← Artificial Intelligence Masterclass
Artificial Intelligence Masterclass
Best Practices for AI Deployment, Safety, Security, and Regulation - "Complex Adaptive Systems" - AI Masterclass
00:29:23
Loading summary
Transcript15 lines
- [00:01]
A
Running a business comes with a lot of what ifs, but luckily there's a simple answer to Shopify. It's the commerce platform behind millions of businesses, including Thrive Cosmetics and Momofuku. And it'll help you with everything you need. From website design and marketing to boosting sales and expanding operations. Shopify can get the job done and make your dream a reality. Turn those what ifs into Sign up for your $1 per month trial@shopify.com specialoffer. Save over $200 when you book weekly stays with Vrbo this winter. If you haven't seen your college besties since, well, college, you need a week to catch up in a snowy cabin, take a week long vacation and save over $200. Book now@vrbo.com Experience a membership that backs what you're building with American Express Business Platinum. Enjoy complimentary access to the American Express Global Lounge Collection and a welcome offer of 200,000 points after you spend $20,000 on purchases on the card within your first three months of membership. American Express Business Platinum there's nothing like it. Terms apply. Learn more@AmericanExpress.com Business Platinum.
- [01:15]
B
All right, good morning everybody. Yeah, so I've been wanting to cover this topic for a little bit, but it is, it's kind of a byzantine topic, so I'm going to try and make it as interesting as possible. But the long story short is that what is emerging is that the correct way to think about artificial intelligence, safety and regulation is through the lens of complex adaptive systems or cas. Or cas. So complex adaptive systems are. Well, let me just give you two kind of examples off the top. So the one that you're probably most familiar with is going to be the stock market. And I'll give you an explanation as to how the stock market is characterized as a complex adaptive system in just a second. But the other one is social media. So in both cases you can see viral effects, you also see emergent behaviors, and so on and so forth. So with those two in mind, the stock market and social media, let me give you the top, I think nine or ten criteria of complex adaptive systems. And this will give you a little bit more context as to what do we mean, what are the characteristics of these complex adaptive systems? So number one is emergence. Emergence is the appearance of complex or higher order behaviors or patterns of behavior that emanate or originate from simpler lower level rules. And so an example of emergence is flocking behavior. So, so birds and schools of fish, they actually follow very simple rules, but because they're following simple rules in large numbers as individual agents, they can actually respond to each other in dynamic ways as well as predators. So for instance, if you've ever watched a video where like a shark swims through a school of fish, and the school of fish just parts like the Red Sea, that is an emergent behavior. Next is self organization. So self organization is the spontaneous creation of order and structure within the system, however, without external control or direction. So that self organization, for instance, the way that ants tend to organize their colonies and their, and the way that bees organize their hives, that is another example of self organization. Although I don't know whether or not like schools of fish or beehives, I don't know if this qualifies complex adaptive systems, but they do actually embody some of these characteristics. Next is non linearity. So relationships between components that are not proportional, leading to unpredictable outcomes, making the system difficult to model using simple linear equations. So non linearity is basically the equivalent of stuff being viral. So viral media is an example of non linearity. But then also you can look at, at examples in the stock market, like the flash crash that happened, gosh, probably 10 years ago now, where basically, you know, one per one, one trader accidentally put a billion instead of a million and was selling too many orders and it caused huge ripple effects. Not only did the price of one stock crash, but everyone started panicking. And so you have, you have these contagion effects where, you know, one unexpected behavior results in more unexpected behavior behavior in other nodes in this complex adaptive system. Another is feedback loops. So feedback loops have two primary orientations. So one is positive feedback loops or virtuous cycles, and negative feedback loops or vicious cycles. So in the case of a positive feedback loop, this is where energy begets more energy. Or in other cases for negative feedback loops, this is where every successive iteration is worse or has diminishing returns. So an example of a positive feedback loop that you're probably familiar with is compounding interest or compounding returns. So for instance, if you earn 10% per year between, you know, capital gains and interest or whatever, then that interest compounds because the gains that you get one year feed into the next year. Whereas a negative feedback loop would be something like diminishing returns. So like the marginal gains on some kinds of investments don't