Emad and Elad Walk Onto a Stage...

Here's your daily briefing:

• We agree with the premise of these tweets about the opportunity cost of sleeping on AI (which explains our decision to write this newsletter). The AI revolution is here whether we like it or not, and its effects far too extensive and wide-reaching to ignore.

• Check out this video of Ai-Da, "the world's first ultra-realistic artist robot" (named after Ada Lovelace, the creator of the first computer program) talking about creativity, feelings, and the meaning of art:

• This demo of Explainpaper looks pretty cool 🤓

We just watched this recent fireside chat between Emad Mostaque, founder of Stability AI, and Elad Gil, entrepreneur and investor.

Mostaque is quickly becoming one of the more interesting and important thought-leaders in the space. His recent sit-down with GIF-lord (and Stability AI investor) Jim OShaughnessy on his Infinite Loops podcast was one of the better conversations we've heard in a while about AI and it's effects on the world and our coming near-future.

Since we know you're busy, we thought we'd do you the favor of watching a few times and jotting down the main points of this fireside chat in an easy-to-read summary style.

So, without further ado:

Gil: How did you get started?

Mostaque:

• Hedgefund manager for years, invested heavily in video games and AI.

• Background in mathematics and computer science.

• Helped organize a Covid + AI initiative.

• Began to see the trajectory of this technology:

  • a) It would be open-source eventually anyway, so might as well do it now, and avoid a potential panopticon.

  • b) It would be great to give this to the world because people from all different backgrounds can take this and make wonderful things.

• "In a few years anyone will be able to create visual communication."

Gil: What was the actual starting point?

Mostaque:

• "As image generation came through, I saw that there was a need for someone to collate resources to extend and expand these. And I saw that that was coming with image."

• "What if this could be available to everyone? What does the trajectory of this tech look like? That's when I started basically funding the entire space, and image generation in particular, just through grants. I gave a grant to help Midjourney...but there was nothing expected in return because this is an awesome technology."

• "A year ago I realized that there was an opportunity to create a sustainable business around the commercial open-source software paradigm. Just like all databases and servers are open-source, if you have commoditized models and open models with open datasets, it's a new kind of programming primitive...scale and service get you to a billion people. And that's a ridiculous new market that's created. And that can be used to fund open-source AI for everyone."

Gil: How did you get involved with latent diffusion (which morphed into stable diffusion)?

Mostaque: 

• It all came together very organically.

• "Supercompute was the issue."

• Saw a massive brain-drain out of academia and tried to fill the gap between it and industry with grants and support.

• "I was talking to Robin, one of our lead coders, and I was like, 'Why don't we take this and accelerate it and see where it goes. I think this will be the breakthrough to a good enough, fast enough, and cheap enough model.'"

• "Open-source is about collaboration. Anyone can participate. We're going to move more and more toward open-source so can bring more people in."

Gil: Beyond the image generation areas, you've done a lot around voice, textual models, bio-science stuff... Can you talk more about the directions you're heading and what you're excited about?

Mostaque: 

• "I think the aim was to have full multi-modality, and I think we've got that now."

• "The model we have is that research is undertaken by communities."

• Funding lots of academic partners, fellows, and PhDs over the next 6 months.

• "Then we have the broader community (100,000+). What if we move from a few hundred to a few thousand people collaborating?"

• "The key thing: how do you make developers happy and give them agency? Every developer has in their contract that they can open source anything they create. There's also a revenue-share for developers. Any model we run, 10% of revenue goes to developers, half of that goes to a pool where the official thing is for the developers to allocate to the coolest shit they can find. This is what has allowed us to scale."

• "Open collaboration means more brain cells per square inch than anywhere else."

Gil: How are you thinking about the business model/monetization?

Mostaque:

• "Scale and service, customization."

• "People are using Stable Diffusion within their own companies now, and they ask us, 'Can you help us create custom models and scale them?' And we can earn nice margins for that."

• "Hopefully, fingers crossed, next month we'll be profitable. Which is unheard of for companies at this stage.... And I think this will go exponential."

Gil: Where do you think things will be in 5-10 years with regard to open-source vs closed source. Do you think it's going to be a patchwork, largely open-source, or more closed?

Mostaque: 

• "Open-source will always like closed-source. Why? Because closed-source can always take open-source and extend it. The advantage is not the models. A lot of AI companies think their advantage is the models and talent. But here's the reality. The amount of money that will go into generative media over the next few years will dwarf self-driving cars ...because of the immediate utility."

