Good evening from icy London,

I had great fun chatting with my friend Henry Ajder today. He’s an authority in this space, and we enjoy geeking out on this together. So we thought we’d record our conversation for you.

Our top picks this week:

Stable Diffusion V3 with training opt-out

Stability AI says it will allow artists to opt out of V3 of Stable Diffusion. Artists can register on haveibeentrained.com, an initiative from Spawning AI. ( Artist tutorial here.)
Emad told me that Clip, a text-to-image model from Open AI, was the source for artist complaints in Stable Diffusion V1 — so they removed it from V2. (See our whole interview here. For this point, skip to 25:07.)
Is the ‘AI training’ debate a red herring? Emad says that future models will likely become ‘zero shot’ -IE, able to generate content with an underlying model comprised of a few million licensed images. No large data sets or ‘scraping’ is needed.
Emad @EMostaque
The AI art debate misses that future models will likely be zero/few shot learners. Model base will have just a few dozen million images, licensed Show it a portfolio of images, write prompt and instant generation in that style See MJ v4 image promoting as an example/dreambooth
2:05 PM ∙ Dec 15, 2022
732Likes96Retweets

Do Not Pay

The ‘Robin Hood of the Internet’, Joshua Browder, has created Do Not Pay — the world’s ‘first robot lawyer.’
Based on GPT3, it’s a chatbot that aggressively negotiates refunds and discounts and gets you off the hook for things like parking tickets.
It’s an ingenious application of text generation, making an expensive and inaccessible resource (legal advice) widely available to consumers.
Joshua Browder @jbrowder1
DoNotPay’s GPT AI bot has something called “tag along” mode. Every time you visit a large company’s website and do something, it aggressively asks them for a refund, discount or credit, if it feels like you deserve it. Just got me back $19 for poor airplane Wi-Fi.
1:15 AM ∙ Dec 16, 2022
4,413Likes245Retweets

Apple’s M1 Chip and Stable Diffusion

Is Apple about to make its big Generative AI play?
We discuss the recent announcements that Apple is optimising Stable Diffusion to run on Apple Silicone devices. (Paper here.)
Stable Diffusion can already generate an image in an Apple Silicone Mac in less than 18 seconds.
Will Apple deploy Stable Diffusion in future products such as its rumoured augmented reality glasses?

Lensa's Soft Porn

Lensa, the viral AI app, generates ‘soft porn’. So while men are being imagined as astronauts, women are being imagined in bikinis. (See my Lensa-generated images below… Henry hasn’t tried it so we couldn’t compare notes.)
CEO Andrey Usoltsev blames the biases in the underlying model (Stable Diffusion) and says they’re building an NSFW filter.
Henry and I ask the philosophical question: what’s better? Models that reflect society (warts and all) or an idealised version of society?

China’s ‘Ministry of Truth’

The Chinese government has prohibited using Generative AI to “engage in activities that endanger national security, damage public interest.”
This is the perfect example of the ‘liars dividend’. It gives the Chinese Communist Party the ultimate authority to arbitrate the ‘truth.’
Will footage of human-rights abuses conveniently be dismissed as AI-generated?

Bullshit or Breakthrough?

Finally, in a quick-fire round of ‘bullshit or breakthrough’ — Henry gives the news about the Apple M1 chip a thumbs up. This one is a breakthrough, not bullshit.

Enjoy your weekend — the holidays are almost here!

Nina & Henry

Apple's Generative AI play?