Frontier Model Forum | LK-99 | Text2Room

AI Daily | 7.26.23 [late]

Welcome to another episode AI Daily. This episode brings together three distinct stories - the inception of the Frontier Model Forum by OpenAI, the intriguing LK-99 ambient pressure superconductor research, and the innovative Text2Room that converts text prompts into 3D point spaces of rooms. The Frontier Model Forum underscores the need for collaboration in AI safety, functioning as a consortium of foundational AI model providers, aiming to lead the industry towards beneficial advancements. Next, we dive into LK-99, a potential game-changer for computing, with its potential applications across various fields, including AI - its authenticity is yet to be confirmed. Lastly, we explore Text2Room, an impressive engineering solution that takes us from textual descriptions to 3D spatial representations.

Quick Points

1️⃣ Frontier Model Forum

  • OpenAI initiates the Frontier Model Forum to foster industry collaboration for AI safety.

  • Serves as a consortium of foundational AI model providers.

  • It aims to instill more trust and potentially lobby for AI advancements.

2️⃣ LK-99

  • LK-99 is proposed as a room temperature, ambient pressure superconductor.

  • Potential applications span across computing, medical, and power grids.

  • Its authenticity is currently under investigation.

3️⃣ Text2Room

  • Text2Room converts text prompts into 3D point spaces of rooms.

  • Uses a 2D model to take images and build a 3D point space.

  • Represents a significant step forward in the field of text-to-3D.

🔗Episode Links:

Connect With Us:

Follow us on Threads

Subscribe to our Substack

Follow us on Twitter:


Transcript:

Ethan: Good morning. Welcome to AI Daily. , and we have three very different but interesting stories for y'all today. So we're kicking off first with the Frontier model form. So this was put out by OpenAI as really another push towards, you know, industry collaboration, AI safety. Really trying to put this like a government body around it, it seems, um, little bit hand wavy of sorts right now, but they are trying to put together pretty much only foundation model providers into this form and say, Hey, how can we all work together, you know, potentially for G P U usage, potentially to kind of collaborate on AI safety, potentially for lobbying efforts.

Fab, anything here that stood out for you or do you see this kind of as just another kind of industry group talking about AI safety?

Farb: I think as a smart move on, on their part, you can. You know, if you assume benevolence on their part, well then this is fantastic. This is them saying, Hey, we wanna be more open about it.

We're gonna be proactively sharing things and reaching out to the world as we take the lead on creating, you know, potentially a g i hyperscale scale, uh, foundational models. So if you assume they're being benevolent in general and aren't trying to build models to take over the world for themselves, then this is exactly what you'd wanna hear.

Uh, if you ex assume they're, you know, not being benevolent or being malevolent, then you're not gonna believe anything they say. So, you know, I, I think this is, this is the smart thing for them to do. I think it will garner more trust for them and will hopefully lead us to a better place that like they're describing.

Ethan: Yeah. Anything that stood out to you, Connor, in terms of what they're at least talking about?

Conner: Yeah, I kind of see it as like the UN Security Council for ai, so it's all these like competing companies working on ai, just like it was competing countries working on atomic weapons. And in the same way that for them, it protected the world from nuclear weapons and protected themselves individually from people saying they're not doing enough to protect the world.

This does the exact same thing. Like it works in two parts of it's kind of smoke and mirror. It's kind of just making it seem like they're trying to do something, but good things will likely come out of it also. So I do think it is really a great move in the end.

Farb: Yeah. The, uh, the United Nations of corporations, the, uh, yeah, techno few to list world I predicted is slowly becoming true.

Finally. It's coming true. It's coming true.

Ethan: Well, it's super cool. Um, our second story of today has, Would have a ton of applications to AI and the future of computing and everything. But this paper went viral, gosh, less than 24 hours ago, and this is LK-99. So this is a. What's being proposed as a room temperature, ambient pressure, superconductor, you know, this has been the kind of north star for anyone in the world of physics for a long time now, is how do we make superconductors that have these properties?

You know, superconductors have applications across computing and medical and power grids and. If this is true, this is a absolutely huge, huge accomplishment in the field of physics and will likely even win a Nobel Prize. So really cool work right now. I know we're working, uh, everyone's working right now to try to replicate these results.

Fab, you've been super into this. How do you feel?

Farb: Well, you know, to be clear, they've, you know, there are some holes that have already been found in the papers. Some weird discrepancies in some of the numbers. But if you're an optimistic person, you should be super excited about this and you should. Wanna see it become real and you should be jumping into help figure out if it is real or not.

Uh, that said, I v Ss spent a lot of the past 12 hours digging into this, coming up with, uh, recipes to do it. I've, I've, uh, got access to a sufficiently powerful vacuum through a friend's lab. We're now trying to identify what type of tube furnace one would need to try to. Synthesize some of this. It's supposed to be doable.

I think the whole world is a buzz with this right now, and that's one of the coolest things I've ever seen.

Ethan: Yeah. Connor, what about you? I, I, I saw on that manifold markets, you know, this kind of prediction marketplace puts it at 20%, I think, of, you know, being real within the next year. Anything you saw, Connor?

