The AI Canvas Newsletter #4

In collaboration with 33 academic labs, the Open X-Embodiment dataset and RT-X model have been introduced. These tools enhance the training of general-purpose robots across various environments, boosting success rates by 50%. This open-source initiative advances cross-embodiment robotics research and fosters scientific collaboration.

Read more with Google DeepMind’s blog here and paper here.

In Microsoft’s fascinating new paper, the capabilities of GPT-4V(ision), a cutting-edge multimodal model are explored. Their exploration showcases its remarkable ability to process a wide range of multimodal inputs, combining text and images seamlessly. Potential applications are discussed and hint at the future possibilities for even more powerful multimodal models, emphasizing the need for continuous self-evolution and incorporating additional modalities beyond text and images.

Read the paper here.

Delve into the world of long-context language models (LLMs) with LLAMA 2 Long, a series of models capable of handling context windows up to 32,768 tokens. Through continual pretraining and position encoding refinements, these models outperform their predecessors and even challenge gpt-3.5-turbo-16k's performance on long-context tasks, all without the need for human-annotated data. The comprehensive analysis sheds light on the key factors contributing to their success, making long-context LLMs more accessible for future advancements in the field.

Read more at _akhaliq’s Twitter post and the paper here.