Sam Williams, founder and CEO of decentralized storage system Arweave, gives an overview of “Mechanism Design,” a field of study that has become newly relevant with the development of Bitcoin and subsequent blockchains that require carefully designed incentives for network participants. Williams uses examples to show that economic incentives, when designed properly, can persuade self-interested people to exhibit useful behaviors at fair market value with minimal central planning. This provides a new tool to bootstrap decentralized networks. He cautions, however, that poorly conceived incentive systems can overpower moral frameworks in ways that can be dangerous. This could be harmful, he says, in decentralized protocols, since self-executing code may not easily be altered to curtail unintended consequences. Williams closes with a case study of his company, Arweave, and the way it created an endowment-style financial incentive system to build a platform where data can be secured forever. This kind of model opens the door to new kinds of community-owned networks that can’t be manipulated by central owners.

Andreessen Horowitz’s Crypto Startup School brought together 45 participants from around the U.S. and overseas in a seven-week course to learn how to build crypto companies. Andreessen Horowitz is partnering with TechCrunch to release the online version of the course over the next few weeks.

Find more Crypto Startup School videos plus additional reading and info: https://bit.ly/2ThsHen

And today I'd like to talk to you about mechanism design, which was an obscure piece of literature kind of study in economics until around 10 years ago with the birth of Bitcoin and that's obviously extremely important to us as people building crypto systems but it's also not something that we were all trained in by default, I would say.

So this is essentially what you might think of as programming humans for engineers.

And this is based on experience and learnings that we've made kind of at Arweave along the way, and also before Arweave, so for the last 8 or 9 years, without, I guess, formal study.

So I was part of a PhD program studying and building decentralized operating systems before I got into the crypto scene.

And as a consequence, it was kind of necessary to learn all this stuff ourselves.

And I would say that most of the literature out there about mechanism design is focused very much on people inside the economics sphere by default.

And so really what I'm trying to do with this talk is give you a bridge from people that are very...

From the place that us engineers are very used to sort of inhabiting to somewhere that is much more kind of economics based.

And I can basically tell you along the way how things relate and also where you can go to read more and get way into the depths of how crypto economics and mechanism design work afterwards.

But really I would say this is the kind of missing link for engineers getting started in the crypto space on the economic side.

There's an introduction which takes you through broadly what mechanism design is, how you can use it, and a few examples both in centralized and decentralized systems.

And then there's an overview of how it can go wrong and the potential pitfalls with this.

I mean, we're talking about essentially programming human behavior.

And so there are a lot of moral and ethical questions that arise from this that I think we need to have a strong framework through which to, yeah, to discuss and act appropriately.

And then finally, I'll talk to you sort of more in depth about some example systems.

So we'll go through Arweave and also Bitcoin and see how we use these mechanism design approaches in practice to build systems that achieve a goal.

So, the fundamentals of mechanism design are programming human behavior through the careful design of incentives.

In this talk, I would like to posit the theory that essentially humans are just like machines that we would normally program, but instead of responding to and running programs in the traditional sense, so applications that are series of instructions that the computer must execute in order to be in line with a properly functioning CPU, instead what we have is people that are led by goals towards, well, through incentive systems.

In a nutshell, We see that humans are machines that want to achieve goals, and they plan their way towards those goals through sub-goals.

So 1 of these goals might be, I would like to go home after this and have a steak.

And as a consequence, as soon as I desire that, my mind starts to think, okay, well, what are the sub-goals involved in reaching that position in life?

We see that on the way to these meta-goals, we go through sub-goals, like with steak, for example, I need to go to a shop.

I need to buy the steak I also need to have a house in which to cook it, right?

So we see there are these sub goals and I Would argue that 1 of the most fundamental of those is money in fact money is the unifying sub goal for almost all larger goals in life.

And before money, we had essentially this mechanism where there was, in order to have a transaction, there must be a double coincidence of value.

So what that means in practice is that if I want to have a steak, I need to go, and I'm a carpenter, say, I need to go and I need to find someone that wants to buy a table from me that I've made and swap it for some amount of steak.

This is obviously a very, very inefficient system.

