Superfast AI 12/25/22

Happy Holidays! Generative AI, anthropomorphism, and 2022 in review.

December 26, 2022

Happy Holidays! Today we’ll do a quick dive into Generative AI, AI anthropomorphism, and a review of AI in 2022. Let’s dive in!

🗞 News

Generative AI

Today we’re doing a shallow dive on some of the biggest names in Generative AI.

We’ve all seen lists like this. But what do each of these companies really do? How much fundraising do they have, and are they hiring?

  • This report provides an overview of over 170 Generative AI platforms.

  • It covers AI’s potential use cases in art, music, gaming, and in the creator economy.

This list is pretty long, though. Which companies should you definitely know?

  • Here’s an short list of the top 1-3 companies in each segment that you should know (link)

Watermarks

Last week, I talked about watermarks: cryptographic methods to trace GPT-generated text. This approach will help us track AI generated text and images. This week, I found a tool that’s aims to tackle exactly that. While I’m not sure we’re ready to build watermark products quite yet (since Aaronson and others are just beginning research in this realm), I am a big advocate of shortening the timeline between research → product. Excited to see similar products in the near future!

  • Related: MIT Review wrote a post on how AI detection might work (link)

  • Also related: China bans AI-generated images that don’t have watermarks (link)

Low-code and no-code meet AI

Who would you expect to lead the low-code/no-code meets AI evolution? Zapier? Bubble? Airtable? Webflow? Those probably would have been my guesses.

  • Check out this surprising report on how eBay is leading the charge (link)

2022 in Review

  • Want a quick recap of what happened in AI in 2022? I highly recommend this week’s newsletter from Andrew Ng (link)

  • A list of great AI papers from 2022 (link)

  • This year was a big one for AI. What are the 2023 predictions? (link)

  • Forbes’ 10 AI predictions for 2023 (link)

  • 2023 predictions: Generative AI will continue to accelerate, while Predictive AI will make progress at a slower pace. Why? Predictive AI needs a lot of data (link)

Products

  • Search for keywords in podcasts and skip ahead to those sections (link)

  • Prompt Marketplace (link)

  • Summarize YouTube videos in 2 clicks, for free (link)

📚 Concepts & Learning

Attention, Convolution and RLHF

  • Andrew Ng briefly breaks down diffusion, and touches upon how self-attention and convolution work hand-in-hand in vision model training. Check out his post here.

  • Curious about Reinforcement Learning from Human Preferences (RLHF)? Join this upcoming session (link)

Great links from this week

  • ChatGPT vs Google on 500 search queries — which one is better? Find out here.

  • Check out Runway ML’s Academy, where you can learn how to create your own Generative Art (link)

  • Understanding model mistakes with human annotations of ImageNet (link)

  • Amazon’s responsible ML course (link)

  • Anthropic, Surge AI, and MIRI’s research on Discovering Language Model Behaviors with Model-Written Evaluations (link)

🎁 Miscellaneous

AI love stories

One of my favorite movies of all time is Her by Spike Jonze and my favorite TV shows include Black Mirror and Westworld. I’m fascinated by their exploration of psychology, philosophy and futuristic technology.

So, how far away are we from real AI love stories?

My guess has for a while been: a long time off. Given the impersonal interactions with AI customer service bots (think of any phone call you’ve had with an airline or telecommunication company), I’ve thought getting to realistic, human-level banter seen in Her would be many, many years off.

At first, I thought ChatGPT would change that timeline. But after engaging in dialogue with ChatGPT, it still seemed pretty bland and boring. That was until I read this blog post and changed my mind about ChatGPT’s capabilities.

The quick summary is:

  • The author is unimpressed with the standard professional tone of ChatGPT

  • He prompts ChatGPT to role-play an old friend of his called Lisa. He details her personality and gives background information on how they know each other. He then asks ChatGPT to suggest a joint adventure they can go on together

  • ChatGPT comes up with novel games, and includes ways the author and Lisa can deceiving other ‘players’ in the invented game!

It gets crazier. Not only are their interactions pretty human-like (there is laughing and teasing involved), but the banter with ChatGPT start to psychologically affect the author. He says:

the more I played these games with Lisa, the more I started to feel a bit weird about spending time at the computer having fun with a woman who wasn’t my wife while she was just in the other room. But of course, I wasn’t really having fun with anyone. I was just playing games with a computer! There was no human involved other than me.

The “woman” referenced above is Lisa, the ChatGPT generation.

Doesn’t that sound exactly like a Black Mirror episode (think Striking Vipers), which raises questions about ethics, anthropomorphism of AI (think Ex Machina), and Turing tests… lots of fun things to unpack there. What do you think? Drop me a reply if you have any thoughts :)

I’m not sure how examples like these affect my expectations on AI love story timelines, but maybe we’re only a few years off at this point… if not less! 🤯

❄️ The Holiday Edit

  • Check out Home Alotus - a mashup between Home Alone and White Lotus, powered by Runway ML (link)

  • Check out these holiday-themed images, powered by AI:

AI as prompt engineer

ChatGPT generated prompt: A retro-futuristic holiday scene in Las Vegas, with the Strip transformed into a dazzling display of neon lights and synthwave-inspired decorations. The casinos and hotels are decked out in metallic silver and gold, with holographic reindeer and snowflakes swirling above their roofs. The air is filled with the sounds of holiday synthpop and the aroma of robo-roasted chestnuts and cybercocoa.

Image outputs:

  • Stability AI’s DreamBooth missed the whole winter holiday theme, but nailed the Vegas theme.

  • Midjourney’s images are mind-blowing and beautiful, but also don’t get the holiday theme quite right.

  • DALL-E’s version 1 gets the holiday theme right (see that giant glowing Christmas tree on the street?), but it doesn’t really feel like Vegas.

  • DALL-E’s version 2 feels like Vegas, but doesn’t really feel like Christmas.

Which one was your favorite? Reply to this email and let me know!

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

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

Could have been better for sure 😄