The AI Daily Brief: Artificial Intelligence News and Analysis - Biden Meets AI Safety Experts As White House Promises Decisive Action

Episode Date: June 21, 2023

President Biden yesterday met with AI safety experts in San Francisco as the White House promises "decisive action" on AI in the weeks to come. CNN reports that senior advisors have been meeting on AI 2-3 times per week since April. On the Brief: OpenAI Is reportedly considering opening an app store for custom models build on top OpenAI; a new LLM beats GPT3.5 with 1/1000th the training data; Opera releases a new AI browser and more. 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 discussing President Biden's meeting with AI safety experts yesterday in San Francisco. The AI breakdown is a daily podcast and video about the most important news and discussions in AI. Like, subscribe and share and go to breakdown.network for more information. Welcome back to the AI breakdown brief, all the AI headline news you need in five minutes or less. Well, once again, Tuesday has lived up to its reputation as the biggest newsday in AI, and we kick off today with a report from the information that Open AI is considering launching a new AI app store. Right now, of course, there are many, many companies that are building on top of OpenAI's models. Enterprise customers, of course, take those general models and customize them for their specific use cases.

Silicon Valley-based publication The Information has reported that last month in a meeting with developers, OpenAI CEO Sam Altman said that they were considering creating an app store that would make it easier for those enterprise partners to offer their custom AI models to other buyers. Now, one concerned that some have brought up is that this might compete with other Enterprise AI app stores such as Salesforce and Microsoft, but it also makes sense from OpenAI's perspective. According to the report, companies like AQUANT and Khan Academy were some of those who had lined up to potentially make offerings on this new marketplace. The information's Jessica Lesson yesterday posted on LinkedIn, seems we've entered a new phase of the AI wars, the fight to be made gatekeeper. My colleagues had an important scoop today on it, made me curious where you all think we are in the AI super cycle. Of the people who responded to her poll, 83% said we are just beginning.

And when it comes to the market's response, I think Amira Fraudy summed it up pretty well. Tired, he writes, chat GPT plugins, wired, customized chat GPT models trained on proprietary data. Next up, an interesting story on the geopolitics of AI. The U.S. for the last couple years has been putting ever-increasing pressure on the Chinese market to deny them access to advanced chips and bring the semiconductor manufacturing industry back home to the U.S. And it appears that that's having an impact in domestic markets. Reuters today released a report about China's underground market for high-end Nvidia chips, where Nvidia's AI chips are going for double what the price is in other places of the world where they are not denied access.

Now, they're not technically illegal in China, but it seems like the local market is still trying to avoid the scrutiny of either U.S. or Chinese authorities. Speaking of chips, Cisco has announced a new AI networking chip, the Silicon One, which is meant to compete in the most rigorous AI and machine learning environments. And when it comes to AI Native products, Opera has announced Opera 1 with built-in area GPT. They're calling it the first AI native web browser. And just based on the discourse we've seen over the last couple of months, there seems like there's going to be a lot of competition in this space of browsers that are retrofitted and reimagined for a chat GPT style interface. Now when it comes to chat bots like chat GPT and the AI that underlie them, one of the big questions is, of course, just how advanced are they?

Ever since 1950, when Alan Turing came up with it, the Turing test has been a key benchmark for just how advanced any sort of computing system is. while now Mustafa Suleiman, the founder of Deep Mind and Inlection AI, has proposed a new kind of Turing test for today's advanced AI. When it comes to the Turing test, Mustafa said, quote, it's totally unclear whether this is a meaningful milestone or not. It doesn't tell us anything about what the system can do or understand, anything about whether it has established complex in monologues or can engage in planning over abstract time horizons, which is key to human intelligence. Mustafa has a new book coming out called The Coming Wave, Technology, Power, and the 21st Century's Greatest Dilemma, and in it he argues that we should be focused

on whether AI has achieved what he calls artificial capable intelligence in which programs can set goals and achieve complex tasks with minimal human intervention. ACI then is kind of what we've all wanted out of our auto-GPT tools. To measure whether a machine has achieved artificial capable intelligence, Mustafa proposes an updated modern Turing test. In his model, you would give an AI $100,000 and see if it can turn that investment into $1 million. Given that, the bot would have to research e-commerce businesses, generate blueprints for a product, find a manufacturer, sell the item, etc., etc. Now, I'm not sure if this is the right way to look at AI or if this is the right test to determine its advanced capabilities. I will only note that as people have gotten excited about autonomous

