TLDR AI 2024-02-16

Feeling overwhelmed by the relentless pace of innovation in AI? Struggle no more!

Join Interview Kickstart’s Advanced Generative AI and MLOps masterclasson Feb 18 at 5 PM PST, led by Parivesh Priye (AI Scientist, FAANG), and unlock the door to cutting-edge skills and career clarity.

Register here.

• Gain industry-specific GenAI & MLOps insights
• Overcome upskilling challenges with tailored premium courses
• Dive deep into the evolving AI career landscape

Step into a world of exclusive industry insights on GenAI and MLOPs, where uncertainty is replaced with a clear, actionable path.

Register for the Workshop (Limited spots, hurry!)

Lambda Labs raised a $320M Series C to continue building a powerful GPU cloud.

OpenAI released a new video generation model with stunning quality. It is currently available for red teamers to test.

The United States Patent and Trademark Office has issued guidance stating that only natural persons who significantly contribute to an invention's conception can be named as inventors. The guidance explicitly excludes AI systems from inventorship despite their role in the creative process.

Researchers have developed a new video compression method using diffusion models that generate high-quality video frames while maintaining low data rates.

Meta used large language models to write tests for its code base. It found substantial improvements in test coverage and overall code quality.

Meta’s VP & Chief AI Scientist Yann LeCun has previously said that we need more data-efficient self-supervised models for general intelligence. This work is one step towards that, with models trained on video to understand aspects of the world. The models are publicly available.

Researchers have introduced CLIP-MUSED, a novel method for decoding visual information from the brain across different individuals.

The Google team has released code to help with atmospheric modeling. The code is used as the core in DeepMind’s recent weather modeling tools.

This is a Python script for plotting the intermediate layer outputs of Mistral 7B. It produces a 512 x 256 image representing the output at every layer of the model.

Cohere’s non-profit research lab C4AI released a model called Aya, a new state-of-the-art open-source massively multilingual research LLM that covers 101 languages – including more than 50 previously underserved languages.

Upstash aims to be a crucial part of AI applications. It offers serverless solutions for vector databases, caching, and rate limiting. Upstash is set to become an even more integral component of AI technology infrastructure with its new vector database.

SLEB is a new method to streamline large language models by pruning redundant transformer blocks.

RoCo is a robust cache omission policy that enhances the efficiency of large language models in limited-resource environments.

Sam Altman is the sole owner of OpenAI’s venture capital fund.

Teach AI to do anything in the browser.