• Even the most open-source model, Meta’s Llama 2, scores only 54% on the new Foundation Model Transparency Index – the index measures whether we know what was used to build a model, how it’s designed, how it works, and what applications it’s used for. - Stanford
• 78% of surveyed mayors around the world are interested in using genAI to improve local government efficiency, but are blocked by lack of tech expertise, lack of awareness, and budget limitations – City AI Connect is a global initiative aiming to boost these efforts. - City AI Connect
• “Regulating research and development in AI is incredibly counterproductive. They want regulatory capture under the guise of AI safety,” said Yann LeCun, adding that current AI isn’t even smarter than a cat, so we shouldn’t worry about AI destroying all humans yet. - FT
• LLMs can infer personal attributes (location, income, sex) just from a Reddit profile, and they can easily extract more private information by asking basic questions, which brings to light a new type of privacy risk related to language models. - arXiv
• AI teaching a simulated robot to spin a pen – Eureka uses LLMs to generate highly performant reward functions for reinforcement learning, enabling a simulated Shadow Hand to perform pen spinning tricks for the first time ever. - GitHub