AI agents, lots of us have explored similar types of use cases. I did a video a couple months ago, for example, called Can AutoGPT Grow a 10K YouTube channel? So it'll be interesting to see if people pick up on this new Turing test as something that we should actually organize around. Finally, today in the debate around whether more advanced LLMs require ever-increasing sizes of data to train on, a new LLM called Phi-1 has achieved better performance than GPT 3.5 with less

than 1-1,000th of the training data of some of its comparable peers. Phi-1 achieved 51% human eval, which is higher than GPT 3.5's 47%, although lower than GPT-4-67% and wizard coders 57.3%. However, as researcher Sebastian

Bubak puts it, any other greater than 50% human eval model is greater than a thousand percent bigger, e.g. Wizard Coder from last week is 10x in model size and 100x in data size. By one achieved its 51% on human eval with only 1.3 billion parameters and a 7 billion token training data set. The secret was apparently training the model on textbooks, and it appears that the quality of that data that it was trained on made a big difference. We will explore more of the implications of this in a future video about data sets and training sizes, but for now, a pretty exciting experiment to see. That's it for today's AI breakdown brief.

If you're enjoying, please like, subscribe and share, and I will be back soon with the main AI breakdown. The White House is promising what they call decisive action on AI in the upcoming weeks, and yesterday President Biden met with AI safety experts in San Francisco. There is no doubt that AI is higher on the White House's agenda than ever before. At the beginning of May, Vice President Kamala Harris met with a number of AI CEOs at the White House. This included CEOs from companies like Microsoft, Anthropic, and OpenA.

but notably did not include Meta's Mark Zuckerberg, a slight which the White House doubled down on in following comments. Just after those meetings, the White House announced a number of new initiatives around AI as well. For example, the White House announced that the National Science Foundation had been authorized to spend $140 million to create a set of new research centers around the country that were devoted specifically to AI. Additionally, they pledged to release draft guidelines for government agencies to ensure that their use of AI safeguarded, quote, the American people's rights and safety. The White House also announced that several AI companies had agreed to make their products available for scrutiny at a public cybersecurity conference. In a public statement following the meeting, Vice President Harris wrote,

advances in technology have always presented opportunities and risks, and generative AI is no different. AI is one of today's most powerful technologies, with the potential to improve people's lives and tackle some of society's biggest challenges. At the same time, AI has the potential to dramatically increase threats to safety and security, infringe civil rights and privacy, and erode public trust and faith and democracy. As I shared today with CEOs of companies at the forefront of American AI innovation, the private sector has an ethical, moral, and legal responsibility to ensure the safety and security of their products. Now, yesterday, CNN gave a little bit of insight into where the White House's push to regulate AI came from. Apparently in April, as Biden was sitting in the Oval Office talking to his advisors about AI, his aides were typing different prompts into chat GPT to show President Biden what it could do. One of those prompts was asking ChatGPT to summarize the Supreme Court's,

New Jersey versus Delaware ruling and to turn it into a Bruce Springsteen song. The genesis of that was that a few weeks earlier, the president had joked with Springsteen himself that the decision in the case, which centered on rights of the Delaware River, gave Delaware a claim to being Springsteen's home. When ChatGPT was able to turn this ruling into a song, President Biden was like so many of us have been when experiencing ChatGPT for the first time, wowed. By the end of the meeting, Biden made it clear to everyone in the room that AI needed to be a top priority. Importantly, a number of the white White House's advisors on AI have told the Biden administration that they believe they are behind when it comes to AI policy. Yes, the AI Bill of Rights that they drafted last year was a good start,

but that still wasn't enough given how fast AI was evolving. According to CNN then, senior administration officials have been meeting two to three times a week to work on AI policy ever since that Oval Office meeting. Those meetings are looking at AI across multiple fronts, including the cybersecurity dimensions, the misinformation dimensions, questions of discrimination and economic equity and more. Interestingly then, yesterday, in San Francisco, Biden held a meeting with another group of AI experts, except this time it wasn't the CEOs of major corporations. It was instead a group of people who are notable for their work on AI safety or AI risk. That included officials from Stanford's Institute for Humane Centred AI, including Fei-Fei and Rob Reich, Dr. Joy Bualam-Winney,

