[AINews] Ideogram 2 + Berkeley Function Calling Leaderboard V2 • ButtondownTwitterTwitter

The section provides detailed information on recent developments in the AI News domain. From new releases and benchmarks of AI models to AI applications, tools, research, and developments, various topics are covered. Discussions include the Microsoft Phi-3.5 models, Meta's UniBench, Llama 3 performance upgrades, Cyberbench cybersecurity benchmark, Codegen tool introduction, and Perplexity Browser development. AI ethics, societal impact, education, and safety debates are also explored. The AI Reddit Recap delves into topics such as optimizing LLM performance, Microsoft's Phi-3.5 model release, and creative AI applications like role-playing and character generation.

This section discusses advancements in AI models, such as the Flux model developed by Black Forest Labs, which includes techniques like Low VRAM Flux and GGUF quantization. It also mentions the release of NF4 Flux v2 and the Union controlnet for FLUX. Other AI models like Google's Imagen 3 and Meta's VFusion3D are highlighted, along with tools like SimpleTuner and AuraFlow-v0.3. The section also covers discussions on the capabilities of AI models, benchmark results in coding, AI applications in industries like filmmaking, and trends in the AI industry.

LiteLLM for LM Code Delegation

A member inquired about delegating LM code to LiteLLM and whether fine-tuning should be separated from prompt optimization. They believe prompt optimization and fine-tuning should be coupled due to their intricate interaction.

DSPy Self-Discover Framework Revealed

The DSPy Self-Discover Framework was discussed with a link to the framework's GitHub repository provided at https://github.com/jmanhype/dspy-self-discover-framework.

OpenRouter has officially deprecated function_calls and functions parameters from OpenAI calls. Some of the recent updates on OpenRouter include the release of new models such as Hermes 3 based on Llama 3.1-70b, Microsoft's release of the Phi 3.5 model family, and OpenAI allowing GPT-4o finetuning. Users have reported performance issues with Llama 3.1 70b on OpenRouter, mainly related to the DeepInfra provider. Additionally, there is a RAG cookbook available on GitHub for creating retrieval augmented generation systems. Links to mentioned content and discussions have also been shared.

Within the OpenAccess AI Collective's general discussion channel, members discuss cutting-edge models such as Phi-3.5-vision, a powerful multimodal model with a 128K context length and robust safety measures. The Phi-3 model family, including Phi-3.5-vision, explores the frontier of multimodal understanding and reasoning. Additionally, topics like fine-tuning methods for models like gpt4o and Mistral are also explored in this channel.

This section discusses various research and development topics within the Eleuther Discord channels. It covers studies on long context reasoning using different models, exploring the efficiency of transformers and generalized state space models in sequence modeling, and introducing the concept of Model MoErging where specialized models collaborate for complex tasks. Papers and surveys related to these topics are linked for further exploration and understanding.

This section discusses various topics related to LangChain AI, including medication extraction, evaluation methods, BERT integration, and accuracy assessments. Users share experiences and seek advice on optimizing and improving the extraction process for medication information through LangChain. Additionally, there is a focus on comparing LangSmith and LangChain for evaluation tasks, exploring the use of BERT in Ollama for accuracy assessments, and determining effective methods to assess the accuracy of extracted data.