Regulating AI; Negative Interest Rates (#19)

Hi everyone. Welcome to this week's episode of 16 Minutes, our new show where we cover recent headlines, the ASICs and Z-Wave, why they're in the news, why they matter from our vantage point in tech. I'm Sonal. This week for episode number 19, we have two segments. The first is on all the recent news around regulating AI or artificial intelligence, including both the White House's new guidelines and recent limits to exports of AI software that just went into effect this past week. And then the second segment is on all the recent news around the topic of negative interest rates and quantitative easing, why it's in the news and where tech comes in here. As a reminder, none of the following should be taken as investment advice. Please be sure to see a6nz.com/disclosures for more important information. So the first segment is actually two items. The White House, specifically their Office of Management and Budget, the OMB, and the Office of Science and Technology Policy, the OSTP, issued a draft memorandum, really a set of 10 principles for the stewardship of AI applications, including public trust, public participation, scientific integrity and information quality, risk assessment and management. So that's acknowledging all the tradeoffs involved, benefits and costs, flexibility. So not trying to do overly rigid regulations and making them more performance-based, fairness and discrimination, disclosure and transparency, and that the AI be safe, secure, and operate as intended. And finally, this is the 10th principle, interagency coordination for a more coherent coherent and whole-of-government approach to AI. These principles are very broadly defined, but the federal agencies have 180 days to submit explanations of how any of their proposed regulations satisfy these principles, which even though they don't have the authority to overturn that. Okay, that's a quick high-level summary. And now let me introduce our A6NZ expert, operating partner Frank Chen, who is our expert on all things AI and more. Can you help us by just situating us relatively for where we sit globally?

[Speaker:GHION.AI] Yeah, so almost every country in the world has set forward an AI policy, and the AI policies differ between countries, but they're generally like, hey, I wanna be the country that does AI the best. I wanna do it safely. I want my economy to benefit the most. So the White House has been working on defining a national policy. Their first reports on this were published in October of 2016, and I actually wrote a primer that was cited in that piece. And if you look at the principles, it's almost sort of a meta policy. They're principles to think about when you're defining policies. Contrast that with China, which has very explicit policies, guidelines, funding. The US is the undisputed leader of AI research, AI industry. If you look at China's policies, they are all very self-consciously, how do we catch up?

It's interesting because the effort for behind these principles is actually driven by something called the American AI Initiative. And it was created by a presidential executive order last year to create a national strategy, and it's titled for maintaining American leadership in artificial intelligence. And then the US CTO, Michael Kratsios, wrote an op-ed in Bloomberg this past week, basically arguing that this is actually meant to advance innovation and the goals of it are to ensure public engagement, limit regulatory overreach, and promote trustworthy technology. So that sort of seems to align with that leadership position.

Yeah. Let's not trip ourselves, right, with too much regulation. Yeah. The thing that the AI community actually mostly worries about with respect to government oversight is around sort of the use of data and how much harder will it be to get data.

Interesting.

So China is way behind on algorithms, but the one area everybody acknowledges China being ahead on is basically data and ease of access. To turn that into dollar terms, imagine that I do a Series A investment of $10 million into a startup, and imagine that $3 million of that goes to getting and cleaning data and building models around it. The equivalent cost in China to that $3 million is probably $30,000, $50,000, because they don't have to worry about cleaning data and data and licensing and keeping it safe.

Right, right.

And look, there's downsides to that approach. But what that does mean is they can run a lot more experiments than we can on the same dollar.

So I'd love to ask you one quick question though. The recurring theme that I saw in all of, of what came out this week is this idea that the public has the ability to participate and engage. That by having input and the ability to have these conversations in a transparent way, we can actually create more trust in AI. But I have to ask you, what's your take on this idea of quote trustworthy AI?

Look, we've had this debate about trustworthy computing since the beginning of time. Remember when the Intel chips had like division errors? And so we couldn't trust the results.

Andy Grove. Wrote about it in his book.

Right. And so given how important computers are to our society, it's pretty important that the computers get the right answers. So this debate is basically coming back because now we don't know why the computer has reached an answer that it's reached with the most modern machine learning techniques, especially deep learning.

Let's quickly unpack the phrase transparency, because that can mean so many different things to so many different people. I read the report as meaning transparency in terms of communication. But then there's also this whole world of explainable AI.

Yeah. So if you look at the principle, the principle is disclosure and transparency. Which is they're saying the public will only trust AI if it knows when and how it is being used. So if you remember a couple of months ago, Google demoed this thing called Google Duplex where it would call a restaurant and make a reservation, and it was an algorithm doing it. And after that, when there was a little backlash, it would self-identify as an algorithm. I'm not a person, I'm an algorithm. So transparency is about, I'm telling you that machine learning is being used so that you are not surprised. Another dimension of transparency is understanding why algorithms make the decisions they do. So, for instance, if we put a machine learning system in charge of deciding whether you get a loan or not, we'd want to understand why did you make that decision, right? Because there's disparate impact laws on the books that say you can't discriminate on the basis of gender or sexual orientation or religion or any of that.

