Ilya and OpenAI Parting Ways, Inflection AI's Leadership Shakeup, and More!

Ilya and OpenAI Parting Ways

Ilya is leaving OpenAI to work on a personal project. He's been a key figure in the organization's success. OpenAI CEO Sam Altman praises Ilya's contributions and intellect, describing him as a guiding light in the field. Jakub will take over as Chief Scientist, with Altman expressing confidence in his abilities to lead important projects and advance their mission of beneficial AGI. (Read More)

Inflection AI reveals new exec team and plan

Inflection AI, facing challenges after Mustafa Suleyman's departure to Microsoft, has announced a new leadership team focused on embedding emotional intelligence (EQ) in business chatbots. The team, led by CEO Sean White, aims to develop empathetic AI tools capable of understanding and adapting to human emotions, distinguishing Inflection AI in a market dominated by giants like OpenAI and Google. They plan to leverage their expertise and unique training data on emotional conversations to build personalized and emotionally responsive chatbots for businesses, with a well-funded runway for the next 18 months. (Read More)

Stability AI discusses sale amid cash crunch

British AI startup Stability AI, known for generating audio and video from text prompts, is reportedly considering a sale due to financial struggles. The company, which lost its CEO Emad Mostaque and laid off staff recently, generated under $5 million in revenue but lost over $30 million in the first quarter of 2024. It currently owes about $100 million in bills to cloud providers and others. (Read More)

Bend: A massively parallel, high-level programming language

Bend is a high-level, massively parallel programming language developed by HigherOrderCO. It features the expressiveness of Python and Haskell and supports higher-order functions and recursion, optimized for GPUs using CUDA. Bend automates parallelism without explicit annotations. Installation involves setting up dependencies like Rust and HVM2. It supports various execution modes: Rust for sequential, C for parallel, and CUDA for massively parallel execution. The language excels in parallel performance but is currently limited to NVIDIA GPUs and Linux or macOS (via WSL2 for Windows). Example code demonstrates simple and complex parallelizable tasks. (Read More)

llama3-from-scratch: llama3 implementation one matrix multiplication at a time

This repository by user "naklecha" provides a detailed step-by-step guide to implementing the LLaMA3 model from scratch. It focuses on matrix multiplication and tensor operations, covering the entire process from initializing token embeddings to constructing and normalizing attention layers and feed-forward networks. The project includes code snippets for loading model weights, applying rotary positional encodings, and handling multi-head attention mechanisms. The ultimate goal is to generate embeddings suitable for predicting the next token in a text sequence. The document concludes with validation of the implementation by predicting "42" as the answer to a classic text prompt. (Read More)