The AI Daily Brief: Artificial Intelligence News and Analysis - The US Wants To Block China's Access to Advanced AI Models

Episode Date: October 13, 2023

The US-China AI battle heats up as the Biden Administration looks into ways to deny China access to key advanced AI models. Before that on the Brief: an exciting new feature from Adobe and Google protects users from copyright complaints. TAKE OUR SURVEY ON EDUCATIONAL AND LEARNING RESOURCE CONTENT: https://bit.ly/aibreakdownsurvey ABOUT THE AI BREAKDOWN 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 Join the community: bit.ly/aibreakdown Learn more: http://breakdown.network/

Today on the AI breakdown, we're looking at reports that the U.S. is considering banning not only chip exports to China, but also Chinese access to important AI models. Before that on the brief, a very exciting new feature from Adobe and some legal indemnity from Google. The AI breakdown is a daily podcast and video about the most important news and discussions in AI. Go to Breakdown.network for more information about our Discord, our YouTube channel, and our newsletter. Welcome back to the AI breakdown brief. All the AI headline news you need in around 5. minutes. We kick off today with a very exciting feature that was just premiered at an Adobe event this week. Of course, at this stage, some of the most exciting advancements in the AI space this year

have been around text image generation. We've seen increasingly advanced models, Dolly 3, integrated into chat GPT, and of course we've had features like Adobe's generative fill, which give creators more control than ever before. Generative fill is a feature in which you can select a very specific part of an image and change it using the same sort of natural language inputs imprompting that you would create a whole image from the ground up with. Now, this feature is no longer just a part of Adobe. You see something similar in Canvas. Mid Journey has adapted it as well, but at this week's Adobe Max event, they actually premiered something they're calling FastFill, which is the same type of feature but for video. So in the example video they showed, they had

a small clip of a man in a suit walking through a city scene, and were able to select and change his tie from not having any tie to having a blue tie, to having a different color tie. And again, this wasn't just a static image, but a moving video. Another example they gave was taking a group of background joggers out of a video whose subject was a woman in the foreground. Now, unlike some of the other things that Adobe has recently announced like their Image Model 2 earlier this week, this is unfortunately still very much in the experimental category. Users can't get their hands on this service just yet.

But it shows how quickly the text-to-image capacities that were quickly getting used to and developing new workflows around are also coming to video. The average response I saw on Twitter slash X were phrases like, next level and holy moly. Now, speaking of the world of images and video, outside of features, a lot of the discussion recently has been around copyright claims. Of course, the data used to train different models has remained largely a black box, and there are several lawsuits going on across a variety of different services from image creation tools to things like ChatGPT, where content creators and copyright owners are claiming that the use of their works to train AI

models somehow violates their copyright. Now, this is a fairly important issue, because one of the big areas of potential adoption for AI is, of course, in the enterprise. However, legal and compliance departments have a strong issue with using tools that potentially open companies up to lawsuits. There are a number of different ways that companies are trying to address that. One of them is training models on proprietary datasets. Getty images has done this in the image space, and Adobe's Image 2 model also promises to be only trained on the set of stock images that they actually own the copyrights to. That's sort of step one in making them a little bit more safe for enterprise use. Step two, however, is taking the proactive step of guaranteeing that if a user

finds itself in a lawsuit based on those copyright claims, that the company will take on their legal defense. Adobe was first to market with that pledge earlier this year, and was followed quickly by Microsoft doing something similar, and now Google has followed them, pledging to defend people who are using their generative AI tools and workspace against this sort of copyright claim. Now, so far it appears that Google's indemnity program focuses on their vertex AI development platform and their duet AI system, which generate text and images in Google Workspace and other cloud programs. Their press release didn't mention the better-known generative program barred, but it wouldn't be surprising to see that as well. To me, this just underscores that the AI market is so valuable

right now that even in this time of a lack of clarity around things like copyright, the big tech competitors are willing to put their money where their mouth is when it comes to legal defense, just to try to continue to move people onto their platforms. Speaking of Google and speaking of that AI arms race, now that both chat GPT and Bing have integrated image generation, Google search generative experience or SGE is getting a similar capability. It's a great example of how something that didn't exist a couple months ago has now become absolute table stakes for people who are in this space. Now, on this theme of enterprise use cases, Gartner has just put out a new report that suggests major, major growth in enterprise use of AI over the next few years. Right now, Gartner says

that around 5% of enterprises are formally using AI tools. In other words, they've made it a specific initiative or they've done some specific procurement process. That's, of course, different than just individual employees of companies who are using those tools, where the percentages would be much higher. However, according to this report, they anticipate that by 2026, a full 80% of enterprises will actually be using things like AI APIs or even spinning up their own models. Over in the world of policy, even as some number of senators and Congresspeople are working on comprehensive AI legislation, we're also seeing an increasing number of specific discrete AI-related acts that try to address or create rules around some specific issue. A great example of this is a new proposed