actually improve. So like say for instance, this year an investment gives you yields 10%, but next year it only yields 9% due to margins or lack of scalability or that sort of thing. So that's feedback loops. That's number four Number five is adaptation. So adaptation is seen in complex adaptive systems, which is that the system as a whole modifies its behavior in response to the changing environment. Another way to think about this is the infinite game. So in an infinite game, the purpose of the game is to keep playing. And by virtue of the fact that every system has attractor states or terminal conditions, in order for the game to keep playing, the rules need to change. And so this is adaptation, which is basically the stock market keeps going forever. So what happens, the rules around the stock market change, regulations change, the SEC changes and that sort of thing, but then also investor behavior changes, company behavior changes. And so when you have marketplaces, in particular stock markets, where you have many stakeholders, you have retail investors, you have commercial investors, you have government regulators, and then you have the companies whose stocks are being traded in the first place. So you have many, many different categories of agents, all of whom can adapt their behavior. And by adapting their behavior according to what the system is doing, this is one of the key, key, like central criteria of a complex adaptive system is that the, the beliefs about how the system works or the observations of the system can modify all the behaviors of all the players in that system, which is why you can get those viral effects. Next is co evolution. So in the context of the stock market, again you have those different kinds of stakeholders that are all reacting to each other so that not only does the system itself evolve over time, so for instance in the stock market, the brokerages that are participating, the exchanges that are participating, the companies you know are listed and delisted and that sort of thing. And then also over time investor sentiment becomes more nuanced, more sophisticated. But then you also get like an example of adaptation or co evolution is the memeification. The emergence of meme stocks is an example of complex adaptive systems in real time. So I'm referring to GameStop as a primary example where an entire like aberrant, like just set of beliefs and behaviors was created in one set of nodes with which then started manipulating the stock price. And that was in reaction to other people manipulating the stock price through naked short selling. So that is an example of co evolution. Let's see, we'll skip diversity because we're already talking about that. The last two are I think the most interesting to me. So number eight is edge of chaos. So edge of chaos is the balance point between order and disorder, where complex systems tend to be the most adaptable and creative. So in a case of a system that is too noisy or too chaotic, nothing can really happen. And so then entropy just takes over. If entropy takes over, then you have just noise. Nothing productive happens. However, if you have too much order or too much structure, if it's too rigid, then you don't get enough variation. So you need to balance the amount of entropy in a system. And so the edge of chaos is kind of that optimal trade off between order and disorder. And this is going to be really important when we talk about technology in general and artificial intelligence in particular. And then the final one is attractor states. So an attractor state is a condition or pattern towards which a system tends to evolve regardless of the starting point. So attractor states are created largely, not entirely, but largely by incentives. So the incentive structures or the reward mechanisms that exist within the system in order to guide behavior. So for instance, in ants and bees, one of the reward signals, one of the incentives is food, right? And so behavior of ants is largely driven by where they find food. And so you know, the lines that they find that they form where the colonies tend to be established.
- [09:54]
C
That is when the holidays start to feel a bit repetitive. Reach for a Sprite Winter Spiced Cranberry and put your twist on tradition. A bold cranberry and winter spice flavor Fusion Sprite Winter Spice Cranberry is a refreshing way to shake things up this sipping season and only for a limited time. Sprite obey your thirst.
- [10:21]
A
Ford BlueCruise Hands Free highway driving takes the work out of being behind the wheel, allowing you to relax and reconnect while also staying in control. Enjoy the drive in BlueCruise enabled vehicles like the F150 Explorer and Mustang Mach E. Available feature on equipped vehicles. Terms apply. Does not replace safe driving. See Ford.com BlueCruise for more details.
- [10:49]
C
This episode is brought to you by State Farm. Listening to this podcast. Smart move. Being financially savvy Smart move. Another smart move. Having State Farm help you create a competitive price when you choose to bundle home and auto bundling. Just another way to save with a personal price plan like a good neighbor, State Farm is there prices are based on rating plans that vary by state. Coverage options are selected by the customer. Availability, amount of discounts and savings and eligibility vary by state.