So 1 of the first mechanism designs that humans made was like, okay, well, what is a global generic sub-goal?

That is, everybody wants money, and as a consequence, we can freely exchange through money for other goods of value.

So I'm a carpenter, I can sell a chair to someone, I get money, I can take the money, and I can take it to a butcher and I can buy some food, and as a consequence, everybody is happy.

And This makes economies much more efficient.

So from our point of view, we can see, OK, if we can design a mechanism that gives people money as an output, then we can insert ourselves on the chain towards reaching basically any of those goals that those humans want.

So it doesn't matter to us whether they want a vegan meat-free burger or a steak or anything else or a chair, a house.

It doesn't matter because the sub-goal in common is money.

So if we can define a way of giving them money as a result of playing an economic game, then we can essentially push their behavior down the route that we're attempting to define.

We design a mechanism that's output is money, leading to eventually the goal of whatever that is that that individual actually desires.

And humans are so attached to this process of see goal, generate sub-goals, execute towards those sub-goals, towards the major goal, that this is an extremely reliable way of pushing behavior in a desirous fashion.

So by designing these economic games appropriately, we can push selfish humans, which are essentially humans that are playing what we would call the dominant strategy.

So the strategy that is most effective inside the game you've created to exhibit useful behaviors.

You get a human, you create a mechanism, this output is typically money, but it can be social reputation, we'll get to that a bit later, but in most cases it will be some form of money which allows them to get towards a goal, and as a byproduct of your mechanism, you have something that you find useful.

In my native Berlin, we have a bottle pickup mechanism design run by the government.

But basically, if you pick up bottles from the street, you can deliver them to drop-off locations, and you'll get around $0.30 per bottle.

So this is a way of giving the local population that don't have any other job a job by default.

But as well as that, it has this useful social byproduct of creating cleaner streets and also requires the local government to pay fewer street cleaners and centrally plan less.

So, 1 of the, I guess, common factors we'll notice when we're looking at mechanism designs in this lecture is that a lot of them really have this amazing attribute that they push to the collective brain of society, the question of how to play the game that you've created most effectively.

Here is an example of the Juicer economy.

So when these new scooter startups came around, they basically had this problem where, okay, so at the beginning of the day, we distribute the scooters across the city.

We put them in places where they're useful to people, and we make sure they're fully charged.

But by the middle of the day, all those scooters, or almost all of those scooters, are in some way that's much less useful and they have half charge.

By the end of the day, it's in complete disarray.

So, the early startups went down the route of saying, okay, so what we need is a central planning system whereby people can...

Yeah, we essentially hire enough people with vans to go and pick up the scooters, to charge the scooters, and then to redistribute them to useful places in the city.

But this is extremely, extremely costly and very ineffective.

It's actually another sort of example of the traveling salesman problem, right?

You have to work out, how do I go around all of these scooters in the most efficient way to charge them and then also redistribute them?

So the smart scooters companies decided, okay, well, let's just not do this.

Instead, why don't we say for every scooter that you charge and you move to a point that we determine, we will give you some reward.

And Then they found a way of marketing that opportunity to people, and as a consequence, people just came along and they would pick up the scooters, move them to the right place, and charge them.

There was no central planning involved, and it's extremely efficient.

If there's only 1 thing you take away from this talk, I would say is that you can take that same mechanism and that you can apply the same reward rules that we do in Bitcoin mining and in cryptocurrency mining of almost all kinds and you can make a mechanism that will give you a fair market value for any task like this.

So let's just imagine that I'm Bird, right?

I don't know how much I should be paying people up front for this job of distributing the scooters around the city.

Like it's actually quite difficult to determine what an appropriate reward for that would be.

So what I do instead is I say, okay, well I'll start off and I'll say, I will distribute $2, 000 per day for people to do this task for me.

And as more people join, they have this sort of competition effect, which lowers the cost of each of these movements and charges of the scooters until it reaches the opportunity cost, at which point that no new people will come to the mechanism and say yes this is a profitable thing for me to do.

But from this point as a startup, you've really essentially found the optimum pricing strategy for Having a service delivered for you.

