The AI Daily Brief: Artificial Intelligence News and Analysis - Are OpenAI Trying for Regulatory Capture?

Episode Date: May 20, 2023

On this edition of The AI Breakdown weekly recap, NLW looks at all the AI news, including Mind-blowing DragGAN photo editing research Blockade Labs Skybox text-to-3D world StabilityAI releases open source StableStudio NYT open source Meta article StabilityAI letter on open source to the Senate Was Sam Altman's testimony just regulatory capture?   The AI Breakdown helps you understand the most important news and discussions in AI.  Subscribe to The AI Breakdown newsletter: https://theaibreakdown.beehiiv.com/subscribe Subscribe to The AI Breakdown on YouTube: https://www.youtube.com/@TheAIBreakdown

On this episode of the AI breakdown's weekly recap, we talk about a mind-bending photo editing tool, a big new open source move from stability AI, and ask whether Sam Altman's testimony was just about regulatory capture. The AI breakdown is a daily video and podcast about the most important news and discussions in AI. If you're enjoying it, please like, subscribe, and share. What's going on guys? Welcome back to the AI breakdown. This is the weekly recap where I talk about some of the most important news that happened this week. We start with a really mind-blowing new photo editing tool that's in the research stage. And basically, this video started exploding all over Twitter on Friday morning. For people who are listening, you're seeing images that someone is clicking on a particular part. So it could be the face of a lion or it could be the face of a dog and actually manipulating them simply by dragging them.

It's photorealistic image manipulation all through AI. Beulawal Sid who writes, whoa, this new AI paper is legit like a dog. though be puppet warp on steroids. Text prompts aren't the be-all-end-all of AI creation. Drag-GAN is a perfect example of giving creators fine-grained control over the AI image generation process. Drag-Gan allows you to reshape images interactively, much like pulling on points in the image to move them exactly where you want.

The system relies on two main parts, one that guides the point you're moving, the handle point towards its destination, and the other that keeps track of the handle point's position using features specific to the underlying technology, GANS. Drag-Gan is super flexible. You can tweak various aspects of your images, such as pose, shape, and expression. It works on a wide range of subjects, including animals, cars, people's, landscapes, etc. Now, in an example of using AI to understand AI, I went to ChatGPT and took advantage of a new plug-in called X Papers.

I asked Chatchapet, do you have access to this paper? And it said, yes, I have access to the paper you're asking about. It's called Drag Your Gan, Interactive Point-Based Manipulation on the Generative Image Manifold. I said, can you please summarize the paper for an audience at a high school graduate level? in no more than five bullets, making sure to explain both generative adversarial networks and drag GAN. Chat GAPT did exactly that. Generative adversarial networks are GANS are a type of artificial intelligence model that can create new data that resembles the original data it was trained on. Imagine it has two artists, the generator and the discriminator, competing against each other.

The generator tries to create convincing fake paintings, while the discriminator tries to tell if the paintings are real or fake. Over time, both get better. The generator at creating fakes and the discriminator at spotting them. Number two, the problem, when we use GANS to create images, we often want to control specific details like the poser shape of an object in the image. But existing methods to control GANS aren't very flexible or precise. Drag GAN is the solution. This is a new process that allows users to drag points on an image to change it in specific ways. For example, changing the pose of a person. Now, importantly, I was also able to ask, can it be used to modify real world images,

or is it only useful for images that have been created with AI? ChatGPT says drag GAN can indeed be used to modify real world images, but it requires an additional step. This step is called GAN inversion. GAN inversion is a process where a real-world image is transferred into the latent space of the GAN. So while drag GAN is primarily designed to work with images generated by a GAN, it can also be used with real-world images through the process of GAN inversion. Pretty Wild stuff all around. Speaking of Pretty Wild, this is a demo from Blockade Labs of their new Skybox tool, which allows you to effectively sketch a 3D world that you'd like to see and give it a little bit of a description and then have it appear exactly as you'd

imagined. I am seeing such an explosion of 3D world building tools. I am really, really excited to see what comes out of it from the standpoint of games and Metaverse and applications we can't even imagine yet. Now, one other product release that I thought was really interesting was Stability AI announcing Stable Studio. Stability AI is, of course, the company behind Stable Diffusion, and you might have used their tool Dream Studio, which is their main web-based interface for interacting with their text to image tools. Stable Studio is effectively an open source version of Dream Studio that encourages people to actually build out these tools in an open source way. At the heart of it is their AI image generator, but they're also bringing in all of their other open source tools as well,

including their language model, stable LMUNA, as well as soon their stable vacuna chat interface for StableLM. In the Stable Studio press release, the company wrote, we believe the best way to expand is through open community-driven development, rather than private iteration on a closed source product. The end goal, they say, is to create an AI interface for users to, quote, fully control. Now, stability AI will continue to build out Dream Studio, but that will effectively be their own internal implementation of Dream Studio, which anyone can now build upon.

The Open Source discourse also got a boost from a big New York Times article this week that focused in on Meta and Jan Lacoon. Lacoon, who is the chief AI scientist that Meta wrote on Twitter, a New York Times article on the debate around whether LLM-based models should be closed or open. Meta argues for openness starting with the release of Lama for non-commercial use, while OpenAI and Google want to keep things closed and proprietary. They argue that openness can be dangerous, but they are just protecting their commercial interests. I argue that closeness is considerably more dangerous than openness.

