TLDR AI 2023-09-20

Before going all-in on AI, you need data that’s AI-ready. The semantic layer has a big role to play here — providing a single view of critical business metrics (e.g. ARR, headcount) and common analytics vocabulary across various groups.

The new 2023 GigaOm Sonar Report for Semantic Layers and Metrics Stores compares the leading vendors in this space including AtScale, Microsoft, Google, and dbt. The report names AtScale as the leader in semantic layers, owing to consistently exceptional performance across several categories — user interface, data source and client connectivity, security, access controls, and native API support.

Download the full GigaOm report (free) ↗️

OpenAI has introduced the OpenAI Red Teaming Network, a group of experts that will assist in assessing and mitigating the risks associated with its AI models.

Google's Bard chatbot can now scan Gmail, Docs, and Drive to retrieve information and perform tasks. Google says these new integrations, called extensions, aim to save users time when searching for specific information within their documents and emails.

Anyscale has unveiled Endpoints, which allows developers to integrate fast, cost-efficient, and scalable LLMs into their applications using popular LLM APIs. The company also announced a new integration with NVIDIA to boost the performance and development of LLMs on Ray, Anyscale’s open-source framework for distributed machine learning.

Turning long texts into images can be tough, especially when there are lots of items or details to show. This new method breaks down long descriptions into simpler steps and makes sure each detail fits perfectly in the picture, setting a new high standard for creating images from text.

Figuring out 3D depth from a single 2D image is tricky because many 3D scenes can look the same in 2D. This method separates camera details from the picture's content, making it easier to guess the depth in various situations.

Hot on the heels of Google's new dataset comes CulturaX. A cleaned dataset of trillions of multilingual tokens will be a great boost for multilingual language models.

Researchers have introduced a new method called DoLa that helps large language models (like the ones that chat with you or write articles) stop making stuff up and stick to the facts. It uses a smart way of comparing different layers inside the model to filter out false information.

PIPS2 is a state-of-the-art point-tracking system that performs dense pixel-level tracking and is robust to basic occlusions and fast motion.

Creating eye-catching designs online is important for getting users interested and helping them understand information. This new method, called LayoutNUWA, makes these designs smarter by using language models to turn coding instructions into awesome layouts.

The world's most powerful speech-to-text (STT) model is here: Nova-2. Compared to alternatives, Deepgram’s Nova-2 offers next-level accuracy and blazing fast speed, to ship voice AI features faster. Learn more or sign up to start building with $200 in free credit.

Google’s AlphaFold played a crucial role in determining the specific structures of two intrinsically disordered proteins that induce previously unidentified structures when they interact. The discovery sheds light on how such proteins function and brings valuable insights into complex biological questions that were previously challenging to answer.

Matting is the process of creating a black-and-white mask to segment out important parts of an image. It is usually used for foreground and background segmentation. The portrait mode on smartphones uses a method similar to ViTMatte described here.

ChatGPT’s training cut-off date has been upgraded from September 2021 to January 2022.

Your personalized AI for meeting notes and scheduling.

Generate any chart you want using a single sentence.