AI News

August 3, 2024

Language Models

  1. Gemini 1.5 Pro Performance Google's DeepMind's Gemini 1.5 Pro surpasses GPT-4o and Claude-3.5 in Chatbot Arena with superior performance in multilingual tasks and technical areas like math and coding. Read more

  2. BitNet b1.58: Efficient LLMs BitNet b1.58, a 1-bit LLM, demonstrates efficient large model operation on devices with limited memory, such as phones. Read more

  3. Llama 3.1 405B Release Llama 3.1 405B, the largest open-source model, now available for free testing, is competitive with closed models and allows enhancements for other models. Read more

RAG Systems

  1. Open Source RAG Implementations A comprehensive open-source repository features various Retrieval-Augmented Generation (RAG) strategies, including GraphRAG, offering tutorials and visualizations. Read more

Fine-Tuning

  1. Distributed Shampoo Optimization Distributed Shampoo outperforms Nesterov Adam in deep learning optimization, marking a significant advancement in non-diagonal preconditioning. Read more

  2. LoRA Training Techniques Discussions on saving and loading models trained with Low-Rank Adaptation (LoRA) in different formats to maintain model accuracy highlight the importance of merging and correct quantization methods. Read more

Security and Ethics

  1. Acquihires in AI Industry Character.ai's acquisition by Google and other similar moves spark debates on the implications for competition, innovation, and talent retention in the AI sector. Read more

Others

  1. FLUX.1 Text-to-Image Model Black Forest Labs' FLUX.1, a state-of-the-art text-to-image model, is available in three variants, with the schnell version under an Apache 2.0 license. Read more

  2. SQLite-vec Vector Search Extension SQLite-vec v0.1.0, a new vector search extension for SQLite, offers vector similarity search capabilities, supporting various platforms including WebAssembly. Read more

  3. GitHub Model Hosting GitHub now hosts AI models directly, providing a frictionless path to experiment with model inference code using Codespaces. Read more