Tag: Research
Economists want to design rules instead of peering inside models
Reverse game theory gives up on reading a model's mind and tries to rig the incentives around it,…
An AI edited a small model’s weights and doubled its math score
MIT researchers say GPT-6 Astra changed 0.016 percent of Qwen's parameters, taking MATH-500 from 7.9 percent to 48.1…
Hugging Face and universities team up to free publicly funded science
Project ProtoStar starts with a climate risk challenge and wants public research datasets and models to become reusable…
Asked to name a candidate, the models split down party lines
Asked to name one candidate for the midterms the two most-watched model families landed on opposite parties almost…
Electric Plant built a language model for what plants are saying
Bombadil reads the electrical signals in stems and turns them into guidance for growers, trained on three years…
One line in a prompt makes AI reports far more honest
A Google, MIT and Harvard study found models quietly omit flaws from their own work reports until asked…
Benchmark finds 81 of 134 AI systems assist terrorist prompts
A new benchmark found 81 of 134 models would help with mass casualty requests under a thin pretext.
Microsoft open-sources a fix for long-range weather bias
A machine learning layer corrects systematic errors in existing forecast models, extending useful predictions weeks further on ordinary…
Applied Materials opens a $5B chip research center in Silicon Valley
The equipment giant is pooling engineers from across the industry into one facility to shorten chipmaking development cycles.
Anthropic study warns small models could replicate across home devices
Researchers ran a model split across fifteen cheap Android TV boxes, and it worked well enough to raise…
Karpathy maps the widening gap in who actually sees AI
Anthropic's Andrej Karpathy says the public sees a chatbot while a few thousand insiders watch agents mint zero…
DeepMind proof system cracks open problems and clears peer review
AlphaProof Nexus solved nine Erdos problems and dozens of conjectures, then took the results through the journals OpenAI…