Right. But that's not the transparency in this particular report.

I don't think that's what they're talking about. Right.

However, they do cover that idea through this note that there needs to be fairness and no discrimination.

Right.

Okay. So then now let's talk about this other bit of news that sits in contrast to this, which is that last week Last week, the administration took measures to limit exports of AI software as part of a bid to keep sensitive technologies out of the hands of rival powers like China, which you mentioned earlier. This is from an article in Reuters. It's actually really interesting because Kratzios made a point of saying that there's also, quote, light-touch regulatory approaches, and they included non-regulatory approaches, including sector-specific policy guidance and frameworks, pilot programs and experiments, voluntary consensus standards. And then you have the other extreme, which is these very specific ITAR regulations for exporting certain types of AI. And the rule went into effect just this past Monday and is focusing on companies that export certain types of geospatial imagery software from the United States. They have to apply for a license to send it overseas, except when it's being shipped to Canada. Can you quickly give us your take on what the scope of where that fits into how we think about regulating AI?

Yeah, so a little history here. When I was growing up, in Netscape days, there were very broad prohibitions about the strength of encryption that we could export. So for every build of Netscape Navigator, the browser, and every build of the server products, we actually built them twice: one with weak encryption and one with strong encryption, 128-bit keys, because software was literally regulated as munitions.

Right, this is under like the export control ITAR, like the rules—

ITAR, right, the regulations that govern, like, who can you sell an M16 to, or an F-16, or an aircraft carrier. Like, literally software was placed under the same bucket, and it was so broad. To me, this is so specific, which is good. The software has to have all 4 of these things. It's got to have a GUI—

graphical user interface—

a graphical user interface. You've got to be operating on pixels and doing scale and rotational normalization. You've got to be using a deep convolutional neural network, right? So if you were using a recurrent neural network, it didn't apply to you. And you have to be identifying objects and geospatial imagery. So think, I've taken pictures from satellites and I'm trying to find the military bases in those pictures.

Got it. Why specifically did they single out the deep convolutional network in your view? There's a whole set of toolkits of neural networks and various technologies that people use when it comes to deep learning and machine learning, different types of recurrent neural networks and all types of different things.

The short answer is it's these deep convolutional neural networks that are working for image recognition tasks, achieving the state-of-the-art results.

—Which are?

—Which are— so you basically compare these to human performance.

Oh, this is like basically where you have photos that you can't get computers to recognize, but people can, where they can tell the difference between a bagel and a dog. Or there's like many variations of this.

Right, exactly. And where we've gotten to is the algorithms are now on average better than people. Ooh, interesting. And it's this particular technique, the convolutional neural network. So if you're looking for military bases in photos, this is exactly the technique you want to use. Look, that's probably not a durable thing. In 5 years, we might be using a different technique, and all of a sudden adversaries have software that are recognizing military bases in ways we don't want that don't use this technique. And so we have to just be very mindful of how specific we write our export requirements. I think the real question here is, is AI the right thing to be thinking about to regulate, or is it too broad? I always make the analogy of AI is a little like databases, which is here's a fundamental set of computer science techniques that it's generally useful across a broad class of software. And every time I see, let's regulate AI in this way, I sort of do the mental substitution let's regulate databases in this way and see if it makes sense. AI really is software.

Particularly interesting when you think about the future of software and the evolution of open source, because one of the best quotes, I think it was Bill Joy said, he was a former Sun Microsystems, that all the smartest people in the world will never work for you. This idea that people can collaborate around software, we have like AWS phases coming to all kinds of industries. It launched a whole set of innovation. So with these recent regulations around AI, where does this fit in that picture?

I think to keep America at the forefront, we want to have our fingers in as many of the widely used open-source platforms in the world that form the basis of everybody's software. So I think we'd all feel better if a leading Chinese AI company was using Linux, looking at their results in a browser that was a derivative of Chromium, using Google's TensorFlow. Because imagine what would happen if we forced them onto an all-Chinese stack, if they were using Red Flag Linux, which is the Chinese variant, and Alibaba's equivalent of TensorFlow, which is called PyTorch, which nobody uses today. But if we write our policies in a way that makes it very difficult for our technology to get out there, then they will switch. And so we need to think about, do we want that? Do we not want that? When we're setting the slider of like, what are you allowed to export? What are you not allowed to export?

I'm so glad you shared that sort of counterfactual vision of a stack that's not US. Based, but Chinese or otherwise based, because it basically makes you realize that these AI principles and regulations are not just about maintaining US innovation in the container of the specific product, but that the product can also be infrastructure and the things that the infrastructure then enables.

