Superfast AI 4/30/23

Hi everyone 👋

Another fun one this week! Today we’ll ask HuggingChat and ChatGPT to tell a joke, watch two DeepMind robots play soccer, and see what Midjourney thinks donut farming looks like.

Let’s dive in!

🗞️ News

1. Hugging Face launches a chatbot called HuggingChat, which provides another alternative to ChatGPTAh… and it looks like LLMs still can’t tell a very good joke (check out some of my earlier newsletters where I tried doing themed jokes that all… pretty much failed). At least ChatGPT’s joke is a bit spicier!

   1. Underlying the chatbot interface is a 30B parameter open-source model

      1. Open question to be tested: how well does HuggingChat perform compared to ChatGPT or other LLM chatbots? It would be great to see another comparison blog like this: ChatGPT vs Claude.

   2. Check out HuggingChat here: huggingface.co/chat/

2. RunwayML launches a mobile app for AI video editing

   1. Why? Creators capture videos on mobile, so it makes sense that they would want to edit and upload their videos on mobile too (à la Instagram?).

   2. What can it do? Transform videos from one style to another (raw footage → claymation style). Here’s an example video.

   3. Limitations? Currently, users are limited to 3-second clips and there are some delays in processing times.

   4. Overall, it seems to be a big, positive move for RunwayML as other video editors begin playing in the space (like Luma AI).

1. Deep fake AI-generated music? Some musicians are embracing it while others reject it

   1. Last week, a viral but fake song by Drake and the Weeknd song called “Heart on My Sleeve” garnered a lot of interest before being taken down soon after. Here’s a preview if you want to check out a clip that hasn’t been taken down yet. Universal Music Group (UMG), the label that represents the artists, issued claims of copyright infringement based on the likeness of the generated music and the original music creations. UMG claims the Metro Boomin producer tag to be an unauthorized sample, which is why the viral videos have been removed from all streaming services, which a YouTube spokesperson verified. What happens when future generated samples are completely novel, but still have the likeness of the original artists? Read a fuller dive on the legal entanglements of the Drake and Weeknd viral song here.

   2. In contract, Grimes has embraced AI generated music and is open to splitting royalties 50-50 with anyone who creates in the ‘style of Grimes.’

1. Google launches a new security focused language model called Sec-PaLM. This model is a fine-tuned sub-model of Google’s original PaLM model. You can read a short blurb I did on PaLM here. The fine-tuned Sec-PaLM model is specifically attuned to software vulnerabilities, malware and threat indicators, and more. You can read a full post here.

Fundraising

• OpenAI closes $300M share sale at $27B-29B valuation (Link)

• Pinecone receives $100M investment on $750M valuation, as vector database demand grows (Link)

• Replit raises $100M, bringing its post-money valuation to $1.16B (Link)

📚 Concepts & Learning

DeepMind trains two robots to play 1v1 soccer

DeepMind trained two robots to play 1v1 soccer.

The process:

• The robots are trained entirely in a simulated world on millions of simulated games, and stress-tested in real world games.

• The robots are given real-time action sensors to give them up-to-date information about body, ball and opponent location, among other factors.

Some interesting rules:

• The reward: number of goals scored

• Penalties:

  • If a player is within a 1-meter distance of the opponent

  • If a player over-stresses their knee joints, especially during kicking and stepping. This taught the robots to move with softer gaits.

Some interesting results:

• Combining skills: The robots are able to combine multiple skills (get up after a fall, score a goal) into a seamless single policy (play soccer)

• Quick recovery: the learned policy can dynamically react to disturbances and recover quickly. The robots can: quickly turn, maintain ball control, position themselves to score a goal, intercept ball trajectories and employ higher level strategies such as positioning oneself between the ball and the goal to block their opponent’s shots. All of this complex behavior is learned from a few simple rules. So cool!

Read DeepMind’s post here or their Arxiv paper here.

Check out a tweet thread here and here or the YouTube here.

In case you’re interested in more projects at the intersection of robotics and AI, check out this post I did a few weeks ago on the research at Google on SayCan and on a startup’s autonomous food delivery services.

🎁 Miscellaneous

Bonus round: AI-generated music

If you loved the AI generated collab between Drake and the Weeknd, check out this sample of Kanye covers on various songs (Link)

The power of plugins

Interested in ChatGPT plugins? Check out this post on how this startup built a YouTube video summarizer (Link).

ChatGPT Prompt Engineering for Developers

Check out this new, free course by Andrew Ng (DeepLearning.AI) and Isa Fulford (OpenAI) on prompt engineering best practices. (Link)

A throwback but a goodie

Check out this 2 minute time-lapse of an hour of Stable Diffusion by Andrej Karpathy (Link).

And finally, Midjourney’s interpretation of A World Where Everything is Grown:

Check out the full, delightful thread here.

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

What did you think about today’s newsletter? Send me a DM on Twitter @barralexandra or reply to this email!

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