Stable Doodle by StabilityAI, Claude-2 with 100K-token limit, Open dataset of 10 million+ 3D objects, Text to geo maps, Shopify's AI assistant for merchants, China's new open-source LLM and more
Greetings and welcome to this week's AI Brews for a concise roundup of the week's major developments in AI.
In today’s issue (Issue #23):
AI Pulse: Weekly News & Insights at a Glance
AI Toolbox: Product Picks of the Week
AI Skillset: Learn & Build
🗞️🗞️ AI Pulse: Weekly News & Insights at a Glance
🔥 News
Stability AI launches Stable Doodle, a sketch-to-image tool that converts a simple drawing into a dynamic image. Under the hood, Stable Doodle combines Stable Diffusion XL with T2I-Adapter, which offers additional guidance to pre-trained text-to-image (SDXL) models while keeping the original large text-to-image models unchanged. Stable Doodle is available on the Clipdrop by Stability AI website and app (iOS and Google Play) [Details].
Anthropic launched Claude-2, a ChatGPT rival, supporting up to 100K tokens per prompt (corresponding to around 75,000 words), with enhanced performance in coding, math and reasoning. It’s available via API and a beta website, claude.ai, for US and UK users [Details ].
Poe by Quora has been updated: availability of Claude-2 with 100k-token window length (including for all free users), ChatGPT-16k and GPT-4-32k models and new file uploading, URL retrieval, and continue chat features. Poe also released a macOS version [Details].
Objaverse-XL, an open dataset of over 10 million 3D objects, was announced by LAION, Stability AI and others. It was used to train Zero123-XL, a foundation model for 3D that displays remarkable generalization abilities [Details |Paper].
Google's chatbot Bard has new features: Python code export to Replit, tone adjustment, audio responses, image prompts, and more. Now available in Brazil, Europe and in 40 languages [Details].
Shopify to roll out Sidekick, a new AI assistant to support merchants by providing insights into sales trends, inventory statuses etc., along with assistance in editing website themes and responding to common queries [Twitter Link].
Vercel has announced the 40 successful applicants for its AI Accelerator, selected from over 1500 applications [Details].
LAION AI released Video2Dataset: an open-source tool designed to curate video and audio datasets efficiently and at scale [Details].
Google launches NotebookLM, an experimental AI-based notebook that can interpret and interact with your Google Docs to provide insightful summaries, answer queries, create document guides and generate ideas. Currently available in the U.S. only [Details].
Elon Musk has announced the formation of a new AI startup, xAI with the goal to "understand the true nature of the universe." Elon in a twitter Space: “I think a maximally curious AI, one that is just trying to sort of understand the universe is, I think, going to be pro-humanity.” [Details].
Google's AI medical chatbot, Med-PaLM 2, is undergoing testing in several hospitals, including the Mayo Clinic. The testers of Med-PaLM 2 will have control over their encrypted data, which Google won't be able to access [Details].
ElevenLabs announced ElevenLabs Voice AI Hackathon - a 3-day online event to build applications powered by ElevenLabs voice AI models [Details].
Meta AI released a Speech Fairness Dataset with 27,000 utterances from 600 U.S. participants, aimed at enhancing speech recognition fairness [Details].
Stable Diffusion XL is available free on PlaygroundAI now [Link].
Shutterstock will supply OpenAI with training data in a six-year extended deal, in exchange of gaining priority access to OpenAI's technology. The deal also includes a collaboration to bring generative AI capabilities to mobile users through Giphy, the GIF library Shutterstock recently acquired from Meta [Details].
Chinese startup Baichuan Intelligent Technology released Baichuan-13B, a 13 billion-parameter model trained on Chinese and English data. This Transformer-based model is open-source and optimized for commercial use. Baichuan-13B is trained on 1.4 trillion tokens, exceeding Meta's LLaMa model, which uses 1 trillion tokens for its 13 billion-parameter model [Details | GitHub].
🔦 Weekly Spotlight
AI companions with memory: an open-source project by a16z to create and host AI companions that you can chat with on a browser or text via SMS [Link].
gpt-prompt-engineer: An open-source AI tool that can generate a variety of possible prompts based on a provided use-case and test cases. The system tests each prompt against all the test cases, comparing their performance and ranking them using an ELO rating system [Link].
PoisonGPT - An article on how one can modify an open-source model, GPT-J-6B, and upload it to Hugging Face to make it spread misinformation while being undetected [Link].
Danswer: an open-source Enterprise QA tool that provides reliable answers to natural language queries from internal documents, supported by source citations. [Link].
🔍 🛠️ AI Toolbox: Product Picks of the Week
💎 Textomap: Generate interactive geo maps from text sources and ChatGPT prompts.
💎 Coframe: an AI-powered A/B testing tool for websites and apps that generates user-specific content and uses feedback to continually improve it.
💎 Afforai: summarizes, translates, and answers your questions about documents, spreadsheets, and websites. It can also scrape through multiple websites and extract key information.
📕 📚 AI Skillset: Learn & Build
What AI can do with a toolbox - Getting started with Code Interpreter [Link].
Free course: Build AI Apps with ChatGPT, Dall-E and GPT-4 [Link].
✨ 📌 If you find value in AI Brews, consider supporting the work behind it. Visit our Patreon page if you'd like to contribute. No obligation, of course. As always, your readership is much appreciated. Thanks and have a lovely weekend! Mariam.