Hey everyone, it's the end of the week! It's been a busy one in AI (and on Twitter in general - is it still working?) Welcome to the 121 new folks joining since yesterday.

Have a great weekend - it's pizza-making for us in Ben's Bites household. I'm not trusted to roll out the dough, strictly pizza-oven duties 🫡 (no to pineapple).

If you want to work in AI, join the talent collective¹³⁴. If you're hiring, post a job + get featured in the newsletter here¹²².

See you next week! I'll try cutting down these emails too.

Prompt: List the topics in today’s email:

📝 Worlde for prompts🤗 Hugging Face demos are now in arXiv📲 A language learning app🎮 AI for game creation

🫦 BEN'S BITES

arXiv collaborated with Hugging Face to make demos related to papers directly accessible on arXiv²⁷⁷!

Search your codebase²²² with natural language.

Why people need to be more thoughtful in building products on top of LLMs³¹⁴.

Get up to $50k to create AI games, NPCs, and immersive experiences³⁴⁶.

Stanford announced HELM -- an AI benchmarking project aimed at improving the transparency of LLMs¹⁴⁰ and the broader category of foundation models.

Create, mint, and trade your AI-created collectibles²⁷⁹ on Revel.

Instead of spending hours crafting prompts, take a look at this simple thread³⁵⁸.

👋 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.

Exploring Google’s wildfire simulation project¹⁴⁸.

Sensible raises a $6.5M seed round³²² to make documents as accessible to software as APIs.

Speak, an app to make learning a language 10x easier through AI⁴²⁸, raised $27M from the OpenAI Startup Fund. The company has been in stealth and the first time it’s been talked about publicly, here’s a thread to give context²⁴⁴.

First place in the lablab.ai hackathon went to this team for their tool that kids can write their stories in simple words, and magically get illustrated by Stable Diffusion generated images²⁴⁰.

VC firm, Base10, created a market map and research article on generative AI⁴³¹. Highlighting early signs of a platform shift that could rival that of Cloud.

Trouble in Galactica, again. They’ve turned it off for now given all the misinformation and concerns from people. While it can present many threats, if used responsibly it can also be a powerful tool for science, as Moreno talks about here¹⁵⁵.

The Data Cards Playbook, by Google²⁴⁹. It’s a toolkit designed to help organizations navigate transparency challenges associated with their machine-learning datasets.

The latent diffusion 4x super-resolution model is now available¹⁸² in diffusers (super-resolution images).

Wordle for guessing prompts^308† from an AI image.

If you remember yesterday I shared Suhail shared his dataset from PlaygroundAI. Well, this is why…people are now building on top of it. This tool helps you to analyze and inspect the images and corresponding prompts from Playground AI Images⁴²⁵.

Meta has released ImageNetX²¹⁵, which is a set of human annotations of factors such as pose, background, or lighting for the entire ImageNet1k validation set as well as a random subset of 12k training images. The dataset is used to study the types of mistakes made by 2,200 current recognition models.

a16z have come out with another post around AI, The Generative AI Revolution in Games²⁵³. It discusses how generative AI is transforming the gaming industry by allowing for the creation of high-quality images in a fraction of the time it would take to generate them by hand. It also outlines the various segments of the generative AI market for games and identifies key companies in each.

The article discusses the concept of emergent abilities¹⁸⁵. It argues that scale is not the only factor that matters in determining these abilities, and that inductive biases (i.e. the assumptions made about the data when designing a model) can also have a significant impact. (Similar to the article I shared on emergence yesterday).

AI-generated game assets⁴⁹⁶. Create high-quality, style-consistent, proprietary assets for your games.

Swift Ventures and AI pub co-wrote a piece “The Next Generation of Large Language Models Will Blow Your Mind and Disrupt Your Business²⁵¹“. The article discusses the upcoming launch of GPT-4, and how it will be a more powerful and disruptive model than its predecessors. Key takeaways:The next generation of large language models will be more powerful, multi-prompt, agentic, and knowledgeable.These models will be able to execute long chains of reasoning and correct their own work.The next big trend after "generative AI" may be "agentic AI."

Contentinator is a new Figma plugin that lets you populate your designs with realistic content^295† of basically anything, with AI (of course).

How large language models can be taught to solve algorithmic³¹⁷ reasoning problems.

The $72M ARR Google Suite of generative AI³⁶². Sacra always write really good, in-depth pieces and this is no exception. It looks at Jasper, Copy.ai and others and how they’re positioned against startups and incumbents.

DiffusionDet is designed to learn to reverse the noising process in order to more accurately detect objects²⁶².

Uni-Perceiver v2: A generalist model for large-scale vision and vision-language tasks.²⁸⁷

A new approach to training deep learning models²¹⁴, using a neural network to optimize the model instead of hand-designed optimizers.

Diffusion models for synthesizing human motion³⁵¹.

A new method for editing images based on human instructions. The method, called InstructPix2Pix, is trained on a large dataset of image editing examples. The model is able to edit images quickly and produce a variety of edits for the same input image and instruction¹⁹⁴.

SPACEx is a method for generating high-resolution, expressive videos with realistic head pose²⁴⁶, using just speech and a single image.

A new model and benchmark for few-shot summarization³³³. The model, called UniSumm, is designed to be pre-trained with multiple summarization tasks and then tuned to excel at any few-shot summarization datasets.

The limitations of existing NeRF-based methods³⁰⁸ when reconstructing high-resolution real scenes, and proposes solutions to address these limitations.

I can't believe there's no images! Learning visual tasks using only language data²⁶⁵.

Conffusion - for generating confidence intervals around reconstructed pixels in an image¹⁶². This method is designed to be faster and more accurate than existing methods.

Andrej Karpathy wrote a short thread on the history of neural language models¹⁶⁰, and how they have only recently become mainstream. The key to their success has been the combination of a powerful neural net architecture (the Transformer) with a simple but scalable training objective (language modelling).

Null-text - inverting images using the Stable Diffusion model⁵¹⁰ by enabling intuitive text-based editing of real images.

🧑‍💻 WHO'S HIRING IN AI

VEED.IO - Simple Online Video Editing. VEED is hiring AI / ML engineers²¹¹ to level up its creative toolkit and make it more magical.

Buildspace - where builders, build! They're looking for an ML/AI instructor to build their new course³²⁶.

Add your company here⁹⁴ | Join the Talent pool⁹⁵

🖼 AI IMAGES OF THE DAY

All these 3 images are from Joshua Larson¹³⁹ using Midjourney - I love them!

👋 SEE YA