 Hi everybody, I'm Dazza Greenwood and I should probably introduce myself as initially a co-organizer of Massachusetts legal hackers and then subsequently when we combined Massachusetts legal hackers and Boston Cambridge legal hackers and another legal hacking group, co-organizer of Boston legal hackers and then when I lived in New York after volunteering for a while, co-organizer of New York legal hackers, I think I'm still on the page so you have to get rid of me there and now I've moved back to my hometown of Oakland where you can find me as a co-organizer of San Francisco legal hackers. Thank you very much, legal hacker and oh how I love legal hacking and so this is what we're looking at right now is a blog post on my blog which is civics.com and you can go there and find the post called Algorithmic Entity Accountability and why am I showing you a blog post for this flash talk? It's because sort of like opening the curtains to a interesting little legal hack that I've been contemplating lately and that soon I'm going to try to tackle and so I'd like to describe it to you and then see if anyone has any interesting ideas or questions because I have not pulled the trigger yet. So the background on this is that the state of Wyoming has recently enacted a new statute that allows for decentralized autonomous organizations to become legal entities, LLC specifically and I was honored to help a little bit on that statute and one of the most remarkable aspects of it is that in forming that legal entity organizers can in the operating agreement have to indicate the extent to which the entity shall be algorithmically managed and that seemed really cool to everybody when it was going through the legislature and then I started thinking well hold on a minute the extent to which it shall be algorithm that one extent an entity could be algorithmically managed is 100% it could be 100% algorithmically managed and that was completely consistent with what some of the advocates of the legislation wanted they wanted a completely autonomous organization that was a legal entity that was an LLC and that's enticing and interesting however it started making me wonder if this is really such a great idea because our algorithms up to the task of managing a legal entity and when you think of and when it was discussed before the select committee that devised the legislation the context was sort of smart contracts that were doing you know basically like investment funds like a classic Dow not an investment fund technically in the financial services sense but you know what I mean you pull your cryptocurrency you invested in projects hopefully you make a profit however in LLC in Wyoming and in any state could do business in any way it doesn't have to be like on the on a blockchain and in fact it doesn't have to be that form of an organization you know it could run a kindergarten or a dynamite factory or whatever a hardware store it could do just anything and so that got me thinking well what are the what's the nature of my misgivings about 100% algorithm and it really so what I looked at when I did on the legal side of legal hacking was I just started searching through the rest of Wyoming statutes to see what are the requirements for managers of LLCs in Wyoming turns out they were similar to Massachusetts LLCs and much of the US one of the requirements is that okay I'm sorry I shouldn't be just scrolling through a blog post I'll just tell you is that is that they have fiduciary duties of care and loyalty to the other members of the LLC and to the entity itself and I thought oh good now this is something I can get my mind around where it's something I associate really with what I would call human level intelligence human level cognition and judgment and one of the big things in AI which is I'd say very closely adjacent to using autonomous technologies on blockchain for organizations is this quest we've had for decades now for general level intelligence and it occurred to me this is an interesting way to describe some of the gap between what we can get now with even very like complex smart contracts and complex modern AI like large language models for example and what we would expect from a human so I started thinking you know maybe what we ought to do is shine some further light on this partly some of the legislature legislators are asking me well why not just give this a try we're a pro innovation state like what would why not let people experiment with completely algorithmically run legal entities and so what I'm planning to say during the next hearings which are in just under a month now when they're looking at amending the statute further is maybe what is to ask the question well how could we test how could we assess and evaluate measure the extent to which a the as such an algorithm could correctly comprehend and apply the relevant fiduciary duties it occurred to me one way I know how to test it is with a standardized test in the legal context it's the MPRE specifically and if if you're a financial advisor and if I do share a rule applies their standardized tests for the same duties there are standardized well calibrated like longitudinally tested exam instruments and so I thought okay well let me get ahead of this a little bit and I talked to some friends at open AI they run GPT 3 and Dolly and some other things in order to get a research API access to GPT 3 which is a natural language model and it's possible I'm still figuring out how which is probably why I want to expose this to everyone see if anyone had ideas Jack or others but I think I can create a test harness and basically apply just the duty of loyalty and care question set to through the API and see if it's if it can well this is where it gets complicated you can't do a multiple choice test in the API there's no such like get or post for that but you I think I could arrange it in such a way to see if it could autocomplete the right way or if I can basically get the difference if it can start to distinguish answers that would demonstrate that it is selecting the scenarios and facts that would reflect for example a manager that is not putting the their own interest or someone else's interest ahead of the interest of the members or the entity itself and so I've been slowly on the other thing I noticed when I