DisCo | OpenAI Superalignment | ChatLaw

AI Daily | 7.5.23

Welcome back to AI Daily! We kickstart the conversation with DisCo, a revolutionary project for real world dance generation that's reshaping how we understand motion capture and dance generation. We follow that with a discussion on superalignment from OpenAI, a cutting-edge project designed to align super-intelligence with human interests.Finally, we turn our attention to ChatLaw, an open-source legal language model that could redefine legal discourse.

Key Points

1️⃣ DisCo

  • DisCo, a collaborative AI project between NA Yang Technological University and Microsoft Azure, generates realistic human dance movements from photos.

  • Despite some initial artifacts, the AI can generate natural and high-quality movements, promising photorealistic results within a year.

  • With its ability to realistically simulate dance, DisCo has the potential to dominate social media content creation.

2️⃣ OpenAI Superalignment

  • OpenAI has formed a new alignment team to solve "superalignment", dedicating 20% of their resources to aligning super intelligence and preventing potential threats.

  • This approach acknowledges that aligning super-intelligence is both a philosophical and a technical problem, requiring significant investment and a dedicated team.

  • The initiative signifies the importance of AI alignment, suggesting a future where AI systems compete, with the winners determining the narrative.

3️⃣ ChatLaw

  1. ChatLaw is an open-source large language model with integrated external knowledge bases, fine-tuned for Chinese legal data, aiming to tackle issues of AI hallucinations in legal contexts.

  2. The team found that relying on a Vector DB alone isn't sufficient to meet the exacting standards of law and could lead to the production of false information.

  3. The model showcases the breadth of AI, with solutions tailored for specific applications like Chinese legal data, contributing to a reduction in hallucinations.

🔗 Episode Links

Follow us on Twitter:

Subscribe to our Substack:


Conner: Hello, and welcome back to another episode of AI Daily. I'm your host Conner, joined by Ethan and Farb, and we have another three great stories for you guys coming today. First up, we have DisCo. This is Disentangled control for referring human dance generation in the real world. It's a collaboration between NA Yang Technological University and Microsoft Azure ai.

Basically, you can upload a photo of any person in a foreground, any background. And then any target pose to move their body, um, usually generated by something like Open Pose and then it can generate a whole new frame of them in an entirely different position. And they showed some pretty cool demos where if you string this together, you can get an entire like TikTok dance of any person doing any type of dance.

Farb, this looks highly realistic. This looks really good, in my opinion. What do you think about it?

Farb: It's really fun to check out. Uh, they have a little demo that doesn't have the video working yet in the demo. You know, I, I thought their v videos were, they certainly weren't convincingly that person dancing.

A lot of artifacts, at least from what I saw in less than a year from now, this will be photorealistic video realistic. I have almost zero doubt about that. It may be in a month, quite frankly. It's not quite there yet, but it's super impressive and I, like I said, I think this sort of thing is gonna be.

Dominating social media in not too long. So these, these may be the folks that are the ones, we'll see if anyone else can catch up to them and we'll see how quickly they get ahead.

Conner: Honestly, it reminds me of early stable effusion. Yes. If you blow it up, yes, if you really look at it, there are some artifacts, but if it's small on your screen, if the video's kind of grainy, you can't see it that well.

It looks realistic. The actual like movements themselves.

Farb: Yeah, the movements are very realistic, natural looking.

Conner: Very natural. Very, very high quality. Ethan, what do you think about it?

Ethan: Yeah, poses and compositionally is, you know, very difficult in this domain. So I think far nailed it in terms of, at the end of the day, we're gonna see these artifacts get removed in the future, and we're gonna see the actual video be of higher quality.

But in terms of being able to estimate poses, it nails it. So really cool to see.

Conner: Yeah. And of course if we think back to like when like the Pope and the white popper went viral, it's little, little artifacts here and there don't really matter if people aren't looking for them. And the overall picture looks realistic, so, yeah.

Yeah, very exciting though. Disco, and we'll see where it goes. We'll see what disco two rings us. So next up we have, uh, introducing super alignment from OpenAI. Uh, a new alignment team from OpenAI, led by Ilia Ver and Jan Leika. Uh, open AI themselves are dedicating 20% to try to solve super alignment.

They're comparing normal alignment to normal AI intelligence, and now super alignment will be their solution to super intelligence and aligning super intelligence. They're very concerned about it. In their intro, they said, uh, the vast power of super intelligence could also be very dangerous and could lead to the disempowerment of humanity or even human extinction.

So they're sounding like they're taking this pretty seriously. Ethan, what do you think about it?

