Frontier models, safety, infrastructure, and where the hype actually lands.
A new wave of open weights releases from Mistral, Meta, and Google DeepMind is narrowing the performance gap with closed frontier models.
Three years after GPT-4, the research consensus is shifting: compute scaling alone may not get us to the next qualitative leap.
The company behind Claude released its most detailed account yet of interpretability research, including new findings on how concepts are encoded in transformer activations.
A European Court of Justice case collapsed after counsel admitted AI-generated citations were invented. The incident triggered an emergency regulatory review.
Compute costs for training and inference are spiraling faster than the industry publicly acknowledges. We obtained internal projections from three major AI labs.
The latest evaluation results from OpenAI's newest frontier model show performance that surpasses domain experts across medicine, law, and materials science.