Subscribe

Hey, Explorer!

Like images and video before it, music is having an AI moment. New models and tools from Meta and others are poised to fill the internet with AI generated music. How will everyone take it? Will we be able to tell the difference? Meanwhile, AI is making its way into everything from White Castle to Tinder. Things are getting faster. Hang on!

In this edition of Strange Magic…

• 🎵 Meta leads the charge into AI generated music

• 🍎 Apple’s alternate path to the Metaverse

• 🖼️ Midjourney tips - we got ‘em

• ✨ Artifact brings AI to your news feed

Read time: 6 minutes

🔮 What’s New

Music is having its AI moment

In the spring of 2022, images had their AI moment, perhaps kicking off the generative-AI explosion that we are currently immersed in. Early examples of text-to-image generation models showed some promise, but it was the June release of Midjourney followed by Stable Diffusion in July, both on the heels of DALL-E 2’s April drop that sprung the movement into action.

The same thing seems to be happening now around text-to-music generation with the primary focus this week on Meta’s open-source AudioCraft release. The new framework of AI models for generating music, as well as sound effects, is now open to the public, and honestly better than anything that’s come before it. Including Google’s MusicLM, still only available in limited testing.

With AudioCraft being an open source code base, we’ll now just have to wait (although probably not long) to see a wave of new music generation tools and apps. We’ll soon have AI generated tracks for our social media videos, we’ll have AI lo-fi study music streams, and we might hear AI generated hit songs. We’re definitely not there yet, but it’s not a stretch to think it could happen before we know it. In fact, we might not know when we hear the first AI generated until someone tells us.

In related news, Suno.ai, the company behind the Bark text-to-voice AI, released an alpha version of their lyrics-to-song (text-to-music) platform called Chirp. Similar to Midjourney, Chirp works by typing a command into Discord, in this case a set of “lyrics”, and in around a minute the model renders out a short song in a random genre featuring your lyrics.

It’s at the same time kind of glitchy, a bit kitschy, and totally mind blowing. With practically no control over the results, it’s impossible to try and “compose” any ideas with Chirp right now, and many results aren’t great to listen to, but there is a random serendipity that occurs every so often when the right set of words and genre render out a real banger, and there’s no doubt a few Chirp generated tunes will appear behind trending TikToks before you can hit mute.

Just like the previous art forms to be gobbled up and reimagined with AI, we all need to be ready for the AI music onslaught, complete with complex questions about copyright and human creativity. But, it’s definitely on the horizon, and it sounds like it’s going to be interesting.

In Other News

• AI with mustard: White Castle is bringing AI to drive-thru’s. The plan for better order management includes adding AI-generated voices at menu boards across over 100 locations next year. At least the AI won’t judge you for ordering that strawberry cheesecake on-a-stick!

• The virtual world according to Apple: Meta has not been able to make the Metaverse a thing despite immense spending. Now, Apple has partnered with Pixar, Adobe, Autodesk, and Nvidia to form The Alliance for OpenUSD. In short, they want to standardize the way 3D worlds are built for devices like their upcoming Vision Pro.

• Tinder is your mom now: In an attempt to “eliminate awkwardness” for daters, Tinder is working on several AI-assisted features like one that will go through user’s photo albums and pick their best images, and another to rewrite user’s profiles to highlight their best qualities. The question is, how will you know if you are swiping her or her?

📖 Good Reads

Yes, AI models can get worse over time

Believe it or not, we live in an age where our software may get dumber. AI models aren’t hard programmed with information. This enables the ability for them to learn over time, but just like humans, they also appear to have the ability to unlearn. A recent study found that OpenAI’s GPT-4 model answered some math questions far less accurately than a few months prior. It was also less verbose and developed new quirks. But, over the same period, it got better with some visual reasoning problems and gave safer answers. We can’t have everything.

Large language models, explained with a minimum of math and jargon

Look, the real magic with AI large-language-models runs so deep that many of the top Ai experts can’t even explain it. But, if you want a pretty thorough deep-dive into how LLMs work without needing a prerequisite PHD, Tim Lee and Sean Trott have you covered.

More Reads

• Apparently Adobe staff are worried about its AI software

• Uncensored and ‘insane’: A look at OpenAI’s secret image generator

• The AI rules that Congress is considering, explained

⚙️ Try These

• Artifact News: This personalized new reader includes a host of AI integrations, including summary generation, AI voices to read you articles, and even a tool that rewrites click-bait headlines to save you from yourself.

• Shop AI: Shopify has added an AI chatbot to its platform-wide shopping app. The addition gives shoppers an improved ability to search using more natural language, with followups and guidance from the bot.

• Perplexity: Perplexity looks a lot like a typical chatbot, but it operates more like a search engine that looks for answers and then summaries them, attempting to give more reliable results than some competitors. The latest updates support file uploads.

🧪 The Lab

• Meta AudioCraft Demo Walkthrough: Before you dive into Meta’s Audiocraft, let MattVidPro walk you though the in’s and out’s of the cutting edge sound model.

• We’ve got two quick tips for Midjourney this week.

  • Using the “—video” parameter in Midjourney

  • Using the “—weird” parameter in Midjourney

✨ We’re Diggin’

• Your new medicine may have previously been extinct: Scientists at the University of Pennsylvania’s Machine Biology Group have used AI to resurrect molecules with antibiotic properties found in extinct Neanderthals and Denisovans. With current antibiotic advances slowing, these scientists are using future tech to look into the past.

• You get an AI meme, and you get an AI meme: Love or hate AI, it’s really great for making memes. Here are some of the best… so far.

🌤️ Fresh Air

Whats happening outside of tech?

• Sun bears ≠ people in costumes: ICYMI, there was an uproar this week over a sun bear at a Chinese zoo that was caught on camera standing and waving at onlookers like a Disney animatronic gone rogue. But, in an online statement, Angela (the sun bear in question) stated, “I will stress again, I’m a sun bear!” ‘Nuff said.

• The 9 most beautiful lakes in the world: These astounding lakes may look imagined, but they are as real as it gets. Meet you in Croatia!

• Chop chop, ya don’t stop: You only need to watch one of Nicole Coenen’s woodchopping videos to realize what a weakling you are, but you’ll keep watching for the inspiration and knowledge.

Gimme that sweet, sweet feedback! We might feature it in the next newsletter.

Well, that’s enough strangeness for one day. Hope you learned something new. If you enjoyed it, please spread the word! Our robot is saving for a Vision Pro. 🤖

And, of course, if you have anything interesting to share, or just want to say hi, please feel free to reach out to us.