The AI Daily Brief: Artificial Intelligence News and Analysis - White House AI Executive Order Coming Monday

Episode Date: October 26, 2023

The Biden Admin's long promised AI executive order is coming Monday and it seems like it will use the power of government procurement to try and set new standards around testing and model evaluation. Before that on the Brief: AI and ads plus updated knowledge for ChatGPT. Today's Sponsors: Listen to the chart-topping podcast 'web3 with a16z crypto' wherever you get your podcasts or here: https://link.chtbl.com/xz5kFVEK?sid=AIBreakdown  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 new reports that the White House executive order on AI is coming next Monday. Before then on the brief, AI for ads, and a new updated knowledge cut off for chat GPT. 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 YouTube, our Discord, and our newsletter. Welcome back to the AI breakdown brief, all the AI headline news you need in around five minutes. One of the biggest news stories of the day is the report that the White House is expected to release its AI executive order on Monday, but that will be the topic of the main show. So for now, let's cover some of the other most relevant news from this morning, starting

with a little update, but a significant one, to ChatGBT. Now, this has been a little less relevant since ChatGBTGT with Browse came to the four, but the knowledge cutoff for ChatGPT has moved up once again. When ChatGPT first came out, the knowledge cutoff was in 2021, and then recently it moved up to the end of 2022. Now, however, ChatGPT seems to be reporting on some desktop users that its information is up to date as of April 2023, which some folks tested by asking what was happening in and around that time, but then other folks are seeing even more up-to-date information. When developer Nick Dobos asked the iOS app, ChatchipT said my knowledge is

current as of September 2023. Now, obviously, this makes it a lot more useful to use even not in the Browse with Bing version, but I think it's also a reflection of just how much competitive pressure is shaping the way that AI labs make decisions. Initially, the knowledge cut off was a safety feature, as was not giving it access to the internet, but competitive pressure made both of those things go by the wayside fairly quickly. Anyways, moving on to the business side of the AI space, the big tech giants aren't just interested in AI because of the ways that it shapes the business models of the future and creates new product opportunities and positions themselves for the glorious years to come, but also because it has implications for their

core ad business right now. For example, yesterday TechCrunch reported that Google is actively looking to insert different types of ads in its generative AI search. This information, like much of what we've had reported from Google this week, comes from their earnings call on Tuesday, where they said that they are experimenting with a variety of different types of ad formats that are better suited to generative AI-powered search. Now, of course, this is hugely relevant because as many new products and different services as Google has, and as many attempts as they've made to diversify their revenue, the majority of their revenue still comes from ads. While we don't really have a lot of information about what the new ad units might look like,

Alphabet CEO Sundar Pichai has made very clear that they believe that there is going to be a new type of native ad format that's designed specifically to work with the search generative experience and is, as he put it, customized to every step of the search journey. Now, speaking of ads, another place that AI is coming into the ad world is that the big platforms that allow self-service ad creation, such as meta and Amazon, are giving advertisers more AI-powered tools to help them create better performing ads for customers more easily. As of Tuesday, Amazon has become the latest to roll out this type of tool set, adding AI-powered image generation to their advertising platform.

Amazon points out that in a survey from March of this year, among advertisers who were unable to build successful campaigns, 75%, three-quarters of them cited building ad creative and choosing a creative format as their biggest challenges. This is Amazon contends, and I subjectively agree, an area where AI image generation can be hugely significant. The example they give is for a toaster. Maybe an advertiser has a standalone image of the toaster on a white background, but they say that when it's placed in a lifestyle context, for example, on a countertop surrounded by fall accoutrements, click-through rates can be 40% higher compared to just generic product images. Again, this may not be the side of AI that gets your gears turning,

but when it comes to the immediate term economic impact, a lot of it is going to be in this sort of changing the way that advertising is made and consumed. And apparently it's working. Yahoo Finance reported this week, meta and Google's earning signal and AI-fueled revival in ads. The piece begins, A rebound in the advertising business of Google, Meta, and Snap signaled that the growing adoption of artificial intelligence was drawing markers to digital platforms even in an uncertain economy.

They pointed out that each of these companies posted positive metrics for their ad business, and that a lot of the shift is around how AI is helping the ad process. Now, one other announcement from Google that I thought was interesting, the company has announced new tools to help fact-check images. Now, you may remember a couple days ago or a week ago, I don't even know, I've lost all track of time, that I talked about how OpenAI had an AI image detection tool, but that they weren't sure when or if they were actually going to release it.

