EthCC 2: Daniel Kronovet

you know they say that if you're nervous in front of a crowd you should imagine everyone as their 20-byte hexadecimal etherium address so it's working bonjour je m'appelle Daniel Ochefu je parle du budget box hello everyone my name is Daniel and today I would like to talk about budget box 22 slides of applied epistemology so this talk is based on a paper that was written along with Aaron Fisher and Jack do rows two of my colleagues at Colony it's called decentralized capital allocation via budgeting boxes it describes a budgeting mechanism an algorithm that we developed to give out money in a decentralized setting you know questions about funding the ecosystem are really big right now in the in the etherion blockchain space and so this paper was a response it was our attempt to answer that question and this talk is going to be sort of an exposition of the ideas in the paper but with a little bit more of a focus on the theory I mean you can read the paper if you want the needy the nitty-gritty details and this is gonna be more of like a fun big-picture theory and we'll slowly work our way towards the concrete as the talk progresses um can everyone hear me okay louder all right I'll just I'll just speak louder okay so we're gonna begin with a little theory a little information theory and so this is this is my favorite result in information theory it's called the data processing inequality and what this basically tells us is that if we have three random variables that exist in a sort of a Markov relationship where a causes B & B causes C then a and C are conditionally independent when you condition on B what that means is that everything that we know about a in C we know through what is captured by B or put it yet another way the amount of mutual information ie how much C tells us about a can never be greater than the amount of information that B tells us about a this is called the data processing inequality because the idea is that if you have a data set and you process it in some way you may be making it more useful but you're not actually increasing the information beyond what was in the original data set so this is like a very interesting result that comes out of information theory and when things get kind of interesting is when we we sort of take this and metaphorically understand the general task of science through this framework and if we have the world which is the sort of complex some would say like infinite object that we want to know about right we want knowledge of the world but we cannot access the world directly we can only access the world via our measurements so we take measurements we get the height of a tree we give someone a standardized test and we try to we try to then take the world that is not representable and represent it in some symbolic way via measurements and once we have that we can then process it we can then do analysis on the data and that will give us something that is useful something that we can make a decision about and kind of if we look at the the kind of the roles involved we see that that first arrow measurement is really the domain of basic science so a lot of what CERN does is measurement right they tried to measure the physical world at increasing degrees of specificity what you know educational boards news they develop instruments to measure the aptitude of students with various degrees of success and the challenges are representation how do you represent the thing that you care about and then noise you're taking a measurement and that measurement is going to be a mixture of some signal and some noise and that's sort of just that that's just that's how that goes and on the other side we have processing which I would say is the domain of statistics machine learning but but more broadly and so the challenges there are are really expressiveness and complexity right you want your analysis to capture as much information as it can and you also want it to be simple enough that you can execute it in a certain amount of time right we talk a lot of machine learning about you know this algorithm doesn't scale like traveling salesman it's an incredible analysis but its complexity is such that it's actually very hard to put into practice and so the people that are doing machine learning statistics are thinking about these problems and then together measurement and analysis measurement and processing we have a way to make sense of the world I guess I call it a meaningful example just to make it really concrete is an average right we have the world which I represent with Omega and this is something that we care about we take a number of different measurements of it each one slightly different because of the inherent noise and then we take the average and we say the average is that is the thing this average is the is the representation of the thing I care about that I'm going to do something with I can make a decision about it and the point is that the average contains less information than the data set does right you lose information about variance who is all the higher level moments of the data but it's much more useful right so you lose information but you gain utility or usefulness right and so just kind of underscore the point all of our processing does not increase information it makes information we have more useful I turned it into a form that we can consume and make actions about but it's not information generating right and so kind of kind of where I'm driving this you know if we see that as we do analysis we lose information but we sort of gain actionability and we gain usefulness if we can find better measurements then we can actually get usefulness basically for free right better measurements plus the same analysis give us in the end a result that's just more useful it's captured more