TLDR Index
TLDR AI

TLDR AI 2023-07-21

5 sections13 entriesinline urlcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01HEADLINES & LAUNCHES

AND WINDOWS

Meta and Microsoft have announced the support for Llama 2, a family of large language models, on Azure and Windows, marking its first commercial release. The partnership enables developers to build, fine-tune, and deploy Llama 2-powered models on Azure while also optimizing for local use on Windows. It aims to bolster Azure’s model catalog, democratize AI, and reinforce Microsoft’s position as a global supercomputing platform for AI, while also upholding responsible AI practices through safety measures and risk mitigation.

blogs.microsoft.com4 min read

GOOGLE REPORTEDLY PITCHING NEW AI TOOL TO NEWSROOMS

Google is testing Genesis, an AI tool designed to help journalists by generating news articles. Despite these innovations, concerns remain over misinformation and bias if such tools are not accurately supervised.

cnet.com2 min read

AI COMPANIES ARE COMMITTING TO SAFEGUARDS

Tech giants including Microsoft, Google, and OpenAI are voluntarily committing to White House-prompted safeguards for their AI technology, including independent testing and addressing societal risks. This pledge will be in effect until Congress enacts official AI regulations.

engadget.com2 min read

02RESEARCH & INNOVATION

DEEPMIND’S LATEST RESEARCH

Google DeepMind researchers are presenting more than 80 new papers at ICML this year, including demos on AlphaFold, advances in fusion science, and new models, like PaLM-E for robotics and Phenaki for generating video from text.

deepmind.com2 min read

MATRIX MULTIPLICATION ONLY USING ADDITION

Much of modern AI relies on large matrix multiplications. When implemented in hardware, this means there are expensive multiply components on the chips. If we could use cheap additions instead, we could fit more transistors on the chip and potentially run with much less power. This is very early research, but the algorithm presented suggests we can use just additions for large matrix multiplications.

arxiv.org21 min read

03ENGINEERING & RESOURCES

APPLE'S AX LEARN LIBRARY

Another deep learning library, this time by Apple and it is built on top of Jax. Notably, it supports compiler-based parallelism for large-scale training, which is a feature not present in several other popular frameworks.

github.comGitHub repository

OLLAMA

Run and package large language models on macOS.

github.comGitHub repository

RAGSTACK

Deploy a private ChatGPT alternative hosted within your VPC. Connect it to your organization's knowledge base and use it as a corporate oracle. Supports open-source LLMs like Llama 2, Falcon, and GPT4All.

github.comGitHub repository

04MISCELLANEOUS

RUST OR MOJO FOR THE FUTURE OF AI

This blog walks through the reimplementation of a thousand lines of C++ in an opencv function to a few lines of Mojo. They discuss the trade offs and speed ups. Mojo still doesn't have a local release, but it is expected to come soon.

mojodojo.dev8 min read

THE BUILDING BLOCKS OF GENERATIVE AI

A nice, high-level discussion around the key components, emerging trends, and industry players working in generative AI. There’s lots of information here around foundation models, compute, frameworks, orchestration, fine-tuning, synthetic data, and more!

shriftman.substack.com34 min read