Hey everyone and welcome to the 190 new folks joining us since Friday's email. You'll be happy to know the modern Indian was delish.

Another week of AI updates ahead, I wonder what we'll see this time. Let's get to what you missed over the weekend.

Prompt: List the topics in today’s email:

🌅 Midjourney v4 is out🖊 Highlight → get explanations🎨 Paint-with-words🗣 Generate a voice

🫦 BEN'S BITES

I spoke with the founder of Humanloop on Friday. It has a playground to build prompts but focuses on apps with traction, using user data to determine how good (or bad) your model is. You can automatically test various models at random with users and let data tell you what’s working. Check it out here³⁸⁶.

One of the best things I've read about AI recently was from Mario who's super smart and writes about tech and the future. He dives into a trend and how it relates to AI³⁵⁵.

Midjourney v4 is out! And the results are pretty effing epic. Here are some examples:

A v3/v4 comparison.

Also, you can create your own app icons⁴⁰¹ pretty easily with it.

👋 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.

Sam Altman did a Twitter AMA about their 5-week program¹⁹³, Converge¹⁴³.

What would happen in a world largely driven by AI content²⁷⁷?

Anne-Laure has collected over 180 AI tools for “The Artificial Creativity Landscape³³⁸”.

StereoPose¹⁶⁰ is a method designed specifically for transparent objects, using RBG instead of object point clouds.

SEO.ai⁸²⁴ does…well, exactly what you think it does. Like Ahrefs with AI baked in.

Generate background scenery^301† with AI.

A neural network that can learn to represent human motion²⁶¹. The pretraining task requires the motion encoder to recover the underlying 3D motion from noisy partial 2D observations. This pretraining helps the motion encoder learn geometric, kinematic, and physical knowledge about human motion, which can then be transferred to multiple downstream tasks.

Create your own worlds with 3D and AI, using your phone. The demo²⁹⁵ is pretty nuts.

Highlight any text and get a simple explanation²⁹².

A neural network to draw dicks²⁹⁶…(NSFW?)

A system that can take an input of multitrack audio and convert it to sound like a reference song.²³⁶

Multiple members of Twitter’s Machine Learning, Ethics, Transparency and Accountability (META) team have been laid off¹⁰².

Stability AI is funding a wide-ranging effort to apply AI to the frontiers of biotech. Called OpenBioML^227†, the endeavour’s first projects will focus on machine learning-based approaches to DNA sequencing, protein folding and computational biochemistry.

FactorMatte²⁵⁵: Redefining Video Matting for Re-Composition Tasks. Ok whut?! Basically, the goal of factor matting is to separate the contents of a video into independent components.

An easy way to design, produce and deliver fully-custom fashion products³¹⁸ & accessories - Cala.

Unofficial Implementation of Paint-with-words, a method from eDiffi that lets you generate images from text-labelled segmentation map.¹⁴¹

There are security risks associated with text-guided image generation models. It introduces backdoor attacks against these models¹⁰⁸ and demonstrates that their text encoders pose a major tampering risk.

Waveformer is a low-latency target sound extraction model¹³⁴ implementing streaming inference. Check it out on Hugging Face here¹¹⁷.

A new method for processing long text with transformer models.¹⁷⁰

Create a voice by tweaking a few sliders ²³¹- ThisSpeakerDoesNotExist.

🖼 AI IMAGES OF THE DAY

A few more Midjourney v4 examples:

👋 SEE YA