Large language models are just the beginning. The real revolution comes when AI gains the ability to think structurally and self-correct like humans do.
New research from Anthropic explores whether AI models can introspect—that is, accurately report on their own internal states—and the findings challenge common intuitions about what language models can do.
Kolmogorov-Arnold Attention, or KArAt, is a new type of attention mechanism for Vision Transformers using learnable activation functions, outperforming traditional softmax attention in some cases.
A team of AI researchers gave six language models a task to design a human study, and the result was an unexpected survey question about the last digit of your birth year.
When AI models read information that contradicts their own knowledge, they often ignore it. This counter-intuitive behavior reveals a fundamental aspect of how these systems work.
A compact 1.5-billion-parameter model outperforms many larger models in competitive math and coding tasks, proving that size isn't everything when it comes to AI performance.