rule called the No Fakes Act. Now, technically, this is called the Nurture Originals Foster Art and Keep Entertainment Safe Act of 2023, but of course, this being Congress, it has to be acute acronym, no fakes. The bill prevents the, quote, production of a digital replica without consent of the applicable individual or rights holder. And basically what this is going for is that if an artist or a musician comes across someone who's creating a digital artificial facsimile of them, such as, for example, using an artist's voice for an unauthorized song, this bill would give that artist civil recourse. Now, one of the reasons that they've focused on this right now is that a lot of the likeness laws which might be extended to cover AI-generated versions are sort of state-by-state and fragmented. This would effectively federalize those laws, which obviously makes it much easier to enforce. The Recording Industry Association of America, who we talked about yesterday as trying to get a voice cloning site on a privacy watch list in the U.S., said that they welcomed the bill.

Others, however, worry about rushing in too fast. Dwayne Morris' partner Jeremy Elman said, Regulating AI is certainly at the top of the list for lawmakers these days, but they should be careful not to rush into creating a new federal IP rights that may conflict with longstanding balances in the IP system. Moving to our next topic, we've talked a lot about how assertive the U.S. military establishment is being in adopting new AI tools, and in particular in building their own. One interesting story then is that the U.S. Space Force has paused the use of all AI tools, including ChatGPT, because of concerns around data security. Reuters got access to a September 29th memo addressed to all Space Force employees

that prohibits them from using these types of tools on government computers unless they receive formal approval. The ban was said to be temporary and was, quote, due to data aggregation risks. Now, of course, this is the type of thing we've seen from a lot of private companies as well, who are concerned with data leakage, and the foundation model companies accidentally getting access to really proprietary and sensitive data, which of course when you're dealing with military applications is an even more heightened concern. Now, it's probably notable that we're also getting lots of information about, for example, the U.S. intelligence establishment training their own foundation models, which obviously are going to come with a higher degree of control than they would

just tapping into something like ChatGPT. And lastly today, one follow-up from our story yesterday about Character AI. As I mentioned, I don't feel like I have a lot of visibility into who's using it. And so because of that, I'm actually paying even more attention to trying to understand who is using it and how and what the implications might be. Well, right after I published that episode, Going Godward on Twitter wrote, Y'all, oh my word, it's bleak. My students are becoming addicted to character AI and are abandoning real-life interactions and friendships for conversation and pseudo-relations with AI. I can see how social media has been the gateway to fully integrating the reel with

the transhuman. These children began interacting with AI from a young age, so it feels natural for them to experience AI as the friend they've always wanted. It's much more exciting and much easier to be friends with AI LeBron James than to develop the kind of character that makes you more real and more into the kind of person who's a good friend. Every new AI technology seems to make us more like a machine and less like ourselves. Too many negative consequences to name. Now, of course, I know nothing about this user. I don't know what they actually do. This could be entirely a fake post. And even if not, there's always going to be this type of response to the advent of a new technology. Previous generations almost see what is lost more than what is gained when it comes

to a fundamentally new behavior pattern. That said, it's not hard to imagine how this could really ring true, especially among young people whose social systems are just still forming. And so I share it with you as one piece of evidence of the ongoing changes around us, wrought by generative AI, make of that what you will. That will do it for today's AI breakdown brief. Thanks as always for listening or watching. Next up, the main AI breakdown. Welcome back to the AI breakdown. Today we are once again looking at the geopolitics of AI. And this is more than just some wonky topic for people who are inside the Washington DC Beltway. The reason that this matters, even if you are just a user of AI tools, is that one of the big considerations shaping the way that U.S.

policy towards artificial intelligence gets made is the nature of our relationship with China. What happens vis-a-vis that relationship could impact what tools you have access to just a year or two down the line. And that, I think, is the context for new reporting from the Atlantic that in addition to AI chip restrictions, the Biden administration is also considering how they might prohibit access to more advanced AI models as well. In a piece called The New AI Panic, the Atlantic writes, For decades, the Department of Commerce has maintained a little-known list of technologies

that, on the grounds of national security, are prohibited from being sold freely to foreign countries. Any company that wants to sell such a technology overseas must apply for permission, giving the department oversight and control over what is being exported and to whom. The battle lines may soon extend beyond chips. Commerce is considering a new blockade on a broad category of general-purpose AI programs, not just physical parts according to people familiar with the matter. Although much remains to be seen about how the controls would roll out, and indeed,

