Ben's Bites
Posts
It's Science

It's Science

🧑‍🔬 An LLM for science 🔢 Formulas made easy 🖥️ Building a supercomputer in 3 days 🖼️ Multi-flow multimodal diffusion framework

Ben Tossell
November 16, 2022

Ello ello, and welcome to the 104 new folk joining us since yesterday.

We've got a big feast today, take it slow, and peruse at your own pace, perhaps you've got a load of pointless meetings today and you want to just look at cool stuff. Well, I've got you.

Let's get to it.

p.s. Emad from StabilityAI is doing an AMA on Reddit today if you wanna ask any Qs (in between meetings, of course 😏).

Prompt: List the topics in today’s email:

🧑‍🔬 An LLM for science🔢 Formulas made easy🖥️ Building a supercomputer in 3 days🖼️ First unified multi-flow multimodal diffusion framework

🫦 BEN'S BITES

Galactica launched with a big bang (😉). A large language model for science that can summarise academic literature, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins, and more. It has been trained with 48 million papers, textbooks and lecture notes, millions of compounds and proteins, scientific websites, encyclopedias and more. Sound smart, be dumb, like me.

Formula God - write formulas in Sheets, in plain English.

Priceloop is implementing GPT-3 custom functions in formulas too.

Add audio to your publication with one line of code.

What will happen when our phones record everything we say 24x7 using AI to process that information? This person tried it out and here are the results.

Those who can’t do, teach. Teaching LLMs to teach themselves new tasks by teaching them to generate few shot examples from high-level task descriptions.

A visual history of artificial neural networks from 1943 to 2020.

Tutorial: Containerizing Huggingface Transformers for GPU inference with Docker and FastAPI on AWS.

👋 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.

Naveen Rao on building the world’s first AI chip and bringing Machine Learning to every industry.

Generate images in Chinese and English text. The space for Bilingual Stable Diffusion is out now.

Daniel Eckler with another mammoth thread: 90 Days of Diffusion, 90 AI Advances.

UPainting can effectively generate simple scene images as well as complex, designed to improve the fidelity and alignment of generated images.

Tutorial: Serverless Machine Learning Applications with Hugging Face Gradio and AWS Lambda.

A demo of an Airtable → DALL-E script.

An all-new version of Descript has been released (it's a powerful video editor) and it announced that the OpenAI Startup Fund will be leading our $50 million series C. You can see the breakdown of all the new features here.

A curated library of prompts.

Notion QA Bot: ask questions on your Notion database and receive an accurate, conversational response back. Similar to Ask My Book we saw yesterday. I think we’ll see lots of these kinds of products - I want one for my Airtable base to easily search links that have been included!

Tutorial: Training video classification models easily.

Turn yourself into an AR character. Using AI to create a digital twin and then bringing it into an interactive augmented reality experience.

The legality of AI and whether the outputs can be copyrighted.

Hugging Face has come up with a new way to calculate the emissions produced by LLMs more accurately.

Testing Google’s writing tool, Wordcraft. How does it stack up against GPT-3 powered tools?

Built on the back of InteriorAI, is InteriorsByAI. A curated collection of AI-generated interiors in various styles. It’s already had over 10k views.

A new pipeline for creating and running Fast Transformer models on CPUs - Fast DistilBERT on CPUs.

Retrieving desired musical instruments using reference music mixture as a query. Essentially, pulling single instrument sounds from a track. For audio samples and demo, visit the website.

QueryForm - zero-shot transfer learning for document understanding. The framework is designed to help reduce the cost of annotating document entities and enable models to learn from structured documents containing various entities and layouts.

Who doesn’t love balloon art?! Now you can use it in your prompts for Stable Diffusion. I could only ever make snakes, worms, sausages, that kinda thing. Not any more!

Large Language Models struggle to learn long-tail knowledge. The number of documents a language model is exposed to during pre-training affects its ability to answer fact-based questions.

Semantic information from language models can be incorporated into self-supervised speech encoders without labelled audio transcriptions. This unsupervised approach achieves similar performance to supervised methods trained on labelled audio transcripts, demonstrating the feasibility of unsupervised semantic augmentations to existing speech encoders.

The use of language models for question-answering in a low-resource setting.

Versatile Diffusion is the first unified multi-flow multimodal diffusion framework. It natively supports image-to-text, image-variation, text-to-image, and text-variation, and can be extended to other applications such as semantic-style disentanglement, image-text dual-guided generation, latent image-to-text-to-image editing, and more. You can use the demo on Hugging Face.

Image compression uses text embeddings to generate high-fidelity images.

AI Image Generator in Notion - tutorial & free template (no coding required).

Cerebras built a supercomputer in 3 days, you just enter a line of code, specifying how many CS-2’s to run it on, and you’re done, and finally, the system demonstrates nearly perfect linear scaling.

Text-guided real image editing does not require fine-tuning or optimization and can be applied to a single real image.