Daily Papers
1.ChatGPT's One-year Anniversary: Are Open-Source Large Language Models Catching up?(paper)
Upon its release in late 2022, ChatGPT revolutionized AI research and commerce by demonstrating a large language model's ability to answer questions and follow instructions across a broad spectrum of tasks, using a combination of supervised fine-tuning and reinforcement learning from human feedback. This sparked heightened interest in large language models, leading to rapid development in both academia and industry, including numerous start-ups. Although closed-source models like OpenAI's GPT and Anthropic's Claude typically outperform their open-source counterparts, there's been significant progress in the latter, with some claiming to match or surpass ChatGPT in certain tasks. This has profound implications for both research and business. On the first anniversary of ChatGPT, this work provides a comprehensive overview of its success, surveying all instances where an open-source large language model has reportedly achieved parity with or exceeded ChatGPT.
2.Evolutionary Machine Learning and Games(paper)
Evolutionary machine learning (EML) has been applied to games in multiple ways, and for multiple different purposes. Importantly, AI research in games is not only about playing games; it is also about generating game content, modeling players, and many other applications. Many of these applications pose interesting problems for EML. We will structure this chapter on EML for games based on whether evolution is used to augment machine learning (ML) or ML is used to augment evolution. For completeness, we also briefly discuss the usage of ML and evolution separately in games.
3.War and Peace (WarAgent): Large Language Model-based Multi-Agent Simulation of World Wars(paper|code)
This research introduces "WarAgent," an AI system using LLMs to simulate major historical conflicts. It assesses AI's capability in understanding war triggers and offers insights for conflict resolution and peacekeeping, potentially guiding future international conflict prevention.
4.Scalable Extraction of Training Data from (Production) Language Models(paper)
This study reveals the vulnerability of machine learning models, including open-source, semi-open, and closed models like ChatGPT, to extractable memorization attacks, where adversaries can efficiently extract gigabytes of training data. It introduces a new divergence attack for aligned models like ChatGPT, significantly increasing data leakage rates. The findings suggest that current alignment techniques do not fully prevent data memorization.
5.TextDiffuser-2: Unleashing the Power of Language Models for Text Rendering(paper )
The paper introduces "TextDiffuser-2," an advanced generative model designed to improve visual text generation, addressing the limitations of existing diffusion models. This model integrates a large language model for layout planning, enabling automated keyword generation for text rendering and layout modification via conversational interaction. Unlike previous models that used character-level guidance, TextDiffuser-2 employs language model encoding at the line level within the diffusion model, resulting in more diverse text images. The model's effectiveness is verified through extensive experiments and user studies, involving both human participants and GPT-4V. These evaluations demonstrate TextDiffuser-2's ability to produce more rational text layouts and generate text with enhanced diversity, surpassing previous methods in flexibility, layout prediction, and style variety.
AI News
1.SHORT COURSE: Building and Evaluating Advanced RAG Applications(link)
2.Notebooks = Chat++ and RAG = RecSys! — with Bryan Bischof of Hex Magic(link)
3.Perplexity debuts real-time LLM API: Introducing PPLX Online LLMs(blog)
4.Stability AI Positioning Itself for Acquisition(news)
5.Together AI nabs $102M to fuel open AI(news)
AI Repo
1.🐱 tryEmoji: Turn emoji into amazing artwork via AI(repo)
2.DeepSeek LLM：an advanced language model comprising 67 billion parameters(repo)
3.Shush: an app that deploys a WhisperV3 model with Flash Attention v2 on Modal (repo)