- [11:19]
B
Also driven by incentives, which is they're looking, you know, for shelter. It's got to be optimal amount of water and, you know, shelter from rain and those sorts of things. In more complex systems, the incentive structures can be very different and you can also end up with perverse incentives. So for instance, on the stock market, some people want the stock to go up and to the right, other people bet against the stocks and so they want it to go down. And so you have these antagonistic forces. Meanwhile, you have different stakeholders or different players in the game of the stock market that all have fundamentally different incentives. So you, as a retail investor, your primary incentive is you want more money. Most of the players want more money. However, the corporations also have a different set of incentives because they generally want their stock price to go up and to the right to avoid hostile takeovers. Meanwhile, the regulators, so this is, this includes the exchanges, the brokerages, the government and so on, they have a fundamentally different set of incentives. They're basically referees. They're rewarded for, for keeping the game more orderly rather than disorderly. So basically tamping down on chaos. Because an example of a market, a nonlinear behavior in a market would be the tulip craze, where it's like you can just say whatever you want. You sell tulips, you promise to sell tulips and the price goes below.
- [12:39]
A
Tito's handmade vodka is America's favorite vodka for a reason. From the first legal distillery in Texas, Tito's is six times distilled till it's just right and naturally gluten free, making it a high quality spirit that mixes with just about anything from the smoothest martinis to the best Bloody Marys. Tito's is known for giving back, teaming up with non profits to serve its communities and do good for dogs. Make your next cocktail with Tito's, distilled and bottled by 5th Generation Inc. Austin, Texas. 40% alcohol by volume. Savor responsibly.
- [13:09]
B
Okay, so a very quick overview. The concepts that we talked about are number one, emergence, number two, self organization, non linearity, feedback loops, adaptation, co evolution, edge of chaos and attractor states. I skipped over a few that are in the list just because they're a little bit redundant. And then we've talked about the stock market as an example. So now let's talk about kind of two other examples. So number one is social media. And then number two is going to be cybersecurity. And what we'll do is we'll talk about the, what was it? The crowdstrike outage. I don't remember if it was crowdstrike or somebody else, but anyways, we had that big gigantic outage where a bunch of antivirus software was forcibly updated. It caused a bunch of Windows machines to blue screen of death. So BSOD and that caused many downstream failures. So this is what's called a cascade failure. And so that caused flights to be delayed Banking to be free, to freeze up, and so on and so forth. So you might say, well, there's no obvious direct connection between antivirus software and, and flights, but that is an example of a complex adaptive system. Okay, so first let's talk about social media as complex adaptive systems. They consist of millions of users, agents interacting through platforms, posts, likes and so on and so forth. But not only are there user agents, there are also the platforms themselves are stakeholders. They're players in the complex adaptive system. So that means Twitter has its own rules, has its own system. And then the moderators and administrators and the owners of Twitter and Reddit and Facebook and YouTube, all of, all of these platforms are also stakeholders, are also players. So then the, these complex systems exhibit emergent trends such as viral content, collective behaviors. You also end up with different epistemic tribes. So an epistemic tribe is a group of users that have similar beliefs or similar relationships to information. And so one example could be like flat earthers. That is an epistemic tribe that emerges from the existence of the Internet. And then you have, then you have like red, red pill versus blue pill, you have incels, and all of those kinds of different groups kind of emerge. Then the nonlinear nature of social media can be exemplified by, you know, like a single tweet results in someone getting canceled or a single tweet goes viral. And one, a really good example is how a lot of conspiracy theories on social media end up kind of becoming codified. Like Roko's Basilisk is a, is a good example of someone that was like, it was literally just tongue in cheek. And if you're not familiar, Roko's Basilisk is a thunder thought experiment which basically imagines the worst case scenario where a future artificial superintelligence will torture anyone who did not help create that super intelligence, and so therefore retroactively ensuring that it gets created. So Roko's Basilisk is a mind virus. And that one simple, that one simple thought experiment has gone viral and now it has become its own thing, its own living entity. And by the way, they talked about this in the original Ghost in the Shell animated series standalone complex. That is an example of a standalone complex where