Aligning AI with AI-Commerce - Tokyo 2024

okay for our next talk we have um aligning AI with AI Commerce and uh we will have Don gosson joining us welcome Don thanks for making that announcement right before I go on everybody get out of here go eat good just here okay don't move all right I can move within this box excellent all right uh Heim y uh thank you for inviting me um to the to the uh hosts of this event funding the commons I was talking about this topic aligning AI with AI Commerce um with David a few weeks ago it's super novel it's entirely hypothetical it's not even theoretical at this point it's basically something that I've pulled out of my butt but I'll talk to why um I was I did that so first off who am I and why am I talking about this my name is Don gson I'm the co-founder and CEO of a company called nevermind we are hyperfocused on AI payments um my background this is basically the culmination of 20 years worth of work um started in Big Data in ml and have advanced through to adding web 3 to that topic um um I'm I guess OG web 3 and AI I co-founded one of the first projects in this space called ocean protocol back in 2017 um and yeah this is sort of the synthesis of the last decade's worth of work um so what do we believe it never mind we actually believe that what we're witnessing today is the rise of a new consumer that will manifest is trillions of AI agents and that these agents are going to change Commerce Forever This is a concept that we have termed AI Commerce um and we believe that it's fundamentally imperative to address the payments mechanism between AI agents in a decentralized fashion uh we see payments between AI agents in particular as being a singular choke point for for all agents um if you know if the payment system is monopolized by one or a handful of organizations like open AI or Microsoft or Google that their ability to deplatform Any Given AI agent will be an existential threat or pose an existential threat to all agents um doesn't matter if the rest of the AI stack is decentralized if it can't transact within an economic ecosystem it might as well not exist um so we believe that decentralizing the payments piece is imperative um what we believe right now we're seeing is it's very reminiscent of the early to mid 90s um for anybody that was around for sort of the early adoption of the internet the way that we're transacting is kind of comparatively uh laborious so back then it was find somebody that you wanted to trade with um through a forum or uh um uh you know um the the numbers of of lists that were being shared Etc you'd negotiate through something like icq or email um and when you came to a price you would hope that that person would send you the thing and and at least in North America you play in a process called cash on delivery where You' give I think he had that in Japan too where you would pay the uh postal worker the cash and that'll get transferred to whoever sent you the item um so and it wasn't until to like PayPal came along um that linked email with bank accounts that uh Commerce on the internet AKA e-commerce actually started to scale so what we've been thinking about for a long time is is making the equivalent to PayPal for AI and being the engine behind uh AI Commerce who in the audience is familiar with AI alignment okay so crash course um basically it's the attempt to ensure that an AI system um is going to its objectives are going to match those of of the designer Andor the users within the system or at least some sort of overarching uh ethical objectives that have been set the ways that this is currently being done is it's kind of um a a brute force uh means of of aligning AI you're doing it at the model level um there's three predominant or dominant means of of aligning AI at the moment one is through ethical guidelines so you're imposing or codifying ethical Frameworks within the model itself uh you've probably experienced this with like uh GPT or diffusion models when you request like not safe for work content and you get back a response that basically says no I can't do that or um maybe you experienced the early Renditions of Google Gemini and their image generation functionality and asked for like an image of George Washington and instead of getting an image of a 6 foot Tu breed head back you got something completely different so this is sort of imposition of ethical standards um at the model level there's also human feedback so I think today there's been quite a few discussions and some of have touched on things like reinforcement learning with human feedback that's keeping humans in the loop where basically you have this externality that provides inputs into the model when it performs well or when it performs poorly and the model address its um reactions to that there's also model constraints so this is basically imposing like decision logic into the model or adjacent to the model in actual fact um so this would be like telling an autonomous vehicle to stop at red lights and it does so the third is what we're looking at you know potentially deploying which would be economic reinforcement and this is less at the model level and more at the agentic AI level and this would be using economic incentives or penalties or Market forces to actually influence the agents themselves um so the question becomes can we align AI agents via economic incentives Andor penalties that are going to work uh in I I think collaboration with the the more direct alignment activities that are currently taking place with the goal being at least in our mind yeah using AI Commerce as a as an AI alignment driver so what do we need to understand first um well we kind of have to understand how these systems work so anybody familiar with multi-agent systems and agents themselves all right good I'll give you a crash course in 30 seconds so multi-agent system is a synthetic environment uh consisting of more than one um actor that works together in in concert and an agent is something that can perceive its environment um and then use those puts to perform directed behavior and usually it's done in an autonomous or a semi-autonomous fashion I think the major constraint in all of this hypothetical outlining is we're assuming that all agents in this economic environment are rational actors if they're not rational then um most of this logic goes out the window so what do we need and what do we have and what do we need so we've got the assumption is we have carbon based agents us uh we have silicon based agents the AIS um we're going to need some kind of communication protocol and then we're going to need incentive mechanisms penalty mechanisms and some means of signaling and discovering what's out there likely through a Marketplace or marketplaces um and imbued Within These marketplaces are probably uh regulatory constraints and some kind of monitoring interface who is familiar with Smallville so this is a study that came out of Stanford last year a a group of computer scientists created this experiment they created 25 AI agents gave them a bunch of personalities and then set them loose in this ecosystem simulated uh world for two days um asked the a bunch of questions after the fact and got a bunch of really interesting answers um so the architecture of these