Now, of course, OpenAI had an AI detection tool for written text that they pulled because of inaccuracy. On the one hand, we've gotten signals from the company that they believe that this new tool for image detection is much more accurate. The CTO in an interview used the phrase 99% accurate to describe it, and yet still there is a lot of concern here. Of course, people are very hungry for ways to better determine what's real in a world where it's getting harder and harder if not already impossible to differentiate between real images, quote unquote, and images that are generated by artificial intelligence. A different approach that some are trying

is the voluntary inclusion of invisible pixels that can be read by detectors but are not visible to the human eye. Adobe announced something similar earlier this month in partnership with companies like Microsoft and Nikon. And for Google, the approach is a new menu option where people can click and select about this image to get more context about where and when it was taken, how long it's been on the web, and more. When it's available, users can also see metadata, including fields that indicate if it's an AI generated image. This is all part of a larger update to Google's fact check tools. Then we come today back to a big theme of the year, which is a course the AI chip shortage, as well as the geopolitics around chips.

It seems like there isn't a day that goes by without some announcement in this sphere. One of those today was that a company called SentML had raised money from giants like Google and NVIDIA to help combat the chip shortage by training AI systems to operate more efficiently. Now what's notable about this is that this is a $27 million seed round. Here's how Bloomberg describes how it works. The software helps predict the time taken to process tasks using different kinds of hardware. It monitors systems to pinpoint areas of underutilization, analyzing cost, power consumption, and emissions, then automatically distributes tasks to try and speed them up.

Apparently, according to research that the company has conducted, the average utilization for GPUs across the market is around 30%. The promise of CentML is that it can quicken systems by as much as 8X they claim. The team is a combination of PhDs from Carnegie Mellon in the University of Toronto, and will be one to watch if it can deliver on these promises. Meanwhile, other former giants of the chip space have had to shift their strategy entirely as they kind of get left behind. In kind of a brutal headline, the information writes, AI-lagged Intel expands effort to help companies build chat GPT like apps. Effectively, the piece

writes that as the company has been left behind by other rival chipmakers, instead, or at least in the short term, it's focusing on selling specialized software and services. Intel they write is working with multiple consulting firms to build chat GPT-like apps for customers who don't have the expertise to do it on their own. They apparently began doing this with BCG earlier in the year, but have added additional consulting firm partners as well. Now, this of course positions them in a similar, if slightly different space to things like Amazon's bedrock and data bricks and all of these companies that are designed to help enterprises either train models from the ground up or customize and fine-tune open source models. And while it might not be the chip space,

it's certainly something where there is likely to be a lot of demand. Now, one last story on the chip space. As we discussed yesterday, Nvidia said that the new US export restrictions, which were anticipated to come into effect about a month from now, instead actually came into effect Monday. Now part of the reason for that is likely concern that every day that goes on, especially after export restrictions are announced, is a day where China may get more access to the advanced chips that the US is trying to deny them access to. Bloomberg had a breaking story this morning that basically said that before certain restrictions

on ASML produced machines could be implemented, that some of those machines were sold to China and that their technology was used to with other tools to produce a highly controversial Huawei chip. The story sort of stands as a risk of delaying these things, and I think for our purposes dramatizes just how significant an issue the U.S. sees this, which I believe is in and of itself good setup for today's main discussion about the crescendo in policy discourse around the world. So that is going to do it for the AI breakdown brief. Thanks, as always, for listening or watching. Up next, the main AI breakdown. And now a word from today's sponsor.

interested in how two top-of-mind trends AI and crypto can work together? If so, I have the perfect podcast recommendation for you. Web3 with A16Z Crypto, the chart-topping show brought to you by venture firm Andresen Horowitz. Web3 with A16Z Crypto is your definitive resource for the future of the internet, whether you're already building in these spaces or simply curious about what's next. If you need a place to start, they recently released an excellent episode with Stanford Cryptography Professor Dan Bonay and former Google Xer Aliya, in conversation with host Sonal Choxi about the intersection of AI and crypto. From fighting deep fakes and proving humanity to large language models like ChatGBT, BT, they cover

