News 4 June 2023 - simon-oz/Weekly-AI-news GitHub Wiki

  1. Google, Princeton, and Stanford published a paper “Large Language Models as Tool Makers”. The paper proposed a closed-loop framework referred to as LATMs, which can create their own reusable tools for problem-solving. The project uses GPT-3.5 as tool user, and GPT-4 as tool maker to reduce inference costs.
  2. Nvidia announced DGX GH200, a supercomputer that is 10 times faster than the current fastest computer in the world. The computer will be used for generative AI language applications. Watch this from 60mins for about 1 min. Google, Microsoft, and Meta will be its first users.
  3. Nvidia’s Neuralangelo project can now turn 2D video clips into 3D structures and scenes. It literally generates detailed replicas of buildings, sculptures, and real objects from video clips taken on a mobile or camera.
  4. Statement on AI Risk has been circulated and signed by a lot. “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war”
  5. OpenAI publishes a research paper, “let’s verify step by step”, by using this process supervision approach, the process-supervised model solves 78% of the problem from the MATH test set. It also aims at attack the Hallucination issues of LLMs.
  6. GPT4Tools – an open-source tool based on Vicuan (LLaMA), and aims to efficiently enable LLMs to decide, control and utilizing different visual foundation models, allowing users to interact with images during a conversation.
  7. Google, OpenAI, Anthropic, etc published a paper “Model evaluation for extreme risks”. An evaluation model is created to evaluate extreme risks by looking at dangerous capabilities and alignment as input and to ensure responsible training, responsible deployment, transparency and appropriate security.