It's extremely effective and it's possible to do in almost every situation But doesn't just have to be economic rewards.

BitTorrent worked with a mechanism design called Optimistic Tit for Tat.

It's extremely simple and extremely powerful.

What happens is, We play a game all together that says, I will give you data if you will give me data.

Then occasionally, we will give data to each other at random.

So that's, if you imagine, if only selfish players are playing the game, at rest when there is no more movement in the structure of the game, then what will happen is everybody shares data all the time with everyone else.

So this mechanism was responsible for the transfer at 1 point of a third to half of the internet's traffic.

What's really interesting about this is that actually it requires 0 global synchronization.

So that is, I just need to remember who I like, and that's it.

And then you can earn reputation with me, and essentially gain a social ranking, just by interacting with me in a nice way.

There's no need for on-chain transactions or anything like this.

But, of course, this leads us to the final example here, which is Bitcoin.

So, Bitcoin is decentralized, ownerless money.

That is, a system of money, or a system of value storage at least, where no 1 party is able to control it, or to change its ownership by force, or print more of the tokens in the network.

We have proof-of-work as a mechanism to reward people for securing the network and also processing transactions.

Again, just another simple mechanism design that was able to do something that was previously impossible.

I mean, from a purely technical point of view, Bitcoin is just evidently a better version of gold than gold is.

There isn't going to be a massive influx of more gold into the system.

But even in that case, it's fairly easy to forge gold, and indeed we don't know how much gold is out there on asteroids and other planets and so on.

Whereas Bitcoin exploits cryptography and mathematics to enforce that scarcity, making it at least at a technical level, simply a better system.

That's the basics of how you program humans.

You find a way of getting between the human and the sub-goal that gets them to where they want to go.

Then you generate a mechanism which rewards them for doing a task that you want.

So powerful that it can overpower the moral frameworks that most people live by in ways that are dangerous and I guess fundamentally disconcerting the first time you would see them, I would say.

Maurice was a governor in the 1920s, 1930s in Canada.

Maurice was responsible for the children's homes at the time as well as the psychiatric wards.

Maurice noticed that the federal government would give him $2 a day for the care of a single child in these homes.

But he also noticed that he got $7 a day for care of mental health patients in their appropriate places.

Then he realized, well, OK, so I could just take these 20, 000 children that are supposed to be in children's homes and I could put them in mental hospitals and I could get 3 and a half times the reward.

It caused 20, 000 children to be misdiagnosed with mental health issues that they didn't have, who were then forced to take medication for those non-existing conditions with all of the effects on their lives that you might imagine.

In fact, this is foundationally terrible.

And this is a result of a mechanism design that was not deliberately and carefully thought through.

But in decentralized protocols it's far, far worse.

So at least the Canadian government could come along and they could say, Okay, this mechanism had a bug, let's change it, and then the world moves on.

The damage is done, but you can at least change what happens in the future.

With decentralized protocols, we enforce games that cannot be terminated or updated.

That is, once the mechanism designer puts it out there into the network, there is no way that they can then go and change it.

They have to get the assent of all of the miners in the network or all the participants.

So, in 2010, 2011, Satoshi Nakamoto commits the 1MB block limit to the Bitcoin Git repo.

Wasn't a big deal back then when there were almost no transactions in the network.

But then fast forward to 2017 when Bitcoin is arguably having its first shot at mass mainstream adoption.

And there are thousands of miners out there, and they are highly economically incentivized by the network.

Like, the mechanism design is really working.

It's enforcing scarcity, and it's rewarding people for accepting transactions.

Well, we start hitting this 1 megabyte limit.

And that means that the users of the network have to start bidding for their places in blocks.

So they're trying to bid higher than each other to make sure that their transactions are included in a block first or sooner or even at all.

And in the peak of the 2017-2018 cycle, the transaction fee inside Bitcoin got to $30 per transaction.

Now, if you're a miner of this network, this is actually great news.

You've got 12 of these happening every second, so in the average block time, you're getting pretty significant amount of money relative to the block reward just from these transactions.

And so you now have no incentive to change the protocol such that the number of transactions that can be included inside a block increases.

