, offering comparative biologists a principled and statistically defensible framework for their analyses. It was widely adopted across ecology and evolutionary biology, becoming a standard tool in the assessment of trait coevolution and the construction of adaptive hypotheses. For several decades it occupied a position of considerable methodological authority, appearing in thousands of comparative studies and forming the backbone of graduate training in the field. Its decline began as researchers examined the sensitivity of the method's conclusions to the assignment of variables to the dependent and independent positions — a choice that, in a genuinely robust method, ought not to determine the outcome. The terminal finding demonstrated that reversing the dependent and independent variables in a substantial proportion of published PGLS analyses yielded inconsistent or contradictory conclusions, revealing that the method had been bearing a causal interpretive weight it was not constructed to support.
It brought statistical rigour to the comparison of traits across species at a time when the alternative was largely informal, and the questions it helped researchers ask remain among the most important in evolutionary biology.
Note
The associations PGLS identified were real enough; the causal directions it appeared to endorse were a different matter entirely.