• Meta is pushing translation AI forward with Seamless – based on the SeamlessM4T v2 model, preserving expression in speech with real-time translation while the speaker is still talking, Seamless provides a state of the art translation experience for several languages. - Meta
• The leader in LLM search, Perplexity introduced two online models designed to provide fresh, factual, and helpful responses – not as good as GPT-4, but better than gpt-3.5 and llama2-70b, especially when it comes to up-to-date and factually correct information. - Perplexity
• Google DeepMind's GNoME (Graphical Networks for Material Exploration) uses two deep learning models to predict structures for 2.2 million new materials, with over 700 created via robotic automation in the lab – this could lead to important breakthroughs in many sectors. - MIT Technology Review
• PANDA accurately finds and identifies pancreatic cancer, which is often found too late in patients – it analyzes simple, non-contrast CT scans, and is better than the average radiologist, with 92.9% accuracy for cancer cases and 99.9% for non-cancer cases. - Nature
• Mozilla released llamafile, a breakthrough tool that simplifies running models locally – it’s a single file that weighs a few GB, containing both the model weights and the necessary code to run an LLM, it works offline across various operating systems and hardware architectures, and runs with decent performance on a standard modern laptop. - Simon Willison’s Weblog
• The European AI Act could be delayed or shelved due to disagreements – France, Germany, and Italy are pushing to let developers of foundation models self-regulate and other countries oppose this idea, they also can't agree on AI definitions or law enforcement exceptions. - Reuters