it all. I highly recommend checking it out, especially if you'd like to learn more about how AI and crypto will impact our everyday lives. Beyond crypto and AI, this show is for creators seeking more ways to truly own their work, for business leaders trying to prepare for the future today, and for innovators exploring trending tech topics. So go ahead, listen to Web3 with A16Z Crypto, wherever you get your business. podcasts. Welcome back to the AI breakdown. There is something of a crescendo right now in the first wave of global AI policy. Ever since the emergence of ChatGBT, GBT, BT, governments around the world have, of course, been getting slowly more serious about this technology. Now, for some,

like the European Union, those efforts had already been underway, and the rise of generative AI was simply a catalyst to finally get things done. In other jurisdictions, in places like China, artificial intelligence has long been not only on their policy minds, but a very direct and specific strategic priority. However, in some places, notably here in the U.S., the rise of this new wave of tools, has required a very fast effort, basically from a standing start, for politicians at the Congressional Senate and executive level. The push to regulate these technologies has gotten even more acute in the wake of the rise of the AI safety conversation prompted by a number of factors this year, including the six-month pause letter, Jeffrey Hinton's

affection from Google to try to raise awareness of the risks of AI, and just generally the lobbying effort to make U.S. politicians aware of what some of the more sci-fi-sounding implications of this technology might actually be. A few months ago, the White House really started to pick up their efforts in this area. They began by inviting a notable group of CEOs from the AI space to a White House meeting that was chaired by Vice President Kamala Harris and which President Biden stopped by, which was also notable at the time for not including Facebook's Mark Zuckerberg. And subsequent to that meeting, we've gotten drips and drabs of information about what the White House is likely to try to do on this topic. Well, now according to reporting from the Washington Post, the White House is planning to unveil their much-anticipated AI executive order next Monday.

So what's the information we have so far about what is likely to be in this EO? One interesting piece of this is that there is a clear lead-by example attempt. In other words, while it doesn't appear, at least as of yet, that the executive order will impose mandates around assesses. that happen before models can be released generally, it does appear that they're going to require advanced AI models to have assessments before they can actually be used by federal workers. This is both an update to the federal procurement process, as well as a way to use that process as an evaluation tool for these advanced models. Now, in order to help different departments figure out how to use these technologies, federal agencies like the Defense Department and the Energy Department would be

required to conduct their own assessments to figure out how AI could help their different mandates and improve the way their agencies function. It sounds as if there will be a specific focus in those assessments on how AI can help bolster national cyber defenses. Now, speaking of bolstering, the executive order is also anticipated to ease barriers to immigration for highly skilled workers. This will be a very welcome update if it actually comes through to those who think that the U.S.'s immigration policies have put us at a strategic disadvantage relative to something that was once are huge advantage in terms of assimilating the world's greatest talent into our scientific establishment. The fact that that's part of the direction also suggests just how much the specter

of geopolitical competition, particularly with China, looms large over this particular executive order and the topic that it touches. Now, the way that the Post found out about this was that they saw an invite to an event called the Safe, Secure, and Trustworthy Artificial Intelligence event. The White House declined to formally comment, but the Washington Post sources said that the order had actually not been finalized, so details or even plans for timing could change. I think that probably the most interesting thing here, and the piece that people will be watching most closely, is exactly what this assessment process for governments using LLMs is going to be. The Post writes that this red teaming is expected to be led by the National

Institute of Standards and Technology. Earlier this year, that body released a framework for managing AI risks, and it is already collaborating with other agencies on this topic. Now, in terms of precedent of using U.S. government purchasing power as a way to influence the shape of a field, the Post points to a 2021 cybersecurity executive order from the Commerce Department as an example of how that had worked in the past. Now, of course, the White House is far from the only game in town, even in Washington. Earlier in the week, Senate Majority Leader Chuck Schumer hosted a second AI Insight Forum, which brought back a number of people from both the tech side, including notably Mark Andreessen, as well as from the academia side and the safety side of the equation, including Max Tegmark,

and the goal, of course, of these forums is to help politicians in D.C. get up to speed more quickly so that when they go about the work of regulating the artificial intelligence space, they actually have a firmer understanding of what's going on underneath. On top of that, there are numerous attempts to bring together more comprehensive AI legislation, as well as a huge number of smaller AI legislative efforts, so lots going on in Washington. Meanwhile, over in Brussels, while the EU had been lauded for moving more quickly than some of its peers, in passing its comprehensive legislation in the form of the AI Act, it's having trouble getting the details right in practice. Earlier this week, the Verge posted an article called

EU may fail to pass its AI Act in 2023. European lawmakers can't agree on how to regulate foundation models. So basically the way that this legislative process works is that after the act is voted, in principle, it goes through a set of debates and discussions around specific implementation details. These are called trilogues, which are three-party discussions between the European Parliament, the Council of the European Union and the European Commission, and there have been three around the AI Act, so four, with a fourth happening this week. Another provisional meeting has been set up for December if the parties aren't able to agree on things this month.

