TLDR AI 2024-05-07

Stack Overflow, a popular programming website, and OpenAI are partnering to provide a data API for OpenAI customers to retrieve real time and vetted data.

Elon Musk plans to enhance X's AI, Grok, to merge live news with social media commentary to provide updates and citations in real time. Grok will generate news summaries from user discussions on X, focusing on engagement and accuracy. The project faces challenges with proper citation and legal concerns.

Video from the founder of Unsloth on how its team uses PyTorch, writes their kernels, and designs their API surface. Unsloth's framework and library are extremely powerful and easy to use.

SlotGAT is a new approach that improves heterogeneous graph neural networks by addressing the semantic mixing issue in traditional message passing.

This new method detects deepfakes by focusing on masked image modeling, especially in the frequency domain. The approach differs from traditional methods and shows significant improvement in identifying synthetic images, even from new AI generative techniques.

The team at Hugging Face has released a new project that nicely packages common tools needed for robotics development.

Researchers have developed "Morph-Tokens" to improve AI's visual understanding and image generation capabilities. These tokens transform abstract concepts used for comprehension into detailed visuals for image creation, leveraging the advanced processing power of the MLLM framework.

Vibe-Eval is a newly launched benchmark designed to test multimodal chat models with 269 visual understanding prompts, including 100 particularly challenging ones.

Automating prompt optimization for AI models suggests a future where manual prompt engineering may become obsolete, pointing towards more efficient, model-driven methods of generating effective prompts.

Vision-Language Models like GPT-4V are advancing rapidly in understanding and interacting with images and text. A recent study uncovers their significant limitations in visual deductive reasoning. Researchers tested these models using complex visual puzzles, like those found in IQ tests, and discovered that they struggle with multi-step reasoning and recognizing abstract patterns.

DeepSeek has released a 200B+ parameter model with 21B active parameters. It performs extremely well on code and reasoning. It's not clear if it is overall better than Llama 3 70B, but it is a welcome addition to the open model ecosystem.

A new report by Stanford tracking worldwide AI trends.

Command R and R+ are state-of-the-art LLMs optimized for long context tasks and excel at key business-critical capabilities.

Microsoft's Responsible AI Transparency Report highlights its advancement in deploying AI responsibly in 2023, including the creation of 30 AI tools and implementing safety measures.