TLDR AI 2024-06-18

Runway has trained an extremely powerful new video generation model. It will power many of the existing features on its platform. Examples are available in the provided link.

DeepMind is developing an AI technology called V2A to generate synchronized soundtracks for videos. It uses diffusion models trained on audio, dialogue transcripts, and video clips to create music, sound effects, and dialogue.

Cerebras, a California-based company, has demonstrated that its second-generation wafer-scale engine is significantly faster than the world's faster supercomputer in molecular dynamics calculations. It can also perform sparse large language model inference at one-third of the energy cost of a full model without losing any accuracy. Both achievements are possible due to the interconnects and fast memory access enabled by Cerebras' hardware. Cerebras is looking to extend the applications of its wafer-scale engine to a larger class of problems, including molecular dynamics simulations of biological processes and simulations of airflow around vehicles.

DeepSeek Coder is a powerful model that gets 90+ on HumanEval while matching GPT-4 Turbo performance on many other challenging benchmarks. It is available through an API and free for commercial use.

The new Depth Anything model was trained primarily on synthetic data and has dramatically improved performance on complex scenes.

Nvidia has released a dataset and recipe along with a high quality paper about training reward models to align model output to human preferences.

LARS is an application that enables you to run LLMs locally on your device. Upload your own documents and engage in conversations where the LLM grounds its responses with your uploaded content.

This paper investigates why diffusion-based image generation models create "hallucinations" — images that never appeared in the training data.

Everything should be differentiable. This post walks through how to write differentiable SVG lite.

Enhance your search experience with AI while maintaining flexibility and deep control over your architecture and UI. Read the blog for a tech overview, or visit the Algolia Code Exchange to access frontend and UI components, and to explore open source AI search projects built by the Algolia community. Get started

The creator of RAGatouille gave an excellent talk discussing how to dramatically improve RAG performance, some of the open challenges, and COLBERT.

What would it take to make a generally intelligent agent and what are we missing? This post explores the 3 ideas needed to make an agent and posits that we are only a few years away. The author is a researcher at OpenAI.

Chain of Preference Optimization (CPO) is a method that improves the logical reasoning abilities of large language models (LLMs). By fine-tuning LLMs using search trees from the Tree-of-Thought (ToT) method, CPO aligns the reasoning steps of Chain-of-Thought (CoT) decoding with ToT's optimal paths.

Install and run your own AI agent service.

MindMatrix is an AI-powered mind map generator that simplifies the creation and visualization of complex concepts.

TokenCost helps calculate the USD cost of using major LLM APIs by calculating the estimated cost of prompts and completions.