Ethan: Yeah, you know, the goal to use AI to help in AI alignment is fascinating. You know, at the end of the day, they just, like you said, from a philosophical angle, they believe, and I think most people deeply in the space believe that, hey, we need to align super intelligence when it comes and we want to align our current models too.

So they're putting some real weight behind it to build this kind of automated alignment. Researcher, as they call it, 20% of their compute is nothing to scoff at. They're gonna build an amazing team around it. And yeah, it's a philosophical problem. It's a technical problem and we'll see how they solve it.

Conner: Yeah. They said they believe super intelligence could arrive within the decade, uh, far. What do you think about this prediction? What's your opinion here?

Farb: This is a really difficult needle to thread, and I think they're doing a, a about as good a job as you're gonna do. There is the rhetoric that you have out there in the world, there is how you actually think about it.

They're never quite the same. Uh, and then there's actually what you can do about it. Which is different from both the rhetoric and what you even think about it. It is not an easy situation to navigate, and they are doing it. Uh, I, I think real really effectively. They're being honest, you know, about what their plans are, what they're thinking, uh, how they wanna approach it.

They are taking it seriously. They're giving real compute power. 20% of your compute power is an insane amount of compute power to, um, devote to any anything. So they're really showing that they care about this. If anyone's gonna move the needle here, they're probably a pretty good candid to do it. Uh, I wouldn't be surprised if we see lots of other people, uh, getting into it and I, I think OpenAI would, would happily welcome that.

So good luck to them. Good luck to all of us. You know, this is the sort of thing where it's not an all or none thing, although maybe people wanna play as though it is that way. You're going to make improvements, you're going to provide some rails and. Just like things can go wrong with any, uh, technology that is potentially damaging, uh, you're never gonna mitigate a hundred percent of all potential issues here, but you probably can mitigate a lot of them.

So you should, and that looks like the approach they're taking.

Conner: Yeah. On, on a more technical side, um, of course ChatGPT itself was. Fine tuned through reinforcement learning, human feedback, but it seemed like they found that I couldn't really scale to alignment in how they wanted it. Yeah. Ethan. Ethan, how does this compare to what we've seen with Anthropic and their constitutional ai and how does that different, and is this leaning more towards that?

Ethan: I think it's just a call out that, hey, constitutional AI might not work. And we're still trying to figure out all the technical possibilities here, all the possible scientific solutions. So they're just saying, Hey, we're gonna dedicate some real team and compute behind this and we're not sure what it is yet, but we're gonna go figure it out.

Um, you know, some of the other players might already have constitutional AI or want to try to scale human feedback, but they're saying, you know, honestly, we're not sure yet and we're gonna go figure it out.

Conner: I think it is a lofty goal and I'm excited to see what comes out of it.

Farb: Absolutely the battle of Theis is coming.

There will be the good AIs fighting against the bad AIs, and hopefully the good ones are better than the bad ones.

Conner: We hope. And of course, the winning AI will write the history book saying that it was the good ai. So absolutely. Last of today, we have ChatLaw, which is an open source legal, large language model with integrated external knowledge bases.

So a little bit similar to BloombergGPT and FinGPT that we've seen and talked about before. This is a legal l l m fine tuned on legal data. Um, but it has, it's mostly fine tuned for Chinese legal data because the entire team, of course, is Chinese and legal models so far don't really have a good knowledge of that far.

What do you think about this? They have some interesting new in integrated external knowledge bases. What do you think about that?

Farb: I, it sounds like one of the major things that they discovered is you can't just, you know, rely on a Vector DB to pull this off. You're just gonna be hallucinating your way all the way to jail.

Like that poor guy who, uh, used some, uh, AI based stuff in court. I don't think he went to jail, but I think he got fined and I think, you know, he got relat pretty well admonished by the judge for bringing in a bunch of false information that he generated from Chad G p T. This is actually a really difficult problem and obviously the standards for law are.

You know, pretty exacting. It's not quite mathematics, but it's certainly not an area where hallucination, uh, is gonna be. Okay. So how do you get the hallucination out? Well, maybe you, you know, do keyword, uh, search as, uh, while using your Vector DB to get some improvements there. This is the type of thing where you're probably gonna have to cobble together all sorts of solutions to get to something viable, you know, may include humans in the loop as well.

And, you know, Probably the super intelligent AI will do a better job of it, but it'll be a while to there. But I think we'll have this solved before then.

Conner: Yeah, law is a very interesting domain. Um, ChatLaw is open source, which is nice to see. Ethan, what do you think, how do you think vector databases and how do you think, uh, keywords search?