whether they will ultimately roll out at all, experts described alarming stakes. If enacted, the limits could generate more friction with China while weakening the foundations of AI innovation in the U.S. Now, from there, the piece describes how, while the AI chip export restriction might be focused on a concern that AI is going to be used to develop advanced autonomous military systems in China, there is a growing second consideration now of how access to frontier models from companies like Anthropic and Open AI could benefit China in other more nebulous ways. The piece discusses how a number of influential white papers have helped shape Washington's thinking in this area and is driving it closer and closer to a model of regulation and export restriction

that while potentially denying China access also creates fairly severe regulatory capture for the organizations at the top of the food chain. writes the Atlantic, the obsession with frontier models has now collided with mounting panic about China, fully intertwining ideas for the model's regulations with national security concerns. Over the past few months, members of commerce have met with experts to hash out what controlling frontier models could look like, and whether it would be feasible to keep them out of reach of Beijing. I think really importantly, the piece identifies what they call a, quote, precarious dynamic playing out in Washington. Basically, they point out that the AI panic has made policymakers, quote,

uniquely receptive to the tech industry's messaging. Combined with concerns about China and the upcoming election, it's engendering new and confused policy thinking about how exactly to frame and address the AI regulatory problem. Said Emily Weinstein, a research fellow at Georgetown Center for Security and emerging technology, parts of the administration are grasping onto whatever they can because they want to do something. So one important aspect of this piece is how they identify the specific dynamics in Washington that are leading the leading edge of the AI tech industry to be in sync with policymakers, potentially at the expense of smaller companies. A second thing articulated in the piece is the likelihood that expanding restrictions beyond just chips would further exacerbate tensions with China. However,

there's a third piece, which if you listen to this show regularly, you will know is a growing topic of conversation. As Matt Shaheen, a Carnegie Endowment for International Peace Fellow put it, if the export controls are broadly defined to include open source, that would touch on a third rail issue. In a really important section, the Atlantic writes, what's frequently left out of considerations as well is how much this collaboration happens across borders in ways that strengthen rather than detract from American AI leadership. As the two countries that produce the most AI researchers and research in the world,

the U.S. and China are each other's number one collaborator in the technologies department. They have riffed off each other's work to advance the field and a wide array of applications far faster than either one would alone. Part of the reason that I think is important to share this is that it's one of the first times I've seen in mainstream media a counterweight to just either the A, full AI safety concern narrative that is getting more and more play, and B, frames the question of open source in terms of value to Americans not just cost, and proliferation towards China. Now, of course, the piece also recognizes that the feasibility of export controls of frontier models could be really, really challenging.

Now, putting a fine point on this difficulty is another article this time from Bloomberg, from about a week and a half ago, where U.S. Commerce Secretary Gina Romando showed just how challenging the AI chip export bands really, are. At a Senate Commerce Committee hearing, Remando basically pointed out that Chinese companies are continuing to advance in their own chips, a development which she calls, quote, incredibly disturbing, and which she suggested means that her department needed more ways to enforce the administration's export controls. Now, among other things, she's asking for expanded authority over technology transactions, which is, I think, exemplary of how this type of issue can lead pretty cleanly to an expansion of government power. Now, just today, we also got news from Reuters

that the current administration is continuing to expand its thinking about how it blocks China getting access to AI chips. Basically, they're trying to close loopholes, through which Chinese companies can get access to chips through partners and other international locations. Wrights Reuters, in the initial round of curbs, the Biden administration left overseas subsidiaries of Chinese companies with unfettered access to the same semiconductors, meaning they could easily be smuggled into China or accessed remotely by China-based users. So now they're basically trying to figure out how they can close that loophole. Now, this follows news last month that the U.S. had also been dealing directly with companies like NVIDIA to restrict chip sales to parts of the Middle East.

That wasn't about a concern that Saudi Arabia would have access to advanced compute, but that those chips would ultimately end up back in the Chinese sphere of influence. What's clear is that the White House is coming closer to announcing some AI policies. Politico also reported today that the White House executive order on AI is getting closer than ever and started to give a few details of what that might include. For example, the White House is expected to introduce guidelines for testing and evaluating AI systems, which could of course be part and parcel of a regime through which certain types of models are not allowed to be used or sold to Chinese companies, and the order is also expected to

require cloud computing companies to track their customers who they think are likely to be developing AI systems. That executive order is expected to come out before the end of the month. So what we've got now, just to sum up, is a United States that is thinking aggressively about China as it creates AI policy, in ways which may tip the scales towards the incumbents, or at least the startup market leaders as opposed to new companies, and which may place serious restrictions on things like open source AI development, and of course, all the while, you have China proceeding ahead as fast as it can as well. This may be less interesting than new features and chat JBT4 vision, but it's going to have an extraordinary impact on the shape of the field in the years to come.

And so I hope this gave you a little bit more of a sense of where we are and what's coming down the pipeline. I appreciate you guys listening as always, and until next time, peace.