On the 21st day of Christmas my true love gave to me, an email talking about 3D DALL-E. Related; when do the 'days of Christmas' actually start?

Ho, Ho, Holy crap I've got to stop with the Christmas references.

Let's get to it.

🤌 Ben's Picks

Self-Instruct: Aligning language model with self generated instructions. So what's the story here? OpenAI can get a large amount of user data and afford to pay many human annotators for their InstructGPT. What if you can't afford to do that? This team instruction-tuned the GPT3 model and tested how good it can be at following new instructions. The result? Self-instructed GPT3 (which barely uses any human annotation) outperformed those trained on other public datasets, performing close to InstructGPT-001. Exciting!! (link⁴⁵⁶)

OpenAI prototype of 3D DALLE (called “Point-E”) is open sourced. (link⁴⁶⁷) Of course, there's already a Hugging Face demo for the new text-to-3D model. (link³⁶⁴) TechCrunch article here²³⁵. And you can read the arXiv research paper here¹²⁴.

Deleted tweet from Rippling co-founder: Microsoft is all-in on GPT. GPT-4 10x better than 3.5(ChatGPT), clearing turing test and any standard tests. (link^1,268)

This Detroit startup aims to shape the future of video creation with AI (Sponsor). 

Small Michigan startup Waymark is trying to break down barriers to video creation with the help of AI. With Waymark, users just type in basic information, and a team of AI models generates a complete marketing video for them in minutes. It looks like this approach is already picking up steam - Waymark has a few enterprise clients under its belt and has been listed in both Hulu and Roku's creative directories.

Check it out here³¹⁷.

🛠️ Cool Tools

• Using Zapier to integrate a Slack channel to talk with ChatGPT. (Zap template³¹²) (link to Jons Substack²¹⁸)

• Ellie - GPT-3 powered assistant that can reply to your emails, has been officially launched and the first 50 users to use the coupon BENSBITES get 40% off. (link⁴⁷⁰)

• GPT-3 + HARO Journo Response Generator - A tool that uses GPT-3 to identify and respond to press requests from journalists looking for sources for a story. (link²⁵⁴)

• Sentenceviz - A streamlit application to visualise sentence embeddings. (link³⁰⁵)

• Kailua Labs Image Search API - An API to easily search images, video, audio, and more with AI. (link^341†)

• Softr Studio now has AI - Generate images and copy within Softr with just one click. (link²⁹⁸)

• Gradio Discord Bot (BETA) - Use any Hugging Face Space as a Bot from within your Discord server. (link¹⁸³)

• Tele-Prompter - An on-device AI for your meetings that listens to you and makes charismatic quote suggestion. (link⁴⁰⁴)

• Project Relate by Google (Beta) - Android app that uses machine learning to offer personalized speech recognition for non-standard speakers. (link²⁴⁷)

• Text Blaze - Chrome extension that makes OpenAI GPT-3 available anywhere you type. (link⁴³⁷)

• Tome - Generate an entire presentation with AI using just a description of the deck. (link^741†)

• Albus - Your personal AI teammate right inside Slack. (link⁵³⁵)

Join Discord¹²² | Hackathon¹⁰⁸ | Partner with us¹¹⁶  | Host or join a meetup¹²²

🎓 Learn

• Transcribe and summarise YouTube video with 2 clicks with AssemblyAI. (link⁵¹⁷)

• What are Hugging Face Model Cards? (link²⁵⁸)

• Ebook: Natural Language Processing (NLP) for semantic search. (link²⁶⁸)

🔬 Research

• Improving video voice dubbing through deep learning. (link¹⁶⁸)

• Instructor: An instruction-finetuned embedder that can generate text embeddings tailored to any task given the task instruction. (link¹⁹³)

• Connect-the-Dots, a new accounting algorithm that uses an indirect approach to accurately discretise privacy loss distributions. (link¹⁰⁷)

• ‘Unnatural Instructions’: A paper that proposes a new method to automatically generate natural language instructions. (link¹²⁰)

• Rethinking the role of scale for in-context learning: An interpretability-based case study at 66 billion scale. (link¹²⁴)

• InstantAvatar: Learning avatars from monocular video in 60 seconds. (link¹⁶⁸)

• Why can GPT learn in-context? Language models secretly perform gradient descent as meta optimisers. (link¹⁵⁶)

• Pretraining without attention. (link¹⁴⁰)

• Length extrapolation: Training on short texts while evaluating longer sequences. (link¹¹²)

• SODA: Million-scale dialogue distillation with social commonsense contextualisation. (link⁹⁹)

• Panoptic Lifting, a novel approach for learning panoptic 3D volumetric representations from images of in-the-wild scene. (link¹⁰⁷)

• MetaCLUE, a set of vision tasks on visual metaphor, and a collection of high-quality and rich metaphor annotations. (link¹¹¹)

• LayoutDETR: Detection transformer is a good multimodal layout designer. (link¹⁵⁷)

• Character-aware models improve visual text rendering. (link¹¹⁷)

👋 Too many links?! I created a database for all links mentioned in these emails. Refer 1 friend using this link and I'll send over the link database.

🤓 Everything else

• Here’s what I saw at an AI Hackathon, an account by Dan Shipper. Some really interesting projects and observations from Dan's time at a hackathon in SF. (link³²⁸)

• How AI is being used in modern marketing: A beginner’s guide. (link⁵²⁸)

• OpenAI human-feedback dataset has been added to the Hugging Face Hub. (link¹⁶⁴)

• Unwrap this Twitter thread with delightful Christmas AI art. (link²²¹)

• How are students using AI? (link²³⁶)

• Petals is creating a free, distributed network for running text-generating AI. (link¹³¹)

• An investigative article on the origins of 15 video stills taken from inside homes by robot vacuums and shared to social media. (link¹⁶²)

• Australian Universities seek to curb ChatGPT use by students. (link²¹⁰)

• Google Search testing search in video. (link¹⁵⁹)

• Predictions: Future Versions of DALL-E 2, Midjourney, and more. (link²⁷⁰)

• A thread with companies using AI in fascinating applications. (link³⁶⁷)

• The near future of AI is action-driven. (link²³⁹)

• An interactive exploration of Transformer programming. (link¹⁶⁵)

• ChatGPT for iMessage, WhatsApp, user feedback, and beyond. (link²⁸¹)

• GPT gets chatty - The future of artificial conversation. (link¹⁹⁷)

🖼 AI images of the day

👋 See ya