Once LLMs become the main channel through which everyone accesses information, people and governments will demand that it be open and transparent. Basic infrastructure must be open. Now, this is all the more interesting in light of the Senate hearing on AI earlier this week. Leading into that, Stability AI had written a letter to senators advocating something pretty similar. The letter goes into some detail about the importance of open models, but is summed up in a quote from Imod, the CEO of stability. He writes, these technologies will be the backbone of our digital economy, and it is essential that the public can scrutinize their development. Open models and open datasets will help to improve safety through transparency, foster competition,

and ensure the United States retain strategic leadership in critical AI capabilities. Grassroots innovation is America's greatest asset, and open models will help put these tools in the hands of workers and firms across the country. I did an entire show about the hearing, but one of the notable features of it was that, whereas there's often a lot of acrimony between senators or congresspeople and their witnesses, particularly if those witnesses come from big tech companies, that didn't seem to be on display in this particular case. Part of that seems to have been the fact that the witness that they were most interested in talking to, which was undoubtedly Sam Altman, the CEO of OpenAI, seemed to be broadly in agreement with them that there needed to be a new regulatory apparatus for AI, going so far as to even say that he would support AI licenses. Here's Senator Lindsay Graham in a key section of that hearing.

Mr. Allman, why are you so willing to have an agency? Senator, we've been clear about what we think the upsides are, and I think you can see from users how much they enjoy and how much value they're getting out of it, but we've also been. been clear about what the downsides are. And so that's why we think we need an agency. It's a major tool to be used by a lot of people. It's a major new technology. We think it'll be. Yeah, if you make a ladder and the ladder doesn't work,

you ensue the people made the letter, but there's some standards out there to make a letter. That's why we're agreeing with you. Yeah, that's right. I think you're on the right track. So here's what my two cents worth for the committee is that we need to empower an agency that issues in a license and can take it away.

Wouldn't that be some incentive to do it right if you could actually be taken out of business? Clearly that should be part of what an agency can do. Now, to get a sense of how many people reacted to this part of Sam's testimony, just look at gphoder.id on Twitter who wrote, Sam proposing licenses for AI training is the most awful thing I've ever heard him say. How disappointing say it ain't so. Antonio Garcia-Martinez writes, I love how quickly we went. from a promising prototype to hysterical clout-chasing opportunists issuing dire warnings to craven corporate regulatory capture in the span of months.

And indeed, this theme of regulatory capture was a huge one. Scott Galloway says, we're falling for this shit again. Altman's CEO of OpenAI calls for U.S. to regulate artificial intelligence. Brad Hatman responded, summing up what Galloway was trying to say, writing, tech execs say, we need more regulations, tech execs think, bigger motes equal bigger rents. denying Sam's sincerity in his beliefs of the risks of AI, at Jay River Long, who has the effective accelerationist tag in his profile said, Sam Altman going in front of Congress to demand AI regulation and disclosure is regulatory capture plain and simple. He thinks he has the business

of the century and wants to ban competition. Sam is dangerous and the AI safety crowd are useful idiots to his monopolist goals. Now, not everyone thought this way. Matthew Barnett, for example, writes, I think regulatory capture explanations are overrated. Sam Altman presumably wants to be seen as a responsible CEO while minimizing the impact of regulations on open AI. This theory explains our observations just as well. Now, this all got loud enough that Sam decided that he wanted to respond. Former open AIer Alethea Power tweeted a clip of Sam and said, I think people talking about regulatory capture missed the part where Sam said that regulation should be stricter on orgs that are training larger models with more compute like OpenAI,

while remaining flexible enough for startups and independent researchers to flourish. Sam, himself quote tweeted that and said, Regulation should take effect above a capability threshold. AGI safety is really important and frontier models should be regulated. Regulatory capture is bad and we shouldn't mess with models below the threshold. Open source models and small startups are obviously important. So I think there are three possibilities here. One is that Sam's being sincere.

He's not trying to use regulatory capture as a strategy. He's not hoping that onerous or burdensome regulations will crowd out competitors because they don't have the resources to comply. He is just genuinely worried about what could happen if AI isn't regulated at all. A second possibility is that he is going for regulatory capture for what he views as good reasons, as in he genuinely believes that there could be harm, and he wants only a small number of companies, including his, to be able to actually have the power to cause that harm or not.

A third possibility is regulatory capture for bad reasons, as in he's insincere about these concerns, and he just wants a position OpenAI to be the leader of the next wave. In a lot of ways, I don't think it particularly matters. In that I don't think we should be making decisions about policies in terms of what Sam does or doesn't want or what he does or doesn't think. We need to consider the issues of regulatory capture carefully, whether Sam intends for Open AI to be a beneficiary of them or not. Regulations do increase the cost of compliance. They do create moats for incumbents.

And there's a very real chance that we prohibit or license the wrong thing. Especially with such a frontier technology, there are real reasons to be concerned about that. At the same time, outside of any consideration of open AI, there are good reasons to have this conversation about whether we should have guardrails and what they should be. I think it's clear that what needs to happen next is specificity in the conversation. Right now, we're speaking in super vague generalities, and that's creating a scenario where people are mapping on their priors and their expectations and their beliefs about people without actually engaging with specific policy proposals, which makes sense because they don't exist yet.

So for now, I'm reserving my judgment because, like I said, I think the conversation is important, regardless of whatever Sam or Open AI think, and I think it's the right conversation to be having right now. Anyways, guys, if you ask me, that is a very characteristic week in AI right now. On the one end of the spectrum, some mind-blowing tools that you couldn't have imagined just about five minutes ago.

And on the other end of the questions, big existential questions that could change the shape of public institutions. If you're enjoying the AI breakdown, please like, subscribe and share. Go follow the podcast and follow the newsletter. And as always, I appreciate you listening and watching. Until next time, peace.