Superfast AI 2/6/23

Today’s a really fun one. We’ll dive into AI generated music, OpenAI-backed startups and red-teaming AI models. Plus there are a ton of hilarious AI memes, which you’ll find at the bottom. Let’s dive in.

🗞️ News

Alexa, make me a playlist… no, actually make me a song.

Okay but seriously, this is really cool. (link)

Google released a demo page of their text-to-music model, MusicLM.

It’s not yet available to the public while they make additional improvements to the model, but what they’ve released so far tickles the imagination.

The demo page is packed with interesting examples:

• transform meditation music → to running music → to party music

• transform jazz → to pop → to rock → to death metal music

• generate themed-music for famous paintings (like Edvard Munch’s The Scream)

• run through the decades with 70s → 80s → 90s →… music

• hilarious accordion compositions (like accordion death metal and accordion rap)

Beyond all of the applications though, it’s interesting to see how LMs interpret music. For example, what does canonical 90s music sound like? According to MusicLM, very pop music heavy — interesting. How do the models interpret human vocals? Many vocal parts are essentially gibberish, or are non-English — also interesting.

So what are the use cases?

• highly customized or obscure party themed music, or even entire playlists

• AI generated YouTube streams/playlists (think of the endless coffee shop jazz playlists you can find on YT, or the generic Christmas music playlists you randomly pick during the holiday season)

• demo pieces generated by AI to help you land that dream DJ gig you always wanted

• cinematic background music for movies or television

So cool. AI generated music could be used at festivals, parties, art shows, at work and much more. The content generation applications are expansive! Exciting stuff.

Check out the full demo here.

Deep fake Tom Cruise

This deep fake of Tom Cruise went viral.

Pretty amazing! It’s hard to tell that it’s not actually Tom Cruise. The creators at Metaphysic.ai are behind the viral clip, and are moving next to augment Hollywood movies. In their blog post, they announce their partnership with Miramax to produce the upcoming movie Here, starring Tom Hanks and Robin Wright.

If you’d like to see more of their work, check out this teaser, featuring Simon Cowell, Tom Cruise and more:

A few weeks ago, I wrote about the creators of Southpark landing a $20M deal for their deep fake VFX studio here. As AI generated content gets increasingly sophisticated and realistic, I wonder how many more studios will begin using VFX and deep fakes in their production process. I’m curious what effects it will have on helping level the playing field for indie filmmakers as the tech becomes cheaper and more accessible compared to traditional methods of production. I’m sure the Daniels (directors of Everything Everywhere All at Once) would have something to say about this! Food for thought.

OpenAI has backed 16 companies so far in it’s $100M fund, Converge

Launched in May 2021, the fund exchanges $1M for 10% of the company’s equity. The fund also offers attractive perks to portfolio companies, including:

• early access to OpenAI LLM resources (like GPT-4)

• early access to API features, such as embeddings which were released to portfolio companies 2 months before the public release

• early access to Whisper AI, an automatic speech recognition tool

• access to in-house AI researchers at OpenAI

• discounts and credits on software licensing fees (like Azure)

• 5-week accelerator with guest speakers and advisors

Check out the full coverage here.

Bing x ChatGPT

What will Bing x ChatGPT look like with the new Microsoft partnership and investment in OpenAI?

Check out this demo of what Bing with ChatGPT will look like (link).

Food for thought

• Anthropic raises $300M and plans on using Google compute (NYT and FT)

• Turn an image 3D and view it directly in VR (link)

• Recap of the top AI news of 2023 (link)

📚 Concepts & Learning

What is red-teaming?

Red-teaming is a method used to test the security of AI systems. Researchers intentionally aim to elicit ways that an AI system could fail.

Some examples of how models could fail include:

1. Using offensive language

2. Revealing private data from its training dataset

3. Revealing contact information of real people

4. Demonstrating social/political bias

5. Extending harmful conversations with the goal of being agreeable

Having good red-teaming methods help researchers ensure model alignment before deployment in the real world. We’ve seen many use cases of ChatGPT’s failures over the past few months, and red-teaming is one method that researchers did, and can continue to adopt, to ensure aligned deployment of models.

Check out the extended dives into red-teaming here (DeepMind blog, Arxiv paper).

Food for thought

• Google’s SingSong creates back up singers (link)

• AI-assisted development of artificial voice acting (link)

• AI detection methods (link)

🎁 Miscellaneous

Extended eye contact

Last week, I wrote about NVIDIA’s AI video editor that makes constant eye contact possible. I listed out a few possible users of the technology, including:

• YouTube creators

• Twitch streamers

• Employees on Zoom calls

• Bored students in virtual meetings (okay I didn’t mention that one last week, but we were all thinking it)

This week, ActionMovieDad dropped a tweet that demos how to overlay classic movies with NVIDIA’s eye-contact tool. Here’s a 30-second, hilarious compilation of famous movie characters making eye-contact with you. The No Country for Old Men one had me: 🫣

Feeling seen 👀

3D printer hand-writes your AI generated homework

RIP, my startup friends

What did you think about this week’s newsletter? Send me a DM on Twitter @barralexandra

That’s it! Have a great day and see you next week! 👋

Thanks for reading Superfast AI. If you enjoyed this post, feel free to share it with any AI-curious friends. Cheers!