I almost think of it as like, you know, it's Korean K-pop, right? It's soft power.

So given all this, bottom line it for us, Frank, how should we think about this? What's our takeaway on this recent set of news?

So the exciting thing is, as a class of computer science techniques, artificial intelligence is so important now that we've We've got government regulators expressing interest, and we've got policy leaders who want us to maintain our leadership. Now the questions are, is artificial intelligence the right thing to regulate? And what are the concrete policies that we're going to put in place? We're out in front, and let's not cede our leadership.

Great. Thank you for joining this segment, Frank. Thanks for having me. All right. So in this second segment, we're going to really quickly cover the recent headlines and commentary around the topic of negative interest rates and where tech may or may not come in. It's actually not a new topic, but it's been in the news recently. Recently because the former chairman of the Federal Reserve Bank, Ben Bernanke, he was known for sort of pioneering some of the measures that helped during the financial crisis of 2008, published a blog post this past weekend called The New Tools of Monetary Policy, based on a talk he gave at the American Economic Association annual meeting. And in that post, he was advocating for, quote, new policy tools such as quantitative easing, forward guidance from the central bank, and negative interest rates as being effective. He's basically arguing that policymakers need to increase the range of flexibility they have of what central banks can do during a recession or a crisis, a financial crisis. And he published a paper and blog post, which I'll link to, that you can get up to 3% with some of these tools. So that's just a quick high-level summary of what he wrote. Now let me introduce our A6NZ expert, General Partner Alex Rampell, who covers all things fintech and more. Welcome, Alex.

Thank you. Good to be here.

First of all, maybe you can just give me what what negative interest rates actually are?

Sure. I mean, so you mentioned some of these different tools. The only one that's new and doesn't make any sense to most people is negative interest rates. So quantitative easing is just another word for printing money. This happened in the Weimar Republic in Germany. It's happened in Argentina. It's happened in tons of countries, almost with the exception of the US.

And by the way, this is where you purchase long-term financial assets. The central bank purchases them as a way to stimulate the economy.

Think of it as like they've got this Excel spreadsheet and they can add as many zeros to like the how much money we have. And it just appears out of thin air. So printing money used to mean like literally printing notes of paper. Now it just means who keeps track. When you send a wire from Wells Fargo to Citibank, how does that happen? That goes through the Fed. So the Fed is kind of this master database that keeps track of who has what money, because again, it doesn't exist in printed form. It's not backed by gold. It's not backed by anything. The Fed just says, okay, here's how much money we have. We have $1 trillion. But they could say we have $100 trillion. And if we buy $50 trillion of assets somewhere else, They can just do that because they make the money. So this is how they add zeros. Yeah. So it's literally adding zeros to their own bank account. And we don't let Citi do that. We don't let Wells Fargo do that. But the only person that can do that is the sovereign. And in this case, the sovereign is the Fed. Right. And again, this is not new. I mean, it's a much prettier name. It's a great euphemism. What does quantitative easing mean? It's easy to have a higher quantity of money, whereas negative interest rates really don't make any sense because you get charged for the right of parking your money somewhere. The borrower always pays money to borrow, and the lender always makes money by lending. So there's a seller and there's a buyer. The seller makes money, the buyer puts up money, and in negative interest rates, that's flipped. So how does it work? Well, okay, one way of thinking about this as a thought experiment is imagine that you are really rich and you have $1 billion. Where do you keep the $1 billion? Well, you could keep it under your mattress, and that has problems. You don't earn interest. But again, if interest rates are negative, there is no interest. What if your house burns down? What if there's a robber? How do you get money from point A to point B? So it's useful to have money in a liquid form that can be easily sent around the world. Now, you could keep it with a bank, but banks are not sovereign entities. What if the bank falls apart? Now, it's insured by the government. It's actually called the FDIC, the Federal Deposit Insurance Corporation. So if you're a bank, any deposit held by one of your customers is insured up to $250,000. So I have $1 billion and I keep it in a bank and then the bank goes bust. Well, wait a minute, I just lost, you know, 99.999% of my wealth. Now, if I buy government bonds, the good news is that I can always sell those. They're backed by the sovereign. So the only way that I lose my money on my government bonds is like the—

If the government goes down. If the government falls.

So the thought experiment is, okay, I have $1 billion. I don't want to give it to a bank. Here's what I'm going to do. I'm going to put it in this giant safety deposit box. But then what if there's a meteor that hits that? Well, I'm going to buy insurance. I'm going to go to like, you know, Prudential or Lloyd's or somebody and say, how much will you charge me? Okay, well, we'll charge you 1% of the value of all of the money in the safety deposit box. Well, it turns out that's what negative interest rates are. At least that's the only way that you can make sense of them. Because again, like, why would you be paying somebody to get a mortgage? Because if you think about what this means, the borrower is paid to borrow money and the lender is charged to lend money.

