We created an 'arm farm' where robots learn to grasp objects by trial and error, improving their grasping abilities through pooled sensor data.

The release of DeepSeek's reasoning traces is significant as it allows for better training of smaller models quickly and cheaply.

We've learned a lot through the process of training this model, discovering what scales well and what doesn't.

We can achieve higher accuracy in translations by training larger models on extensive datasets, potentially exceeding human quality for some language pairs.

Google has open-sourced the Gemma model, which is designed to leverage the massive matrix-multiply unit of TPUs, resulting in a vocabulary size that is larger than necessary for its model size.

DeepSeek's model is open-weight, meaning its model weights are available for public download, allowing users to run the model independently and control their data.

The first AI red teaming competition collected 600,000 prompt injection techniques, which became a significant dataset for benchmarking AI models.

We've made substantial progress in reducing reward hacking in our models, cutting it down by about 80% in the new iterations.

GPT-4.5 is OpenAI's largest and most humanlike model to date, representing a significant step in scaling up unsupervised learning and achieving a deeper understanding of the world and human experience.