Because if people are paying you $30 per transaction now, and they're going to pay you, say, a cent, something like that, when you increase the block size, then you have to have 3, 000 times as many transactions in order to get the same reward as a miner.

So you simply don't have an incentive to do so.

And this 1 simple error in Bitcoin led to the fragmentation of the community and arguably a missed shot at mainstream adoption.

But this is actually the nice side of it.

That is, every time we have a block, we guess a random number.

This random number is hashed with a challenge that's created around the block, and we see whether it is preceded, the output of this is preceded by a certain number of zeros.

And this simple mechanism is very effective at ensuring that an amount of work goes into producing a computational output.

It's extremely simple, it works very well.

But of course it burns enormous amounts of electricity producing very little useful work.

Just that we happen to know if you put this and this together, it produces a hash that has a certain number of zeros.

There are many other systems that could have been used.

We could at least have put that computation to something valuable, like, I don't know, folding proteins.

Protein folding has the same structure of very hard to find the right way to fit the three-dimensional shapes together.

But once you do find a way to do it, then it's very fast to verify.

In the same way that if you take, so it's hard to find these mathematical oddities, but it's very, very easy to verify them.

So you just hash it together, it's a single hash, you compare the value.

Instead, we just burned electricity doing nothing.

Again, massive, massive incentive design flaw.

But I think that this could just be the beginning of how these mistakes go in the future if we're not careful.

And that's why I put so much stress in this lecture.

I think as an industry, it is our responsibility to make sure that we're doing this in an appropriate and a moral way.

In AI safety, 1 of the key things that we look at is how do you make sure that you're going to be able to turn off a generally artificially intelligent machine If you make 1, if it starts to do things that are against human safety and good.

Well, the last thing you do is you put it, quite literally perhaps the last thing you do, is you put it inside a decentralized network where humans are selfishly incentivized to keep it running, to keep it powered on.

If we do this with an artificially intelligent entity, essentially, that we don't know how to control and that isn't working in our interests, the effects of that could be absolutely enormous.

So I would just argue that we should think very carefully before we turn these machines on and we should make sure that our mechanisms are elegant and morally appropriate.

And in fact, there is already a paper out there that studies the Bitcoin blockchain and comes up with some reasonably good arguments that people have already experimented with putting AI inside the Bitcoin blockchain.

But I'll leave a reference to that at the end.

That's why you should be careful why you do it or how you do it.

And this is what it looks like when you actually go through the process.

So when you're designing an economic game, it goes through 3 stages.

You need to choose a reward mechanism, how you're going to make sure that the players in the game have some kind of output that pushes them towards 1 of their sub-goals or goals in their life that they're attempting to achieve.

Then you need to create a function that apportions that reward to people in a way that will keep them satisfied while also achieving the hopefully pro-social goal that you've set for the network.

The goal of Bitcoin's mechanism is to maximize the network security and avoid double spans.

1 interesting thing about cryptocurrency history is that cryptocurrencies have been around since the early 1990s.

They all had this basic structure of some kind of centralized database, and then keys that were held on the person's computer or in whatever device that they were using and You would essentially sign a transaction send it to the database Which would update the ownership records of who had how many tokens of which kind?

And that would be it, but the problem here of course is that this centralized party the database could be messed with and no 1 would really have a record of it.

So some of the early advancements were like, okay, well what if we use hashes to create work chains that showed that actually this couldn't have been altered, or at least couldn't have been altered without the record also being altered.

But of course, it's not a fundamental answer to the question.

And so then along came Satoshi Nakamoto with the idea of a blockchain in 2009 that says, okay, fine, what if we replicate that database across all the machines in the network and then we create a mechanism design, so an incentive, a game really, where 1 of the helpful byproducts is that they're incentivized to accept transactions and to follow a version of truth, follow a version of history in the network, where there aren't 2 transactions that take the same coins and spend them twice.

Tokens are then distributed relatively to the security contribution of each miner inside the network.

This is a pretty simple way of pushing those miners to contribute as much security as they can for the tokens.

If we think back to that example I gave with this generalisable mechanism of mining and how you can use it in centralised start-ups to push down the price of a service essentially towards its base level, this is exactly what's happening here.