information about the world we can make better decisions using this new results so sort of the search for for better measurements leads us to better better outcomes and that was kind of the point I wanted to drive here but there's there's a little bit more there's a little bit more to the story you know I'm talking about the importance of measurement and developing better measurements better representations better symbols that's really not gonna it's not gonna get us that's not going to get us there in the in a really in a real way and um so these people are here we have alfred whitehead and here we have curt girdle who were two mathematicians in the twentieth century kind of what they did was alfred whitehead along with his student Bertrand Russell tried to really put math they say put mathematics on strong foundations they really wanted to get mathematics what they saw as the perfect representation of the world the best representation of reality to be as complete and perfect as possible it had quite a lot of work on this and then what Kirk riddle did was he came along and said actually that's what you're doing is impossible there's really no way that you can ever get a set of symbols that will perfectly capture the rules as it is your set of symbols will always be incomplete there will always be a hole or a gap that exists between the worlds as it is and the symbols you've chosen to assign to the world there's always gonna be a gap no matter what you do and so like a lot of people know the story you know I put I put Douglas officer here who's a cognitive science who wrote a book called girl Escher Bach and eternal golden braids 979 won the Pulitzer Prize and he talks a lot about this most of my knowledge of this comes to via him really like a kind of like a really wonderful like wonderful thinker very deep very deep guy I think that the thing that happened next is really even more interesting is gonna be relevant for what happens next in this talk where's that Alfred whitehead basically changed careers you know after after this this thing he devoted himself to just was proven to be totally false he then became a philosopher and he developed a school known as processed philosophy where he basically said you know what I think this whole idea of representing things in a static fixed way is like the wrong approach because the essence of things has changed the essence of reality has changed you wrote a book called process and reality we're basically said we can only really think about things as being in process nothing is ever really fixed and when we represent something as fixed we're gonna represent it slightly wrong this was this was a big insight you know the Greeks say you never step in the same river twice it's very much the same idea and so I kind of made this slide which I thought was a joke but I think it's my favorite slide of this talk you know so kind of what are we striving for in a right representation right measurement right processing and that will give us that we can then use to make good decisions about the world so that's kind of that's kind of the theory and now we're gonna talk about I guess the information practice how we take these ideas and we build essentially a budgeting algorithm that we think has some nice nice properties let's leave it at that um and where we begin going back to measurements is with what I call the the minimum viable measurements which is really just a pairwise preference for those not familiar a pairwise preference is really just a choice between two things you know given two things right you can say do I want water bottle or a phone and i say i want phone that's basically all that it is and there's a there's two things about this that i really like one it's the it's the minimum vial of measurement in a very technical sense and that you were presented with one bit right there's no measurement that you can actually represent in less than that right there's no unit information that's smaller than a bit so in a sense of pairwise preference is the smallest thing that can be represented you know and also that sort of makes it the atom right in a sense you could you could really think of anything that you as representable via pairwise paraffins right even like a string of bits is a series of pairwise preferences right all the data on the internet is just a bunch of pairwise preferences in a sense between zero and one so i really you know if we're looking for the best measurement a place to start would be with the smallest measurement and then the other thing that I like about the pairwise preference is that it kind of represents subjectivity in a way that I would say other measurements do not right now I sort of have this like really personal irrational like prejudice against representing subjectivity in a numeric way like I find like 5.10 point scales look like all these scales I like I like I don't like them like they have a upset me because you're taking this thing that is subjective a feeling and you're trying to fit it into a number right and you you can do it and people do do it and it works fine it's not a bad thing but I'm gonna say that it's not really the right way to think about what's happening and so when we measure something wrong you know that noise then goes upstream through all of our analyses and means that our result is gonna be a little bit less useful than it otherwise could be you know and I want to suggest that in as much as a preference between two things is subjective right I just want this instead of that and we can represent that as a bit we actually can represent something that is subjective with an objective symbol in a way that is really not as lossy not as noisy you know I feel this I represent it this way and we've really captured the thing in a way that I find to be like really neat the next