Now, one piece of the AI Act requires that foundation model developers, so think the open AIs of the world, be required to assess risks, do more extensive testing throughout the development process as well as red teaming, examine biases and training data, and just basically imposes a set of processes around these tools in order to try to drag them to a safer place. Now, one of the contentious issues is that the way that it's framed is sort of a dragnet that will catch up not just the big companies that have the budgets and the time for compliance, but also smaller open source projects, which A, aren't really the main concern that's trying to be addressed here, and B, don't necessarily have the resources to actually comply. The battle,

then, is to get the rules to reflect the diversity of projects in the space and have a better or clearer articulation of how different rules apply to those different types of foundation models. Now, on top of the open source question, there's also just other disagreements about specific implementation details. It sounds like from Loiter's reporting that Spain, who holds the EU presidency currently, has been trying to push through compromises that would have a tiered approach for regulating foundation models, which have 45 million users or more, as well as additional obligations for what they call very capable foundation models that includes regular testing for potential vulnerabilities. Apparently there are opponents of those plans that think that smaller

platforms can be just as risky. Now, of course, part of the context for all of this is the upcoming AI Safety Summit in the UK next week. British Prime Minister Rishi Sunak actually held a speech this morning discussing the issues underlying that event. In that speech, Sunnick did acknowledge the potential for society to lose control over AI with potentially cataclysmic results. He also discussed other risks including heightened cyber attacks, more advanced fraud, child sexual abuse, and potential changes to the labor force. He discussed the capacity of AI to be used by terrorist groups, quote, to spread fear and disruption on an even greater scale. He said, as the independent summed up, that human extinction risk from AI is on the same scale as pandemics or nuclear war,

and that we shouldn't put our heads in the sand when it comes to these issues. At the same time, he said, the opportunities are immense, and this speech really laid a lot of that out. He talked about how much potential AI had to catalyze economic growth and to enhance productivity, and he said that he was generally optimistic about the potential of AI to transform people's lives for the better. When it comes to regulation, on the one hand, the Prime Minister was very clearly trying to put UK in a leadership seat on these issues, but he also said that they would be in no, quote, rush to regulate AI, because as he put it, it's hard to regulate something you do not fully understand. Now, of course, perhaps the biggest issue surrounding the summit is the decision to invite

China to participate. This was initially controversial in the UK, in part because of a Chinese spying scandal that was happening at the same time that the summit was being announced. Beyond that, of course, there is just a sense that when it comes particularly to artificial intelligence, the West and China are much more competitors than collaborators, even though, of course, to look at cutting-edge research is to see a very different story. So Nuk took those concerns head on, saying that it wasn't an easy decision to invite China, but that it was in his words the right thing to do, because the threats and opportunities of AI are too big to have Balkanized strategy across the world. Now, in his speech, the British Prime Minister said that he wasn't sure if China was actually even going to attend.

However, Deputy Prime Minister Oliver Dowden said that China had accepted the invitation. He told the BBC, it is the case they've accepted, but we will wait to see everyone who actually turns up at the summit. As things stand, yes, we do expect them to come. Meanwhile, there had been some concern that there was a growing wave of leaders who were deciding not to attend, potentially minimum minimizing the significance of the event, but at least when it comes to the U.S.'s participation, whatever concerns they might have had behind the scenes around the China invite, they decided to get over them and Vice President Kamala Harris is still expected to attend as had been the plan. Indeed, given the timing of this executive order, it's pretty clear that the White House is

trying to align or perhaps even assert themselves in the context of this UK event where that government is obviously trying to position themselves as leaders as well. The one thing is for sure is that next week will be a big one for AI global policy. And of course, I will bring you all of the updates as they happen here on the AI breakdown. Until next time, guys. Peace.