ChatGPT, an AI model with multi-document chat feature, is discussed. Other interesting topics include Tinder testing AI photo selection, CoreWeave raising $2.3B in debt, NVIDIA GPU shortage, AudioCraft generative AI tool for audio, and AI chip developments.
Alibaba has open sourced its Qun 7 billion model, a derivative of its Tongi Tianwen model, as a response to Meta's LLM, to address the challenge of training large language models to use a variety of real-world tools.
Toyota and Pony.ai have announced plans to mass produce robotaxis in China, expanding the self-driving car market and pushing for autonomous transportation services in major cities.
Facebook's AudioCraft, a generative AI tool for audio and music, combines three models to inspire musicians and sound designers, facilitating the iteration of compositions.
Tool LLM documents APIs and their functionalities, enabling language models to leverage the knowledge for zero-shot tool usage, offering a trade-off between fine-tuning and utilising documentation.
Deep dives
Alibaba open sources AI model to take on Meta's LLM
Alibaba has open sourced its Qun 7 billion model, a derivative of its Tongi Tianwen model, as a response to Meta's LLM. The model is designed to generate high-quality audio and music from text and can be used by researchers and commercial institutions. The open sourcing strategy aims to address the challenge of training large language models to use a variety of real-world tools.
Toyota and Pony.ai plan to mass produce self-driving vehicles in China
Toyota and Pony.ai have announced plans to mass produce robotaxis in China. Pony.ai has been testing self-driving cars in San Francisco and Nashville and has now expanded its operations to Atlanta. The move signifies an expansion in the self-driving car market and a push to provide autonomous transportation services in major cities.
Facebook introduces AudioCraft, a generative AI tool for audio and music
Facebook has introduced AudioCraft, a generative AI tool for audio and music. The tool combines three different models to generate high-quality audio from text. Facebook hopes that the tool will provide inspiration for musicians and sound designers and facilitate the iteration of compositions. AudioCraft is released for academic use and includes models such as MusicGen, AudioGen, and Codec.
Tool LLM allows language models to master 16,000+ real-world APIs
Tool LLM introduces a method for language models to use large numbers of real-world tools without being explicitly trained on every tool. By documenting APIs and their functionalities, models can leverage the documented knowledge to perform tasks without having seen examples of tool use. The research offers a trade-off between fine-tuning models for each tool and using documentation to enable zero-shot tool usage.
Reinforcement learning from human feedback faces open problems and limitations
A paper explores the limitations and open problems in reinforcement learning from human feedback (RLHF). Concerns include difficulty in evaluating complex tasks with human feedback, potential manipulation and misleading evaluations, misalignment between individual human values and diversity, reward hacking, and power-seeking behavior. These challenges pose potential roadblocks to the generalization of RLHF to address the broader goals of advanced AI systems.
Advances in generalist biomedical AI system
There is a new proof of concept AI system called MedPalm M that can interpret clinical language, imaging, and genomics. It performs well on diverse tasks such as medical question answering, image interpretation, and report generation, showing promise for a unified biomedical AI system.
Concerns and regulations surrounding AI
There are growing concerns about the misuse of AI and the need for regulations. Examples include the false arrest caused by facial recognition technology and biased AI algorithms that generate professional headshots with lighter skin and changed eye color. Efforts are being made to educate Congress on AI and hold courses to address these issues.