Conner: I think it's. Honestly, pretty likely. I think some of the holes about how much energy and how much wattage you can actually put through is a pretty big limitation. But Alex Kaplan from a friends over at Commandeer put together a pretty great thread on it, that this isn't like other superconductor papers that have come out in the past that are very like long shot and very like convoluted to set up and make, and everyone kind of knew we're fake.

This is a very simple paper, a very simple process, takes only a few days to make and. Semi available lab equipment, so. We will see in like a week if people can replicate it. But honestly, I think it's likely, I think it's something at least.

Ethan: I think it's really amazing that everything, you know, to make something like this, if this is the path forward for these types of superconductors, could have been made in, you know, 19 hundreds industrial equipment.

Um, so really cool to see stuff like that. But our third story of today is Text2Room, , similar to ish, what we talked about yesterday of kind of LLMs in 3D space, text to room is, hey, how can you, Create a text prompt and generate a 3D point space of a room. So they take images, they use a 2D model in between, similar to yesterday, kind of take pictures and build this three D point space of a room.

So another like really cool engineering way to tackle. How do we go from text-to-3D? Connor, anything that stood out to you?

Conner: Yeah, the output of this is kind of similar to Nerf except the big difference here. Instead of using a like neural radiance fear field, that's based on like light and how everything on Nerf Nerf works.

This is based on RGB, which makes it a lot more available because of course all images that are from text image or RGB, so this seems something more capable. This seems something that works a lot better and it's a big step forward in text-to-3D.

Ethan: Very cool. Farb?.

Farb: You know, I think we talk a lot about how some of these papers almost seem like engineering pipeline solutions as opposed to fundamental discoveries.

And uh, I wanna clarify that, you know, that is not to downplay it in any way. What is critical and what is, you know, spurring this sort of Cambrian explosion of AI outputs and, and tools and things that can actually do stuff, is the fact that. The underlying, you know, ML or calculations that you're doing to achieve these results are now doable by much wider groups of people.

There's some you guys are hearing, uh, an echo. I'm hearing a bit of an echo on my voice. Hopefully that's not coming through on the recording. Uh, The, the computation required to do some of these un the underlying calculations on some of these sort of engineering feats are sort of only really possible to the, the breadth of people who are doing this stuff, uh, recently.

So it really is this combination of, you know, incredibly powerful processors that are more affordable, more prevalent, doing some of the underlying, you know, machine learning and AI and processing so that someone can. Do some engineering to piece these things together and actually get a result where a small team can produce it, um, without taking a decade to do so, without having access to, you know, building size supercomputers.

So this is actually what we wanna see, a ton of endless amounts of engineering solutions that take, you know, what can be done with processing and, and, and really do cool outputs with it.

Ethan: I completely agree. Yes, super cool work by them. Definitely is gonna, you know, help paint the picture of text-to-3D even more so.

As always though, what else are we seeing? I got to see this bit tensor language model, which is a 3 billion parameter language model that's built to run on mobile and edge devices. You know, we've seen that llama's gonna try to work with Qualcomm to get that working, but. You know, these are the future of language models.

Not every language model, not every use case is gonna need to be run on a supercomputer in the cloud. So people working on getting the, working on getting the working, there's a great application of it, of it. I, I love stuff on edge computing. So excited for that. What about y'all?

Farb: I love that you've turned into Max Headroom.

That's the, that was wonderful. That's the coolest part. I dunno, your, your audio started like repeating itself, like you were, uh, nobody knows Max headroom. The, uh, I've seen, um, my, uh, my, uh, one of my investors and, and, and friends, the incomparable Paris Hilton, uh, posted a tweet asking about what car she should buy, and, uh, I replied to her with, don't buy a car, make your own car.

Uh, use AI to design a car. Use 3D printing to print the exterior and the interior and, and work with a EV platform like Rivian, uh, for example, Iona Tesla. Uh, but Rivian has like an EV platform that they're, that they're working on, uh, and, and put it on top of that. And so I used Mid Journey to generate some cool.

Paris Hilton looking style cars and I used runway ML to turn it into a little video. Uh, and then I used BER to sort of create this, uh, you know, backwards in time inspiration of, you know, the original types of cars and bugattis and types of vehicles that kind of inspired the, uh, final outfit. So it looks like the Hot Batmobile.

I loved it. That's awesome. Yeah, it looks like a, it looks like a hot Batmobile. And, and I got a, I got a, that's hot from Paris, which is. What everyone wants from Paris.

Conner: Beautiful. Yeah, I saw the GPU song, um, by weird ai Ya Chip. It's kind of a rip on. Uh, we didn't start the fire called GPU's Our Fire and that was basically the whole song.

Just ripping on everything in the world nowadays will link a blow. Pretty funny. That's

Ethan: beautiful. That's a beautiful song. Beautiful. Oh Dave, we'll see you again tomorrow.

Farb: We will see you tomorrow.

0 Comments
Authors
AI Daily