• "The amount of talent coming to the space is in the hundreds of thousands... By 5 to 10 years, this will be proliferated everywhere."

• "There's a place for open and closed source. It all depends on what you want."

• "I do believe the open frameworks should be open. I believe every country should have their own benchmark models. So that's what we're trying to do, we're trying to create some standardization around this. We're basically being the Switzerland of AI. We're open to collaboration with everyone and we actually listen."

Gil: How do you think about AI safety from a global lens and how do you think about globalization of AI more broadly?

Mostaque:

• AI alignment:

  • "I think AI is less likely to kill us all if it's designed to augment our capabilities and it's diverse. I think if it's very narrow and it's trained on Reddit it probably will."

• AI safety around offensive content:

  • "Imagine all content is generated...it still needs to be curated. So the curation aspect kind of eliminates the extreme edges."

  • "If you don't want it open...who is it controlled by? Unelected private companies, which isn't good."

• Geopolitics:

  • "The bad guys have AI. They're going to create deep fakes of the highest quality. This is why we have a $200k open source deep fake detector competition."

Gil: How do you think of government regulation of AI?

Mostaque:

• "One of the understandable reasons people want to regulate this is, you know, it's so powerful.... The known unknowns and unknown unknowns are crazy."

• "Within the EU there is a push to make the users end use of the model, the liability of the user is on the creator of the model, even if it's academics."

• "There should be an open discussion about it. Regulation is okay, as long as we have a democratic discussion about it. Like, I'm not anti-regulation, I just want to have proper regulation, proper thought-through discussions around it, with a wider group of people discussing than a few people in Palo Alto."

Gil: What are some of the use-cases for generative AI that people aren't working on that you wish existed?

Mostaque: 

• "It's going to be amazing when you can create anything. That means you can 'what ifs'. Like what if you did Muhammed Ali vs Mike Tyson live. Or Lebron vs Michael Jordan?"

• "You'll be able to make your own worlds."

• "My personal target: I'm going to remake the final season of Game of Thrones."

Gil: What technology directions are you most interested in?

Mostaque: 

• "I think there's some very interesting architectures that could potentially be alternatives to transformers... I don't think scaling is everything."

• "We need better data, we need more structured data. I'm very excited about what that looks like, but then taking that and extending it out to models that run on the edge. The combination of deep learning and reinforcement learning now allows us to take the big models and make them super usable."

• "I'm really interested in small models that are customized as opposed to large language models."

• "How can we optimize this? And get it out there to as many people as possible, as well as different countries."

Gil: From a country specific model perspective, do you think that's going to be pursued by individual governments or open-source groups?

Mostaque: 

• "We are talking to dozens of governments at the moment to see how we can get national broadcasting data and others to create national models and communities around them. So going in and funding the PhDs."

• Focus on building communities around the world to collaborate on these models.

• "The governments need help, because nobody else is willing to share the technology for us. It should be standardized but it should be shared."

Gil: How far away do you think we are from really performant short-form video generation?

Mostaque:

• "If you look at recent breakthroughs...The chaining together of all these models is what's been under-appreciated."

• "I would suggest that we can make decent short-form video within two years, at a high-resolution quality."

Gil: Why do you think the first wave of AI didn't really seem to create a lot of successful start-ups?

Mostaque: 

• "I think the complexity was a lot lower... A lot of the AI that's been done so far was glorified data science. And now you have a new type of AI. Is it AGI? No. It's another component of the brain that enables this to happen. Will all the value remain? Well the value is in transforming entire sectors. Is there a media company that by next year won't have a generative AI structure? No."

• Is there an ability to disrupt entire industries?

• "By the time you get to university, you won't need to write essays because the AI will do it better."

• "What a start-up needs to do is find and occupy a moat. So where is our moat? Where are we aiming? We're aiming to be a layer 1 standardized infrastructure layer on this. Just like a server or a database."

• "You come to us if you want to scale or you want to customize."

Gil: How far do you think we are from AGI?

Mostaque: 

• "I don't know man. What I do know is we will never be less than 18 months away from AGI."

• "My timelines are variable."

• "You don't need an AI to be AGI to be dangerous."

• "Our focus is not AGI. Our mission statement is to build a foundation to activate humanity's potential."

"two AI experts on a stage talking about the future"