thing I really like this slide is because I find types of mathematics to be like really aesthetically pleasing but there's also like this there's a point to this slide which is that you know we take all of these pairwise preferences we just take a whole we gather a lot of people's opinions about this or that this or that and then we aggregate them together and the idea so a semigroup is basically it's a set of objects and there's really one rule which is any two objects in the set can be combined together to make another object in the set right if you have x and y in the set s you can like mush them together to get Z which is also in the set s and if you if you like I guess in computer science terms it's like reducing this is the same thing right if you have a reducer you have a bunch of items in an array and like a reducing function you just like reduce them all same concept semigroups oh and the reason why it's important here is that we have all of these pairwise preferences we can kind of just add them together to create like a big big set of preferences representing a lot of people right I mean furthermore you can you can kind of scale them or tweak them to add additional information right so say I have all my preferences on a variety of things but then I have a reputation and so all of my preferences are then scaled by my reputation and then added with others preferences or you could say I know a quadratic voting is like a really popular thing in this space you could say that every preference is then scaled by some like quadratic voting input that I'm willing to allocate amongst my various options or you could say and this is the idea that I'm like the most excited about you could even say that the longer someone has held a preference consistently the more it retains a value right like we talk about politicians always flip-flopping they're flip-flopping their decisions but what if like the longer that I maintained this preference the more sort of acquired strength if I were to flip it you would like reset back to zero to have to start over right so create like a incentive to have conviction and a slight penalty to being a flip-flopper right you can you can sort of incorporate all this additional information and all of these additional rules into this like semigroup reducing process and then what you get is like a set of pairwise preferences the simplest unit of information but like a lot of them well like a lot of information with additional information about length of time reputation whatever you want so we really just have a ton of information it has been sort of accumulated and then we just and then the next step in the last step is we throw in some like mathematical machinery to basically take all of this information just can't like run it through some like linear algebra magic and just like get at the end or a result you know and so what's happening here is that this is a very very kind of standard set of classic set of techniques you know we have the graph here and actually I really like this graph because it's got this really beautiful motion about it if you start it you'll see it it's like you see the UH you look at all the kind of the pairwise preferences like who's winning who's losing and you kind of see like there's a motion where you know see is the loser but then see gives its preference to D and B and then D and C give its preference up to a and to B and then B gives all over to a right we take these pairwise preferences we just kind of like turn them into a matrix following some rules which represents sort of how likely one pair Israel is likely to win over another and then we take this matrix and we just take like you know the technical term is like the first eigenvector you can think of it as running these interactions again and again and saying okay if this is our likely someone just to beat someone else what's like the steady state what is the at the end like who's gonna be the winner based on these pairwise interactions this is a technique known as spectral ranking it's been around for a while and this is how we take all of the data that we've just accumulated all the information we've gathered and we analyzed and turned into a very simple result and the last step which is like kind of I think in a sense the simplest but in this context like maybe one of the more important if you just take this distribution over the outcomes which is really just a bunch of percentages which add to one and we just say that this is now a budget that we're gonna use to govern our organization right I really I have the Vidkun Steen duck rabbit thing here because nothing has changed we just we just see a probability distribution as a budget allocation and then we can move forward like nothing really happened here is just like a kind of a flip of the perception and so here we have budget box which is a simple flexible and powerful mechanism for measuring and processing financial priorities and I emphasize financial right those of you here from at ons talk saw that in the Dow stack sense they're taking a very general problem which is decision making writ large coming up with like quite innovative ways to accelerate that via approximation whereas the difference here is that were taking a specific problem financial decision making allocating budgets which is a much simpler than general decision making and developing a mechanism they can solve that problem taking in more information right so by adding constraints we're able to develop a specialized solution which is not as general but does allow for a broader based input so I would say that these solutions are quite complementary and orthogonal and that they approach the same problem in slightly different ways and so in conclusion you know we talk a lot about governance in this space and when we talk about