Parse notes
OpenAI is launching a grant program to support cyber security professionals using AI and combating AI enabled systems. They are accepting applications and have a number of ideas to show the types of projects they're looking to support.
Chinese tech giant Baidu is rolling out a 1 billion yuan ($145 million) AI venture fund to back startups focused on content generated by artificial intelligence applications. The fund will also launch a competition for developers who build applications off Baidu's ERNIE large language model (LLM) or integrate the model into their existing products.
Google has recorded every change to its code base for decades, including rich descriptions, changes, and fixes. They treat this as a sequence modeling problem and create a set of robust internal tools that can help software engineers be more productive. They find it to be surprisingly effective, even at novel tasks like “history informed generation”.
We’ve talked about precision before, that’s how much space it takes to store a number. And how many unique values that number can represent. When you reduce the precision, AI models take less space on the computer, but weird things start to happen because you represent fewer numbers for the parameters. This paper studies what happens when you quantize before training at scale. They start to poke at the magical 6B parameter performance barrier.
Creating realistic talking digital people that sync with any speech audio is an important challenge in the digital world. GeneFace++ is a new method that improves lip-syncing, video quality, and system efficiency, making it the first of its kind to achieve real-time talking face generation.
Large AI models can be expensive and bad for the environment, so researchers have found a more efficient way to create smaller models using a method called knowledge distillation. This approach is often better than spending money on extra data and helps create compact models that still perform well in various language tasks.
This research introduces 'Occ-BEV', a new pre-training method for autonomous driving that focuses on making better use of multiple camera views, rather than just one, thus capturing a more accurate understanding of the surrounding environment. The technique uses a unique 3D decoder to predict the 3D geometry of the environment.
The new GPT4Tools system can manage multiple visual foundation models by analyzing language content, allowing it to automatically decide, control, and use different models based on the user's needs during a conversation. Uniquely, it enables users to teach their own language model how to use tools more effectively using a simple self-instruction and LoRA approach, making conversations about image-related tasks much smoother and more efficient.
A program that provides LLMs with the ability to complete complex tasks using plugins.
Sponsored
Let's discover how to use these damn probabilistic models in the best ways possible, without sacrificing the necessary software design building blocks. Join the MLOps Community and >50 speakers
The recent MosaicML MPT models are still gaining popularity. One of the big contributions was the “Storywriter” model which could use up to 65k tokens in context. Inference with this model was painfully slow. By adding a Key/Value cache we can speed up generation by almost 10x.
Hamilton Nolan argues that CEOs are becoming increasingly unnecessary as AI becomes more sophisticated. He proposes that we should automate the CEO position and distribute the power and wealth that CEOs currently control to other stakeholders, such as workers and shareholders.
TikTok accounts are posting horrifying artificial intelligence-generated clips of murder victims — mostly children — describing their own ghastly demise.
Amazon is rolling out artificial intelligence across a dozen of its largest warehouses to screen items for damage before orders are shipped to customers.
Infrastructure to build, manage, and run useful Autonomous AI Agents.