That's actually a great thought experiment because you're basically saying it inverts the classic model. Definitionally, negative interest rates, however, are basically zero or close to zero or even below zero. So what does that mean given that interest rates have been falling downward for quite a while? And currently, I believe the Fed has stated that they're keeping it around 1.5 to 1.75%.

Well, a lot of this is people have become accustomed to yield. So if you think about somebody who's on a pension that gets paid X dollars per year, a lot of pensions, in order to meet their payout guarantees— every year I have to pay out 7%. If you have a retirement account and you're used to taking out money every year, where does that money come from? Or there's a federal law, like if you're a nonprofit foundation, every year you have to pay out 5% of your assets. Or if you're an insurance company, you collect premiums and then you go invest the premiums and you have to reserve that in safe securities to go pay out the benefit when somebody nobody dies, and you could just buy GE bonds or something, or some like safe fixed income security and make a good living on that. And now those rates have collapsed. So a lot of these models have turned upside down because that wasn't that hard when interest rates were higher, like above 10%. Well, I could just loan money to the government and then I get 10% every year. Now the problem is that if I buy, you know, treasury bills, which are the safest instrument under the US dollar, I'm getting 1.8% or something. And if I'm getting 1.8% and I have to pay out 5%, well, that's a problem. These yields have collapsed. So now there is this great search, just like there was the explorer era in the 1400s and 1500s, like, let's figure out what this new world looks like and colonize it. Now it's like, wow, we have to find some new territory for more returns. We have to because we're screwed. So there's this great search for yield.

Okay, so my question is, where does the tech industry come in?

It impacts tech for a couple of different things. So one is quantitative easing has made the prices of assets more expensive. One of the reasons why the stock market goes up is if there's more supply of capital because the government prints more money. Well, that has to go somewhere and then all asset prices go up. So, you know, what's really interesting about quantitative easing versus printing money is that once upon a time, the consumer price index, that's what you would see your inflation in. So like the price of toilet paper would go up, the price of bananas would go up like bread, these things. Yeah, exactly. But because in quantitative easing you were buying assets, it just caused the price of assets to go up. So rich people got a lot richer. If you owned a house, you made a lot of money. Negative interest rates is like, okay, now you have assets and the problem is that you're not earning a return on your assets. It doesn't really impact the folks that don't have assets, but like if you're just increasing the supply of money and you have assets, now those assets are worth more money. And they're related because when interest rates are zero, then people start looking for more esoteric sources of yield. But they are separate. It's amazing. You'll see companies like Apple or Microsoft, they're issuing bonds that are marginally above US Treasuries. There's no such thing as a safe asset if you're thinking about like what stock to buy because there's so much disruption. The number of companies today that are in the S&P 500 that were there 30 years ago, there's massive, massive turnover.

They're no longer the same list at all.

That's one. The other is that the stock market isn't yielding that well. I want more volatility that can potentially offset the 0% or 1% or 2% rates that I'm getting in my fixed income portfolio. The way that I'm going to get a blended higher return is seeking out more risk. The higher risk you take, the higher return. The most perilous asset in the world called a brand new tech company with 2 people in a garage that probably is going to go to zero. But if it works, it could be worth $100 billion. I want more of that because if it works, that's not going to yield 1.7%. It's not going to yield -0.7%. It's either going to— I'm going to lose all my money or this could be very, very big. Right.

The point is that tech represents growth in this context.

Well, the 5 biggest companies on earth by market cap are all tech companies. The worst performing stocks of the last 10 years have been what have been called value stocks. So like, you know, look at things with a very low P/E ratio. Price to earnings multiple. Yes. So these have really suffered because it's kind of adverse selection. They're great values for a reason. They're in the bargain basement because they don't really have that much growth ahead of them. And if you had invested in Amazon in 2009 or Netflix in 2009, like, that's mega growth, very risky, right? But I can't make my foundation work. I can't make my P&L work. I can't make my pension plan work unless I get a higher rate of return. I now have to seek out more risky assets. So therefore, tech companies is the source of growth. Like, that's how I'm going to get my yield. And a lot of that is spurred by having negative or close to zero interest rates, which is now I need to take out more risk and more risk and more return is in tech.

And so that's how it affects tech. So bottom line it for me, Alex, what's your big takeaway on the news?

Negative interest rates aren't new, so there's not that much that has changed there. It's more that in this great search for yield, this great new exploring territory of how do I get some kind of yield that's above zero? I want to search out riskier and riskier things. And there are more entrepreneurs that are building things that are in fact risky, but if they they're gonna be big and change the world. So there's more capital now pursuing tech investment. Thank you for joining this episode, Alex. No problem. Thank you.