governance we also talk about legitimacy how is our government legitimate and the thing that I kind of want to suggest is that by having a budgeting mechanism that can gather as much information as possible from the world we can really get a wide set of preferences we can incorporate a lot of arbitrary additional information into this set of int into an entire essentially our matrix then we can turn that into a budget this budget is really more legitimate than other budgets or other allocations just because it represents and it contains so much information right I mean I sort of like I think that I have this like little fantasy I kind of like looking at as far as far as this as far as this could really go I kind of I kind of really wonder if this is actually pulled off quite well and like you know we figure out all the right knobs to tune and the right weights and the right inputs and the right questions and all that we can actually create at the ends like a data artifact which is like thus it's like the source of truth right like this this outcome this distribution this budget this is the source of truth for what we want and that can then become the beginning of a lot of other things right so some was like the Carl Jacobi invert always invert instead of the budget being the outcome of a process it can be really the beginning of other processes because it is the legitimate thing this is like my dream it's not real yet but like like thinking ahead like that would be neat that would be a cool thing if that happens so to talk about applications there are there quite a few right like you know we said we wrote the budget box paper as a response the problem of funding open source and so you know the first application would be funding open source right a lot of projects we need to choose between we want a lot of people in a decentralized setting to be giving their input so people perhaps have reputation they give inputs about what projects they think they should relatively be funding we then calculate a budget at the ends we omit amounts of funding another thing that we could use is people talk a lot about participatory budgeting you know in municipalities and cities right we want people to be able to have input over how their money is spent so we could use this for a petition participatory budgeting where there's this much money all the people in the neighborhood can give us your input give a shrimper give us your input organizations so you want to say we have to allocate money between engineering marketing product we can have the people in the firm make decisions about whether we should spending money for this thing for that thing the other thing matchmaking you know like finding people that you want to work with or spend time with you can just say I would rather spend time with this person or that person like a lot of things like we put in the pairwise language we can use this technique for and I guess sort of like the bullet points at the end um you know simple measurements give us more information she's that our analysis more closely reflects the world which means it's more legitimate ease of aggregation analysis makes it possible to update in real time which means we can quickly incorporate new information this is something I didn't actually mention earlier and I should remember I mentioned the sort of Kurt girdle alfred whitehead every symbolic every symbol always being complete the way that we sort of respond to that you say okay if our fixed static thing is always going to be partial then what we need to really build in in the first-class way is the ability to make change right we know that whatever we do will be slightly wrong so it's building the idea to make change we can easily update in real time new information coming in we can update the budget it can always be kind of like a moving target they can always update new information and that means that we don't have to worry about getting it wrong budgeting doesn't have to be this thing that we have people do once it takes a lot of work and this is the budget we really have just like say we have to see that we can we can come up with a budget and then start to allocate the funding and then oh we realized okay we actually to like move a little bit more this way okay we can allocate a little more funding over here maybe a little bit less over here so it also becomes a way to provide oversight you know we feel like you're not really using your funding so well so we're gonna allocate some of it over here instead of over there it becomes a dynamic kind of moving moving thing instead of this like fixed thing that we have to really worry about getting right and the last thing you know instead of being this thing where everyone gets in a room and like talks about line items there's like political trade-offs there's like horse trading there's like well if you fund this off on that you know because only a few people can do it cuz it's really hard because something that's very easy anyone can just come in and say I really want to I think that we should be putting our energies towards this instead of that and it becomes a tool for a for a really direct oversight because it's a simple thing so by making something simple that was previously hard and many people can participate which really changes changes the experience of budgeting from being tedious to being I would I would dare to say potentially engaging and as I said before with the semigroups there's a lot of ways to extend to incorporate new information so it really it's kinda it's like it's it's quite structured but also highly flexible which means that there's a lot of a lot of knobs we can turn a lot of things we can do with it so that's my talk thank you so much and all right I don't think we understand for questions unfortunately thank you [Applause]