How do you think this compares versus just a normal BloombergGPT?

Ethan: I mean, I think they showed that it was more effective. So that's a really cool way of attacking some of these hallucination problems, like you both mentioned. And secondarily, I think it just speaks to the breadth of this space. You know, we're seeing a Chinese legal open source, L L M, that's something that doesn't fit into open AI or Claude, or any of the other big players.

It's completely different kind of language, completely different fine tuning data set and the pieces around it to reduce some of the halluc hallucinations. So I think it speaks to how big the space is as well.

Conner: It's very exciting. Well, those are three stories. Today we covered disco and then we covered OpenAI.

Super alignment and then lasting. Now of course, we just covered chat law, so very exciting. What have you guys been seen far? What have you seen lately? I've, I'm seeing this. Ah, oh, is that an ai? Is that an AI cap?

Farb: It's fully AI. You might be the ai, super intelligence. We've all been waiting for, uh, I thought Midjourney dropped, panning.

You know, they did Zoom a little while ago, like last week probably. I say a little while ago. It was probably three days ago. Uh, they just dropped panning. I went a little ham on Midjourney yesterday, making all sorts of cool July 4th, uh, photos, you know, kids on Mars celebrating July 4th. All these. The images that it created from this set of, you know, I was asking it to create people celebrating July 4th at block parties in different parts of the United States.

And, uh, they are shockingly realistic. Um, I was kind of blown away by it and I didn't spend a ton of time dialing them in. Maybe it was, you know, 30 minutes or something. And I, I guess I'm, I'm probably more used to writing the, the prompts than, than most people are, but, you know, the, the stuff that comes right out of the gate on Mid Journey is pretty fantastic.

I. I highly recommend checking it out. If you're into the space, you just gotta make a little account. It's not too expensive. It does cost some money if you want some of the faster, uh, services and, and things like that, but they keep adding stuff to it and, uh, you know, they're printing money over there.

They may be also, they may be also lighting it on fire on their, for their GPUs, but most likely, yeah.

Conner: I, I'm pretty sure they use TPUs actually. But besides that, sorry. Yeah, TPUs. Yep. Yeah. Be beautiful images, beautiful thread, um, far. Will link your July 4th thread below. Everyone should go check it out.

Ethan, what about you? What have you seen?

Ethan: Uh, I saw that playground raised a new round 40 million. And you know, to me their blog post they put out was actually really, really interesting around this. Just speaking on the state of computer graphics, you know, Photoshop and most ways you interact with images or hand-built human algorithms and the space of AI affecting that.

How you deal with pixels, how you build images, how you do spatial relationships. Really cool blog posts. So we'll link it below and congrats to 'em for raising round.

Conner: Mm-hmm. Yeah, I saw a very interesting paper out of Wuhan University. I think we've all heard of Wuhan before so far from the past few years.

Yep. Um, but their paper they released was BatGPT. Uh, I think we all get the connection there between bats and Wuhan diseases.

Farb: They're just full sending it here. Do you, do you know something that we don't know? Uh, you've got definitive proof of, uh, what went on over there. I mean, we'd love to see it.

I'm sure there's lots of folks you could call up. We get R FK Jr. On the next episode. You and him can, you and him can figure out all the problems in the world in probably 15 minutes.

Conner: I have no information on this, but the Wuhan University team at least thought it was funny. It works. They got their, they got their, I got a sense of humor about it.

Farb: We're, oh, I can't wait for the comments on today's episode. It's gonna be some good stuff here. And, uh, all set you guys in the comment section. I'm,

Conner: I'm not some heavier, I'm not sure how good the mo the model is. Uh, all the benchmarks they measured it on were Chinese benchmarks. But interesting trick they did use, they reversed the sequence, uh, order during training.

So they got it to predict both the next token and the previous token. Oh wow. So they flipped the entire data set, both in and out. So it can predict both ways. Not sure if that actually helps on anything cuz you want a model to predict the next token or not the bravest token, but maybe it did. Either way bad g Bt we're talking about it.

Well done, Wuhan University.

Farb: That's clever. That's clever.

Conner: Indeed, indeed. Well, thank you guys for watching today. Another great show. We will see everyone tomorrow and don't forget to like and subscribe.

Farb: Hopefully we'll see you tomorrow. We may all disappear between now and then. Thanks to thanks. That last story.

Good luck everybody. If you, if we never see you again, we loved you and it was a great run. Thanks for joining us.

Conner: Have a good one.

AI Daily