Emad Mostaque, a prominent and divisive figure, has resigned as CEO of Stability AI to pursue decentralized AI. Stability AI's COO and CTO are now Co-CEOs as the company searches for a permanent figure.
GitHub introduced a beta for code-scanning autofix, merging GitHub’s Copilot with CodeQL to detect and automatically remedy security vulnerabilities in JavaScript, TypeScript, Java, and Python.
This post contains a deep dive into how to use ChatGPT (or any language model) to write code in an unfamiliar language for a challenging task. Its author uses ChatGPT’s code interpreter to write, compile, and download new bindings for the popular database SQLite.
Open-Vocabulary Attention Maps (OVAM) modify text-to-image diffusion models like Stable Diffusion by allowing the generation of attention maps for any word, bypassing previous limitations.
HETAL introduces a new approach to transfer learning that ensures data privacy through Homomorphic Encryption, marking a significant step forward in secure AI training.
Optimal Transport theory is a rich area of mathematics and computation that deals with distances between probability distributions. This repository contains an extremely comprehensive reading list for scaling Optimal Transport in both speed and accuracy.
Mistral recently held a hackathon. As part of it, the company announced v0.2 of its 7B model and released code on how to fine-tune its language models. The code is clean and concise.
AdaIR introduces an all-in-one image restoration network that leverages frequency mining and modulation to tackle various forms of image degradation such as noise, blur, and haze.
The recent acquisition of Inflection's talent by Microsoft reignited debates around the biggest tech companies gaining too much control over AI development. While their resources allow rapid progress on core language models, there are valid worries about the concentration of power limiting innovation and transparency. This article explores the complex tradeoffs as the commercialization of AI moves forward.
The Databricks Mosaic team has continued to innovate on their language model training. In this post, they discuss the training stack used for fp8 and some of the benefits that come from reducing precision.
This project introduces the Hash-grid Assisted Context (HAC) framework, which achieves over 75X compression of 3D Gaussian Splatting (3DGS) data, outperforming current standards.
A new method called "Quiet-STaR" enhances AI reasoning by giving systems an "inner monologue," allowing for improved performance on reasoning tests, as shown by its application to the Mistral 7B language model.