So we're saying I'm going to give you 12.5 bitcoins, and you just give me as much work as you can to satisfy that 12.5 bitcoins.

And then the price of the bitcoin determines how much work is going to go into satisfying that.

Proof of work, block acceptance, and fork avoidance and recovery.

So this idea that we're putting extra security, we're producing blocks which have transactions in them, and then we're also critically incentivized to take the most worked on chain.

Otherwise, there's essentially this incentive just to ignore everyone else's blocks and only mine your own.

And then, of course, everyone in the network is running their own blockchain.

And of course, that doesn't hold any value.

So it's very important also that the incentives push people to accept each other's blocks.

Okay, now let's think about Arweave, which is, I would say, sort of later generation in the same kind of field of decentralized networks and mechanism design.

So our goal, the thing that we want to get out of this, a pro-social byproduct, is to secure data forever.

The idea here is to build the new library of Alexandria, where everything from Documents, PDFs, web pages, applications, really anything can be stored and perpetuated for future generations to come so that we don't forget what has happened in the past and we can replicate ideas and essentially memes in the non-cat gif sense, but also those as well, into the future.

So the helpful byproduct we're attempting to achieve is the permanent storage of data.

Users contribute funds to an endowment when they add data, and then miners are rewarded from the endowment over time.

So of course, if you want to store data forever, you have to have some economically sustainable mechanism of making sure that the data is, or rather the reward is available indefinitely.

And so for this, we do something that's actually quite simple, but it sounds a little strange to begin with.

So when you put data into the Arweave network, you pay for 200 years worth of storage of that data up front.

This, you would think, is expensive, but storage is so cheap that actually this is like half a cent per megabyte.

So if you're thinking about web pages, PDFs, applications, this kind of thing, it's extremely affordable.

Then over time, that 200 years worth of storage, as the cost of storage declines, you gain essentially a kind of interest on that principle in the form of storage purchasing power.

So, over the last 50 years, the cost of storage has declined at a rate of around 30.5% every year on average.

The Arweave network assumes that the cost of storage will decline at some rate above 0.5% on average, and if it does, at the end of each given year, you end up with more storage purchasing power at the end than you did at the beginning.

And consequently, you get this sort of compounding effect where the available reward for storing a piece of data actually increases over time relative to its burden to you rather than decreases.

So it's essentially like an endowment for university scholarships, something of that ilk.

And now we've got to work out, how are we going to adjudicate how that reward is distributed?

So this is just another piece of mechanism design.

So, a normal blockchain works by including the hash of the previous block in the production of the next block.

And when there's a fork, you get 2 blocks that have the same hash of the previous block, and they essentially, yeah, they fight on the network for the 1 that can accumulate the most work, and then because of these fork avoidance mechanisms as described earlier, they all sort of, well, they conjoin on the same block and they move forward from there.

So it has this fork avoidance mechanism too.

So with our weave, what we did is we took that same fundamental structure of a blockchain and we modified it.

So instead of just including the hash of the last block in the production of the next block.

You also include a random proof of access to a previous piece of data in the network in the production of that next block.

So you can't take part in the production of new blocks unless you can prove that you have access to old blocks.

And we can stochastically, so statistically check the data in the network at random.

And as long as people can't predict which parts of the network are going to be essentially recalled for any given block, there's no better way to play this game.

The dominant strategy in this game is to replicate as much of that data as you can.

This mechanism lays the groundwork for a network that provides the possibility of business models that you just couldn't pursue previously.

There's simple stuff, like let's make sure if we store, if we make a contract together that we can store an immutable copy of that contract embedded in the blockchain so it has this kind of time stamping that's extremely effective and reliable, as well as the permanent ability to recall that contract.

So we can tell it's essentially the shared record of the truth between us.

But you can also use it for academic publishing, right?

So when you want to contribute to science, what you're trying to do is just push a piece of information to as many people as possible.

Well, just like the internet connects people over large distances, the Arweave connects people over long distances of time.

See all a16z crypto transcripts on Youtube

Sam Williams: Mechanism Design 101