TLDR AI 2024-06-12

Builders and GenAI developers, this one's for you!

DataStax is hosting a high-energy hack party on June 24 in partnership with Unstructured, LangChain, Microsoft, NVIDIA, TechCrunch and others. Here's what you can expect:

🍺 Food and drinks

🤝 Networking with the SF AI community

👩‍💻 Hands-on hacking

⚡ Tips on how to boost your development speed by 100x.

Team up with DataStax, Unstructured, Microsoft, LangChain, NVIDIA, and Voyage AI to connect with RAG professionals, see demos of AI in production, and enjoy dinner and drinks while networking!

This event is in-person and space is limited. RSVP today to be considered

Elon Musk has dropped the lawsuit against OpenAI and Sam Altman that claims they have deviated from their original mission in favor of profit. He has the option to file again in the future.

Mistral, a Paris-based AI startup, turns 1 year old today. It announced a monster series B funding raise.

Together has introduced Mixture of Agents (MoA), a novel method that combines multiple LLMs for superior performance that achieves a 65.1% score on AlpacaEval 2.0, surpassing GPT-4o. MoA uses a layered structure where initial responses from various models are refined by aggregators in successive layers, enhancing output quality through collaboration. Despite higher accuracy, MoA faces latency challenges. Future directions include optimizing model architecture and reducing latency.

Most compute for Transformers is spent in the Linear layers. This work builds on muP and Monarch matrices to create a structured representation that uses less compute and has better scaling laws than naive dense layers.

PHATGOOSE is a method that significantly improves an AI's ability to generalize and learn new tasks without prior exposure by smartly routing between different specialized language models for each part of a task.

CARES is a comprehensive framework that evaluates the trustworthiness of Medical Large Vision Language Models (Med-LVLMs).

The Mistral.rs Rust-based inference framework has improvements to its speed, GGUF support, and more in the latest release.

LE3D is a method for real-time HDR view synthesis from RAW images. It is particularly effective for nighttime scenes.

The Human Gaussian Model (HGM) is a diffusion-guided framework for creating 3D human models from a single image.

In this webinar: AI thought leaders Steve Nouri and Arun Nandi, and AtScale CTO Dave Mariani, will discuss ways to use generative AI to increase data and analytics accessibility for all enterprise users. Key topics will include NLQ, co-pilot features, and model suggestions. Join the conversation on June 26

Researchers have developed a novel framework that improves the grading of Diabetic Retinopathy (DR), a condition that can lead to visual impairment.

The inaugural Mozilla Builders Accelerator is embracing Local AI, which brings AI models and applications directly onto personal devices to enhance privacy and accessibility. Key innovation areas include tools for developer productivity, local-based AI agents, dynamic UIs, adaptation of fine-tuning, retrieval-augmented generation, and advanced function calling. The initiative seeks participants to develop a decentralized, open-source AI ecosystem that prioritizes user empowerment.

How overfit are popular LLMs on public benchmarks? New research from Scale AI SEAL revealed that Mistral & Phi are overfitting benchmarks, while GPT, Claude, Gemini, and Llama are not. The researchers produced a new eval GSM1k and evaluated public LLMs for overfitting on GSM8k.

CaseMark, an AI startup revolutionizing legal workflows, has secured a pre-seed investment led by Gradient Ventures to expand its AI offerings for the legal industry.

OpenAI plans annual opportunities for shareholders to sell equity, but there are concerns over the company's restrictive approach.

Haize Labs announced on Twitter a suite of “haizing” (i.e. red-teaming and fuzzing) algorithms against industry-leading, safety-trained AI companies.

Modern authentication for AI applications. Bot and abuse protection included in every plan. Industry-forward pricing, including only charging for users after their first day of activity. Learn More