agents was basically what you see they had the means to perceive but they were also imbued with the capacity to memorialize the those perceptions into memory and then recall that memory which they can then use to plan and reflect on their activities um and those influence the actions of each of these agents some interesting activities ended up taking place one of which was like instead of these agents would go to lunch right and instead of going to a cafe they would go to a bar um so it was also almost like they were day drinking um and this lacked you know outside influences from say a commercial point of view where you know if we all went to a bar for lunch and got ripped and then went back to work we'd probably get fired um so you know what we're talking about is is actually introducing these types of economic incentives where in this case you get fired as a penalty for performing certain actions in the system um I think this is quite a novel approach uh perhaps more harmonious um and and possibly easier to influence agents than through direct management at the model level um I think we can also Leverage existing uh precedent so I mean I think it's pretty appropo that I'm delivering this presentation at the you know un University because this is right up the un's alley um but what are we actually talking about what are the patterns that we can leverage to uh enforce these economic um uh patterns obviously some form of international relations um you know Aid uh or trade to influence certain behaviors uh the alternative to that would be trade sanctions and embargos uh you perform a certain way you're going to be removed or isolated within the system at a more discreet level you know we have systems like labor unions and employment um that ensure certain behaviors within given constraints and then more recently uh obviously from a from a decentralized point of view there has been the emergence of dows and decentralized governance all of which I think we can take cues from to help influence the way that we align AI um just looking back historically you know we can go back Millennia to look at discret explicit examples you know Athenians Banning marara Traders um within their marketplaces and basically annihilating the marara city state in the process Napoleon had the continental system which uh tried to block trade with Great Britain um Marshall Plan etc etc so what can we do well we can look at the incentives um what we're talking about here is likely rewarding these agents either agent AI agent to AI agent rewards which are already emerging in um video games games and uh in the web 3 space gamei you've obviously got direct compensation so paying for certain behaviors to be invoked by agents as well as general uh gamification through tiered incentive structures I think the advantages here is that uh the major one being you introduce probably proactive uh beneficial Behavior the disadvantages incentive programs for anybody that's um been involved in like uh points mining in the crypto space you know that this stuff can get gamed pretty easily um then there are the penalties obviously you can introduce fines and sanctions for certain Behavior you could restrict access um based on certain behavior and you could introduce reputational damage if an agent acts in a certain way um the advantages here obviously you can employ strong deterrents that are going to um oppose sort of harmful and unethical Behavior the disadvantage of these types of systems is that they lead to likely adversarial Behavior I think we see this at a macro uh level from a from a um you know a political standpoint there are examples of this already taking place um the State of Arizona worked with Tesla on aligning um autonomous vehicles and uh following you know proper uh driving protocols the motivation there was to reduce uh insurance premiums um we've seen this in hrrp in the United States which is um H Hospital readmission Reduction Program so these are penalties if patients get um um if they return to a hospital after a diagnosis algorithms have been deployed um in order to combat this and reduce those penalties and then more recently we've seen the emergence of personal traders that are basically AIS that get a commission on the returns as well as AI content creators so um making content go viral making it stickier Etc in order to have higher retention on sites that are content related um challenges that we need to consider in all of this obviously in there are going to be a lot of challenges in designing the incentives they need to be fair they need to be effective but they also need to be practically from a technology they need to be practical from a technology point of view we want to avoid any collateral damage or unforeseen problems that are going to come out of this uh economic alignment um system obviously there could be a lot of unintended consequences um we want to balance the short and long-term objectives of both the agent and the overall system and then there are obviously the ethical considerations we do want to you know limit biases Etc so I'll leave you with this what do we need to happen next uh from from our point of view there needs to be uh a focus on commercially viable agentic AI so we need to start producing um AIS and agentic AIS that are productive in nature we're seeing this start to emerge um once that happens imbuing these agents with the capacity to pay and get paid will then give us the means to start to invoke these um commercial incentive schemes and lastly I think what we also need to do is recognize our limitation on the influence that we're going to have with these systems in all likelihood these AIS are going to operate at uh you know on a dimension that we're just ambivalent to um so recognizing that probably the best we can all do is ingratiate ourselves in the system and not try to control it is the way to go and I've got a minute for questions so opening up the table I don't know if you need a microphone do I go to him I think he is a wonderful sharing um it was really interesting to see that using the economic incentives to like um strs in the AI alignments and I have a few questions about how to setting up these kind of penalties or the instances because you know that there's a lot of different values uh that might have contradicted with each others or maybe you have to have priorities on each other so how to set up these kind of standard Fair standard or very nice standard to like decide it who can decide it that these kind of behavior should be finded or these kind of behaviors be stren with the kindon incentives good question my answer is I don't have an answer yeah I think uh like we can we I think we can look at crypto in a sense to see how you can govern like um the chaos of the masses through potential tokenization and align incentives that way that's one means of doing it uh I think the the concern is that like biases get baked into the system based on who has like the the largest economic influence on these agents right um yeah at this point I don't have an answer I mean this is something we've only been thinking about for I don't know a a quarter or two but if you have an answer by all means let me know anything else all right I don't got the go thank you so much thank you