was researching this is there's actually a bunch of people or there's a number of people that have started to apply IQ tests to AI and I thought oh I got some good ideas for how they created test harnesses to apply though to apply different types of IQ test to different different models and so I think this might be possible to cobble together a way in summary to apply a to apply a test to a to AI specifically to see whether or the extent to which it is capable of of demonstrating answers that would be considered passing grades shall we say for the MP for the something like the MPRE at least that subset of it I think my assumption is that when I do this what we'll see is it doesn't pass the grade and that's okay because I think some of the idea here is to ask the question where's the gap what is to be engineered next what would be minimally sufficient to rely more and more on this beneficial technology for business and financial purposes and something that we can build toward anyhow so that's the hack and I talked to my colleagues at MIT where I also have a post about exploring this there they thought it was a great idea and so over this fall and spring semester of this academic year I'm going to do some kind of research project that will be an open legal hacking project in collaboration with our friends in Boston legal hackers in San Francisco legal hackers to explore this and that means you my dear fellow legal hackers are all invited to this jolly romp and if you have any ideas I hope you will contact me I hope that you'll talk to me at lunch and see if we can find a way that's legit and where we can really learn something from trying this some interesting test thank you very much I ran the clock you are how do we participate how do we participate in January can we do it remotely and what tools are you using to allow us to do that I hereby promise that there will be some kind of way that you can participate remotely and so how that's going to happen exactly I don't know but it'll ultimately be available at law dot MIT dot edu and until we figure out how that's going to work if you go to civics.com to this blog post I know I can at least update this blog post with a link to how you can participate the best way to participate is just like grab me I'm at that table right there and I'll be at lunch and I'll be at dinner and let's just talk about it and you can participate please participate. Thank you members of the committee I'm Dazza Greenwood and I'm here in my MIT capacity today to if it's in scope and I should ask this of the of the chair first it's basically to provide a quick to let you know about some research that we're starting this semester on the Dow LLC statute that follows up on the algorithmic manager part of the statute it'll be short but I don't know that it's necessarily in scope for the identifier topic. Now this is a good time for that Dazza thanks. Thank you so very quickly we are launching a research project at law dot MIT dot edu this year this semester actually on the Dow the Wyoming Dow LLC project and particularly is to explore some of the issues and the options with the Dow LLC algorithmic manager part of the statute and it's especially in the context where some activities under this legal entity will be happening off-chain not completely under a smart contract and you know that will be happening in commerce or other parts of the economy so specifically we're going to be looking to test the capacity of some example Dow LLC algorithms to correctly perform fiduciary duties that are owed by managers of Wyoming LLCs to other members into the entity itself under section 17-29-409 of the Wyoming statutes covering standards of conduct for members and managers and the idea that we had was to take the standardized tests that lawyers for example have to take for to show competence to understand and correctly apply the fiduciary duties of care and loyalty which are the two that the Wyoming statutes require of LLC managers it's known as the MPRE the multi-state professional responsibility exam and possibly also the fiduciary rule exams that financial advisors are required to take and basically got an idea to create a test harness to to see how good or not good a job on some of these algorithms are able to do on these tests and hopefully learn from it and see well just basically see what we see what we expect is that there may be a gap between what some of these algorithms are able to do and what we would expect of a human manager of an LLC and that maybe there'll be some learning from that that could help be an input to the committee in further amendments down the line for the statute perhaps ensuring more human control for the types of activities that would require typically fiduciary duties or maybe limiting the scope of activities such an entity could do or or something else or maybe just kind of going for it in view of the fact that it's a new innovation that you want to encourage but anyway we'd at least like to start to clarify what these things can and can't do as against the criteria expected of humans for similar LLCs in Wyoming under Wyoming law and that's that's what we're doing we hope that it will be helpful to you and if you have any questions well I should say if you want to see the research it's actually on a blog post now at civics.com and if you click on the blog and scroll down there's an algorithmic entity and fiduciary duties post and soon in a week or two we'll have a research page at law.mit.edu as well we'd welcome any feedback questions or concerns the committee may have on this. Mr. Chairman, Mr. Co-Chairman. Thank you. First of all thank you so much for bringing that to your attention and that's a fascinating research and it does get into then a little bit about this this public available identifier concept and the fact that you're going to need to have something which is sufficient to link to that you can actually run that research on and I think it's almost a good test case circumstance where if you can't find what you need to find to run your research then it's probably an insufficient public identifier with regard to what we're looking for statutorily so I think we'll really look forward to hearing how things turn out for you on that and I'll stop there but we really appreciate your work on that Deza.