⚡ Why Can't AI Get Hands Right?

⚡ Why Can't AI Get Hands Right?

Reading time: 3 minutes

Happy happy week! I hit 100 subscribers. Thank you, feels electric, feels like a dream ⚡

I imagine myself sitting in a (big) cozy living room with you all. We're sitting in a circle and having a glass of wine (white bc it’s getting warm). This is what I'm telling you:

• 1. 🙌 Q&A: Why is AI so bad at hands?

• 2. 🛠️ Tool: Wordtune

• 3. 🕵️‍♂️ Concept: AI has a personality

• 4. 💬 Conversation starter: Unreal Keanu Reeves

• 5. 🏖 Text to video is getting good: stormtrooper level

1. 🙌 Q&A: Why is AI so bad at hands?

• AI is hands-down horrible at generating hands. That's because it doesn't actually know anything.

• Think of our savant toddler: dumb as hell but will do well if you tell it exactly what you want. When you feed it images, it thinks that's what you want.

• Savant toddlers don't make 3D scans. It's not creating a model of what you look like. It's guessing based on previous data (input images).

• So 2 main reasons:

  • 1. Data labeling - Photos don't always show all five fingers equally; holding objects can make it seem like we have 8 or 6 fingers. The data isn't labeled to reflect that. So when our savant toddler is fed 500 images, it doesn't make a model. It just sees 8 fingers and thinks, "cool, sometimes humans have 8 fingers."

  • 2. Margin of error - I've had DECADES of experience recognizing hands. The opposite is true with, for example, aliens. I accept any representation. With hands, I have a fixed idea. Low margin of error. Hands are my jam.

• Interesting: HK media institution South China Morning Post has started using AI-generated images. With bad hands.

• DM/email me questions for next week!

2. 🛠️ Tool: Wordtune

• ChatGPT for proofreading with nicer integration.

• Wordtune is possibly the best tool I've seen so far for coherent paraphrasing.

• Rating: ★★★★☆

• Cost: 10 uses a day for free, then $10 a month for unlimited. I don't pay for it, but could see myself doing that in the future.

3. 🕵️‍♂️ Concept: AI agents form a society

• Think of the news. Is unbiased news possible? Even choosing what to publish is, to some degree, biased. Similarly, is unbiased AI possible? Even no personality, is a type of personality. A savant toddler has a tone. And every toddler is different.

• For example, Bing and ChatGPT have slightly different ways to represent themselves. Bing is occasionally “unhinged”, according to the Verge. GPT is formal, professional.

• Take that concept of AI personalities, and extrapolate it. Imagine 25 different ChatGPTs all with different personalities and memories. Unleash them in a town, and see how they interact with each other. 

• That's the base of a paper about 25 AI agents who were given different personalities and unleashed into a virtual town. 

• Researchers gave the agents the ability to retain memory, reflect and make long-term plans. The AIs were able to use those to discern actions towards their goals.

• They had jobs, families, friends. They even organized a Valentine’s party by themselves - the event was not pre-programmed by the researchers.

• And get this - the craziest part - at the end, the researchers conducted interviews with the agents, asking them about their life, daily schedule, past events. Not only were they able to answer, they were able to “synthesize their past experiences into meaningful reflections that can enhance their future behavior”. Not even all humans can do that (haha).

• The implications are hard to digest and imply a gamut of possibilities from Westworld to AI society research.

• Interesting throughout, 3-minute video here, full paper here, and watch the whole experiment here. They detail current actions per character, conversations, … Recommend.

• BTW, remember last week was about autonomous AI agents who can define tasks and plan next ones. This goes further. It’s like the autonomous agent’s objective was to be human.

4. 💬 Conversation starter: AI Keanu Reeves

• Icebreaker: did you know about AI Keanu Reeves? He’s got 9 million TikTok followers. 225k on Instagram.

• That gap shows to me partially the tech discrepancy of Gen Z vs millennials. I of course discovered this through a reposted TikTok on IG reels. Time comes for us all.

• How will Gen Z or Gen Alpha (post-2010) see generative AI? Is this like when I, a millennial, talk to my parents about using Instagram filters? If it is different, why is it different?

5. 🏖 Text-to-video is getting good. "a storm trooper vacuuming a beach" via Nvidia’s latest demo.

See you in our little circle next week. Stay hydrated, eat vegetables. I wish you nice human things like sunny skies and solid sleep.

Cheers,
Tanya

PS.: I launched themultiverse.ai, a selfie-to-headshot platform this week. It transports you to a multiverse where you can be bad at posing and still get great portraits. We can have it all. Check it out here.

Previous editions:

• Beginner’s Guide to Generative AI ⚡

• AI agents and game designers losing jobs to AI 🎮

• How to use ChatGPT in Excel 💻

📧 You’re still here? If you stayed to the end you're probably cool and I'd like to connect. Give me a book recommendation or just say hi on Twitter or LinkedIn.