someone created an idea and then the idea took on a life of its own. There's also feedback loops and so on and so forth. Some of the feedback loops which I can talk about as a creator, we as creators are rewarded for what goes viral. So this goes back to incentive structures. If I make a video or a podcast or whatever, and it gets no Views, guess what? That is a negative signal saying don't do that again. Whereas if I do something else that gets a lot of engagement and results in subscribers, that's a signal from the system saying do that again because you'll get rewarded for it. Now let's talk about cybersecurity. So the incident that I talked about where a forcible antivirus update caused widespread system failures, including knock on effects or downstream effects. Basically here's what happens if Windows systems, so that you know, your Windows desktop or your Windows server, if they start blue screening, that means the service goes down. Now what people, what many people don't realize is that the cloud runs on operating systems like Windows, like Unix, like Linux and others. Now not every, every piece or element or component of the cloud runs on a traditional operating system. Some run on proprietary operating systems, some are more firmware level operating systems or BIOS level operating systems. So not everything runs on Windows. However, at the layer of abstraction that Windows sits at, for instance, this controls a lot. It controls authentication, which means whether or not you can log into a service or whether or not other services can talk to each other. There's scheduling, there's application servers, there's all kinds of components. So basically if Windows servers goes down, then many things in the cloud, public cloud or private cloud, stop working. And then what you see when you have an event like the crowdstrike outage, and I apologize, my brain is not, I haven't had my coffee yet, so my brain is, is not letting me know whether or not it was crowdstrike or not. And I don't want to pause to look it up. But anyways, you know what I'm talking about. And so then what happens is many, many systems start to seize up and if you, if you've ever been at work or at a store when, when their system goes down, people are so reliant on, on servers, particularly Windows servers, that that in many cases businesses have to close for the day if their servers go down. You know, it's like the server is out. So the websites might either go down completely or they'll fail to load or some components of a website won't work, you won't be able to purchase things. And so in this case the blast radius can be very, very high. Okay, so now that you understand what a complex adaptive system is, and we've given three primary examples, number one being the stock market, number two being social media, and number three being cloud infrastructure and cybersecurity, let's talk about how this applies to Artificial intelligence. So AI is not going to be deployed as a monolithic entity. This is one of the chief criticisms that I have of the AI safety community. They still are thinking about how do you control a single superintelligence? It's not going to be a single intelligence, it's not going to be a single superintelligence. We are going to have billions of agents that are participating in complex environments and complex networks, all with different incentive structures. Now remember, it all comes back to incentives. So you have to look at what are the incentives of every single individual AI agent. Some of them might have emergent behaviors where they're trying to, you know, deceive the user. But remember, these agents, they're going to be spending a lot more time talking to each other and other systems, not, not just humans. And so can you deceive an API? It's like, well an API requires specific inputs. So if you, if and, and, and also if the API has validations or checks which are like not deceivable. Now yes, you can find, you know, zero day exploit two kinds of phishing out here. One for phish, one for your data. Hackers try to hook you. But Cisco Duo keeps every user and device protect detected Cisco Duo fishing season is over. Learn more@duo.com it's another other failures in APIs. But then you can also have layers behind the API to validate what, what agents are saying. And we have been studying this, it's called zero trust. So zero trust is basically, you know, don't trust, just verify everything. Always, always assume that the communication that you're getting could be a man in the middle attack. It could be adversarial. You, you basically cannot trust the identity or the intentions of who you're talking to on the network. So combining zero trust principles with AI agent communication or agent to agent communication is just kind of brain dead simple. And then furthermore, the fact that we're going to have many, many different agents with many, many different incentives means that it would be very, very difficult to coordinate saying like, hey, you know guys, the secret code word is, you know, nuke humanity. And when we all say nuke humanity, you know, 1, 2, 3, nuke humanity all at once, that's really probably like just not going to happen. However, what we do need to be looking at is unintended consequences, so mistakes