In today's development of artificial intelligence (AI), the speed of data transmission has become a significant bottleneck hindering its progress. To break through this barrier, a research team led by the University of Michigan (U-M) is developing a brand new chip connection system that uses optical waves instead of traditional cables for data transmission. This innovation is expected to address the 'memory wall' problem that limits computational speed and promote further growth of AI models. The project has received a $2 million grant from the National Science Foundation's Future Semiconductor program, with participation from the University of Washington and Penn State.

Traditional video understanding models face numerous challenges when processing long videos, including the complexities of understanding the extended context. Although considerable research has been conducted to enhance video understanding capabilities, effectively overcoming the issues of low training and inference efficiency remains difficult. To address these challenges, the research team utilized HiCo technology to compress redundant parts of video information, significantly reducing computational demands while retaining key information. Specifically, HiCo achieves hierarchical compression of the video by segmenting long videos into shorter clips, thereby reducing processing time.

At the recent CES conference, NVIDIA CEO Jensen Huang stated that the performance improvement speed of the company's AI chips has surpassed the historical standards set by Moore's Law. Moore's Law, proposed by Intel co-founder Gordon Moore in 1965, predicted that the number of transistors on a computer chip would approximately double every year, consequently doubling chip performance as well. However, in recent years, the pace of progress of Moore's Law has significantly slowed. Image source note: The image was generated by AI, image licensed by Midjourney.

Recently, xAI announced a major upgrade to its Grok AI chatbot, promising significant enhancements in performance, even claiming it can compete with Sonnet3.5, particularly in prompt-following capabilities. The upgraded Grok-2 is now available for free to all X platform users, with paid users benefiting from higher usage limits and additional features. According to xAI, Grok-2 is three times faster than its predecessor and has also seen notable improvements in accuracy and language support.

Dr. Alan Becker, a deep learning scientist who previously led the development of Snapchat's 'My AI' chatbot, has recently come back to launch a new company called eSelf. This startup is focused on creating a real-time, video-based conversational AI assistant platform and has successfully secured $4.5 million in seed funding, officially entering the fierce competition in the field of AI assistants. eSelf's early goal is to provide a more realistic and interactive AI experience for verticals such as education, sales, financial services, real estate, and healthcare.

Generating high-quality, temporally continuous videos requires a significant amount of computational resources, especially for longer time spans. The latest Diffusion Transformer model (DiTs) has made significant progress in video generation, but its reliance on larger models and more complex attention mechanisms has resulted in slower inference speeds, exacerbating the challenge. To address this issue, researchers at Meta AI have proposed a training-free method called AdaCache to accelerate video DiTs.

Recently, OpenAI launched an important update that introduces the 'Predicted Outputs' feature for the GPT-4o model. This innovative technology significantly enhances the model's response speed, achieving up to 5 times the original speed in specific scenarios, providing developers with a completely new efficiency experience. This feature, developed in collaboration with FactoryAI, has a core advantage of bypassing the process of repeated generation of known content. In practical applications, especially in updating blog posts and iterating on existing content.

Large Language Models (LLMs) like GPT and Llama have sparked a revolution in the field of artificial intelligence, but efficiently training these massive models while aligning them with human values remains a challenge. Reinforcement Learning with Human Feedback (RLHF) has become an important training method for LLMs in recent years, but traditional RLHF frameworks have limitations in flexibility, efficiency, and scalability. To address these issues, the ByteDance Doubao Large Model Team has open-sourced an RLHF framework called HybridFlow.