Hugging Face got hacked

The video discusses the vulnerabilities in Hugging Face’s infrastructure exposed by Wiz Research and the implementation of safer storage formats to address security concerns. It also touches on the re-evaluation of GPT-4’s performance, Amazon discontinuing its “Just Walk Out” checkout concept, the rise of automated software engineering projects, and new tools and libraries in the AI and machine learning space.

In the video, the speaker discusses a blog post by Wiz Research that exposed vulnerabilities in Hugging Face’s infrastructure, particularly related to the use of the pickle method for storing models. This method allowed for arbitrary code execution, leading to potential security breaches. Hugging Face has since implemented safer storage formats like safe tensors and model scanning to address these security concerns.

There is also a discussion about the re-evaluation of GPT-4’s performance on bar exams, with a new paper showing varying results compared to previous claims. Despite the discrepancies, the speaker acknowledges the impressive achievement of AI models like GPT-4 approaching human performance levels in professional domains.

The video also touches on Amazon discontinuing its “Just Walk Out” checkout concept in grocery stores, which relied on surveillance and machine learning for automated checkouts. It is revealed that over a thousand people in India were involved in monitoring and labeling videos for accurate checkouts, highlighting the human involvement behind seemingly automated processes.

The speaker then explores the rise of automated software engineering projects, such as Deva, SWE Agent, and GPT Pilot, which aim to automate coding tasks and bug fixing. These projects showcase advancements in AI capabilities to assist developers in various software engineering tasks, potentially impacting the future of software development processes.

Lastly, the video highlights new tools and libraries in the AI and machine learning space, including Lightning Thunder for faster PyTorch model compilation, GPT author for generating novels, Maestro for orchestrating AI agents, and Tactics 2D for reinforcement learning environments. These tools demonstrate the ongoing advancements in AI technology and its applications across different domains.