Next-Generation AI

We would like to maintain an up-to-date list of progress (papers, blogs, codes, and etc.) made in Next-Generation AI (GPT-4, ChatGPT and other AGI), and provide a guide for some of the papers that have received wide interest. Please feel free to open an issue to add papers.

Table of Contents

• For Beginners

For Beginners

In this section, we present various resources in the form of papers, blogs, and videos that are designed to aid beginners in acquiring a rapid comprehension of Transformer (include BERT and GPT Series), Instruct-GPT, ChatGPT and GPT-4. It is not advised to initially refer to the original paper as its technical content may be hard to follow for beginners.

• Groundbreaking Work (Transformer).

  • Video: A Detailed Reading of the Transformer Paper
  • Blog: Details of Transformer
  • Code: Pytorch version
  • Paper: Attention Is All You Need

• Evolution with Pretraining and Finetuning (BERT).

  • Video: A Detailed Reading of the BERT Paper
  • Blog: Illustration of BERT.
  • Code: Tensorflow version, Pytorch version
  • Paper: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

• Zero-shot Predictions with Pretraining at Scales (GPT1, GPT2, GPT3, GPT4).

  • Video: A Detailed Reading of the GPT-1,2&3 Paper
  • Blog: The Evolution of GPT Series. In addition, we offer materials concerning the LLM.
  • Code: GPT-1, GPT-2
  • Paper: GPT-1, 2, and 3 have been accompanied by academic publications, while GPT-4 is currently only documented in a technical report without any model or training details.
    • Improving Language Understanding by Generative Pre-Training (GPT-1)
    • Language Models are Unsupervised Multitask Learners (GPT-2)
    • Language Models are Few-Shot Learners (GPT-3)
    • GPT-4 Technical Report

• Learning with Human Feedbacks (Instruct-GPT, ChatGPT).

  • Video: A Detailed Reading of the Instruct-GPT Paper
  • Blog: Details of Instruct-GPT/ChatGPT. In addition, we offer materials concerning the fundamental principles of RLHF and PPO algorithms.
  • Code: ColossalAI. In addition, we can refer to the Meta version of ChatGPT, LLaMA.
  • Paper: InstructGPT: Training Language Models to Follow Instructions With Human Feedback

• Applying GPT to Various Tasks.

  • Prompt Engineering. In order to elicit specific information from GPT, it is imperative to devise an appropriately constructed question (prompt). The process of developing such prompts is commonly known as prompt engineering, and has been the subject of extensive research [Blog, Paper].
  • GPT for Embodied AI. The potential applications of GPT can be extended to embodied artificial intelligence, including Robotics [Blog, Paper], Object-Goal Navigation [Paper] and Task Planning [Blog, Paper].
  • GPT for Computer Vision. GPT has also demonstrated promise in the field of computer vision, specifically in the realm of VQA [Paper] and video [Paper].