TLDR AI 2023-11-13

Google is in talks to deepen its relationships with Character.AI by investing hundreds of millions of dollars.

Zapier introduced AI Actions, a tool for developers to let any AI platform run Zapier's 20,000+ automation actions. AI Actions works by letting users send natural language commands to the AI platform, which then performs the desired action. The service supports several AI platforms, with simple setup and inherent API integrations.

Figma has incorporated AI assistance into FigJam, its digital whiteboard tool, to simplify and enhance design collaborations. Utility-oriented enhancements, like those derived from the AI-powered project Jambot, help users collaborate more effectively on a virtual canvas. Figma’s goal is to broaden the applicability across various user requirements by tapping machine learning capabilities for visual design.

This project introduces the Dual-Guided Spatial-Channel-Temporal (DG-SCT) attention mechanism that enhances pre-trained audio-visual models for multi-modal tasks.

This blog post is about different attacks that are emerging against language model systems. It contains excellent information about different types of attacks and some mitigations that teams have found to be effective.

This research presents the 3DStyle-Diffusion model, a novel approach for detailed stylization of 3D meshes, integrating 2D Diffusion models for added control over appearance and geometry. It works by first parameterizing a 3D mesh's texture into reflectance and lighting, using implicit MLP networks, and then using a pre-trained 2D Diffusion model to align the rendered images with the text prompt and ensure geometric consistency.

Rigorous unit testing for ML models increases trust in your AI/ML systems, and enables deeper insights into model performance in a fraction of the time.

Discover how it’s done in this upcoming webinar by Kolena. You’ll learn how to design a model quality management process; standardize test coverage; and deal with robustness, fairness, and biases. RSVP to watch live or get the recording after the session

Audio super-resolution is the process of increasing the quality and fidelity of any audio, real or synthetic. Most super-resolution systems are task-specific, with single models trained for single audio data types (e.g., speech vs music). This new work is an amazing step forward where a single model can serve to increase the quality of audio across tasks.

With the advent of powerful new vision models, many groups are attempting to build agents that use vision to interact with web elements. Tarsier toolkit introduces a standard set of tools (e.g., element tagging). You can use any vision system to understand the web page and take action. It also includes utilities for non-vision language models to browse.

With the recent release of the excellent Zephyr language model, the HuggingFace team showcases how you can train personalized models built on top of the few powerful pre-trained open source models available.

The LLaVa model is an open-source model that combines both language and vision. This new version allows the instruction-tuned model to use tools for image editing, generation, and more.

An example of using Bark text-to-speech system to generate coherent, fast, and extra-long audio output.

AI town is an amazing experiment where hundreds of agents live out their daily lives as a prompt state in a language model. The RWKV model is a linear language model that requires fewer resources than standard Transformers. This repo uses this cheaper model to run AI town on your local machine.

Personalized AI content for brands.

Poe has updated its AI platform to support creator monetization through subscription revenue shares and per-message fees, aiming to empower bot developers and small AI companies. Monetization is live in the US, with global expansion on the horizon.