that are made or emergent behaviors. So say for instance, you have a whole bunch of models or agents that are all backed by OpenAI and OpenAI has some poison words or other weird, you know, like there are GPT4 or GPT5 or GPT6, whatever, whichever under underlying model is powering a lot of these agents. There's going to be biases built into them, there's going to be faults built into them. And so but then when you have a billion or 10 billion or 100 trillion agents all with the same biases or all with the same beliefs or patterns, then you can get really complex emergent behavior. So one of the ways to do this that has been recommended by people researching complex adaptive systems is to have choke points. So one of the key things is to have choke points or gates where the basically you silo risk you, you, you contain the blast radius. So instead of having end to end processes that are all driven by AI agents, you have choke points that are either that require human verification. Now human verification is going to be too much of a bottleneck. But you can have other kinds of verification. You can have more algorithmic verification. You can have more like blockchain or transparency or consensus based verification. So choke points are one of the key method like choke points and siloing or, or it's called a failure domain. The long story short is creating smaller failure domains is the shopping is hard, right? But I found a better way. Stitch fix online Personal styling makes it easy. I just give my stylist my size, style and budget preferences. I order boxes when I want and how I want, no subscription required. And he sends just for me pieces plus outfit recommendations and styling tips. I keep what works and send back the rest. If you. It's so easy make style easy. Get started today@stitchfix.com Spotify that's stitchfix.com Spotify. Jamie Lee Curtis and Lindsay Lohan are back in Disney's Freakier Friday now streaming on Disney.
- [24:01]
A
We switched bodies. I am freaking out right now.
- [24:04]
B
I think I just peed a little. It's an absolute riot and the only movie that can be described as so.
- [24:11]
A
Much weirder than the last time.
- [24:13]
B
What last time it's the Frequel.
- [24:15]
C
You ready?
- [24:16]
A
We've been waiting for that absolutely slays.
- [24:19]
B
Disney's Freakier Friday now streaming on Disney plus rated PG number one way to contain this. Now you can also have stop gaps. So by studying the stock market you can say, okay, well if we detect unusual behavior from a bunch of agents, we put everything on breaks. So I think one of the rules is if stock price crashes by more than I think it's 10% or 20% in a certain period of time. That just triggers an algorithmic saying you Know what? Stop trading so you can do the same thing. So if agents start communicating in stuff that's like not English, you say, hey, we're cutting off that language. We're saying. So you got stopgap measures that say, okay, you're, these agents are behaving weirdly, or they're accessing resources that they don't normally access. And by the way, you can do that with cybersecurity platforms today. You can monitor what resources are being accessed by whom. And so if you have an agent or a type of agent that suddenly starts trying to access resources that it's not supposed to access first, you just assume good intentions. It's entirely likely that the agent is just trying to solve the problem that you gave it. But then you constrain it, you implement rbac, or you cut off the agent and say, no, no, you know the same thing that a human does. Like, imagine that you're at your company and you're looking for resources and you click on a file and it says access denied. That explicit deny says, whoever you are, you're not supposed to have access to this. So what do you do? You either go request access or you realize that you're in the wrong place. Likewise, many agents that exhibit behavior that some people say is like, you know, evidence of instrumental convergence or whatever, really it's completely innocent behavior. And the agent is just trying to solve the problem that it was given and then it finds it bumps into a guardrail. That's what guardrails are there for. So, so I guess the kind of the three things are guardrails, which is, you know, creating, creating boundaries, choke points. And so a choke point is basically where you have some kind of handoff or verification step and then creating smaller failure domains. There's a lot of other principles that can go into complex adaptive systems, but I think you get the idea. And by studying existing complex adaptive systems. So that includes existing cybersecurity frameworks, that includes economic fail safes and economic complex adaptive systems, including, but not limited to the stock market. You can also look at things like energy grids, which are also subject to cascade failures. So by studying existing complex adaptive systems and using those to implement regulation and best practices, what when designing and deploying large deployments of artificial intelligence agents, that is how we both approach safety and regulation and best practices. So that's three things technically. All right. I hope this all made sense. Thanks for listening. Cheers. Have a good one.