The Bunny That Broke the Transformer
Three researchers in Nanjing proved that Vision Transformers can't learn 3D spatial reasoning. The ceiling isn't data or training or scale. It's the architecture itself.
A research studio for frontier science
and contested ideas
Deep investigation. Adversarial analysis. Storytelling that earns your attention.
Latest
All posts »
· Tools
How to Raycastify Your Raycast
Raycast wraps your AI models in a system prompt that restricts their behavior. Here's how to override it and get the full model you're paying for.
· Physics
The End of Friction
A Tsinghua team grew single-crystal graphite films that slide with nearly zero friction at the macroscale. Two orders of magnitude past where physics said it should fail.
· Physics
Using Cosmic Rays to Catch Nukes in Space
An MIT researcher proposed using cosmic rays and a shoebox-sized satellite to verify the Outer Space Treaty's nuclear weapons ban for the first time in 58 years.
Who Decides What a Safe LLM Means
1,095 people rated AI safety. They couldn't agree. Keeping disagreement in training data beats throwing it away by 53%.
The Geometry of Refusal
An independent researcher built a framework where AI hallucinations aren't minimized but mathematically impossible. Evidence on a sphere. Contradiction collapses the region. Refusal becomes geometric necessity.
· AI
I Asked AI to Protect Me from AI
AI is changing the world faster than anything in history and it's also the same thing helping me stay myself. Here's how I built a system to protect my voice.
Deep Learning Might Have Been Lying to Us About Resolution
A new paper argues that flat error curves across resolutions aren't generalization. They're a frequency ceiling nobody was looking for.
Static Parallelism Is Killing Your GPU Budget
Current LLM training frameworks pick one parallel strategy and pray. ParaDySe hot-switches strategies layer-by-layer based on actual input length.
AI Kill Switch
Researchers in South Korea weaponized prompt injection for defense. The more capable the AI attacker, the stronger the kill switch.
Harvard Crosses Quantum Threshold
For three decades, quantum computing errors piled up faster than you could fix them. Harvard just built a system where errors get fixed faster than they're created.
· AI
Why Babies Are Still Smarter Than AI
Francois Chollet dropped a reality check. AI is 3-4 orders of magnitude less sample efficient than a baby learning to crawl.
Don't miss the next breakthrough
Get the research that matters, delivered weekly. No spam, no fluff.