who's the founder of the Algorithmic Justice League. Dr. Joy tweeted, I am looking forward to this meeting with POTUS on the dangers of AI and what we can do to prevent harms already impacted. everyday people seeking mortgages in housing, in need of medical treatment, encountering workplace surveillance, and more. The meeting also included Center for Humane Technology co-founder Tristan Harris, whose AI dilemma talk for March is actually a really good starting place for people who want to explore the risks and dangers of AI. At the meeting, President Biden said, we'll see more technological change in the next 10 years than we've seen in the past 50 years, and maybe beyond that. And AI is already driving change in every part of American life.

We need to manage the risks to our society, to our economy, and our national security. Now, when it came to the tone of the meeting, Dr. Lee told Politico, quote, there was recognition of how the technology, if not used responsibly, could have very negative implications and how it could have constructive and positive implications. Lee said that she had urged Biden to take a, quote, moonshot mentality when it came to investing in the public sector to ensure what she called responsible stewardship of the technology into society. According to CNN, while senior administration officials recognize that there will need to be a legislative dimension of AI policy, in the short term they're hoping to shape the regulatory framework through executive action.

Some of the things that they've been working on in those multiple White House meetings per week include creating an inventory of government regulations that are already on the books that could be applied to AI, working with companies like Google Microsoft and OpenAI to announce privacy and safety commitments that were created in coordination with the White House, and guidance from the Office of Management and Budget for federal agencies on the use and procurement of AI. There's also an effort from National Security Advisor Jake Sullivan and his team to develop policies that relate to AI and cybersecurity. Now, on the legislative front, there has also been a lot of chatter recently. For example, yesterday, a group of bipartisan lawmakers have proposed a blue-ribbon AI commission. Representative Ted Liu and Representative Ken Buck of California and Colorado, respectively, introduced a bill on Tuesday that would have Congress in the White House appoint 20 people across government, industry, civil society, and computer science to an AI commission. That group would issue three reports over two years with recommendations for, quote,

mitigating the risks and possible harms of AI, while also protecting U.S. tech innovation. Lou said that AI could be, quote, disruptive to society, from the arts to medicine to architecture to so many different fields, and it could also potentially harm us, and that's why I think we need to take a somewhat different approach. Now, while the goal of the commission wouldn't be to delay any new rules or regulations until after those reports had been released, Democrat Lou did say that it would be, quote, prudent for lawmakers to hold off on overarching legislation until that group could weigh in. He said, I just think we need some experts to inform us and have just a little bit of time pass before we put something massive into law. It's worth noting that

Ted Lou is one of the only folks in Congress who has any sort of computer science background. Until we get more information about what Biden and the White House are actually thinking when it comes to policies, it's hard to know exactly how much they're going to come down on the side of risk and safety versus coming down on the side of innovation and opportunity. However, what does strike me as clear is that I think President Biden is giving more than lip service to this issue. He said in a news conference earlier this month, I don't ever think in the history of human endeavor that there has been as fundamental potential technological change as is presented by artificial intelligence. It is staggering. When I see him say stuff like that, I actually take it

at face value. I think that he does have a sense of the magnitude of the change here. But of course, the devil will be in the details. And as we discussed on yesterday's show, putting new regulations onto the books and implementing new policies can be an extremely thorny, challenging, and delicate endeavor. In the meantime, however, if you're interested in a lighter take on AI in politics, there is currently an endless debate going on on Twitch TV between AI Joe Biden and AI Donald Trump that is unhinged, offensive, definitely not safe for work, and pretty hilarious. It's powered by GPT and Play.HT's voice cloning, which is what I used to clone my voice a couple weeks ago for Long Read Sunday. And it's definitely worth a couple highly offensive minutes

of your time. Anyways, guys, that is it for today's AI breakdown. If you're enjoying the show, please like, subscribe, and share. Hit the notification button so you don't miss an episode. Go check out the podcast in the newsletter version. And until next time, peace.