The recent surge of general interest in blockchain has resulted in my having many conversations with people about building decentralized institutions.

Here, I’ll:

  • Define institutions, and provide some examples to fix intuition.
  • Make two straightforward-but-underappreciated points:
    • Decentralized institutions won’t be that intelligent.
    • Institutions that do desirable things will probably be bad at surviving.
  • Vaguely discuss some implications for actually building decentralized institutions.

0. Background

0.a. Institutions

We’re often interested in doing some thing. So, it would be nice if we could build an institution that:

  1. does the thing (Implementation)
  2. for a long time (Survival)

These two conditions provide a trivially broad definition of institutions that we’ll use here, in order to draw attention to commonalities among all types of institutions:

  • Institution := a wrapper that implements some behaviors, minimally survival.

Examples:

  • Some ‘designed’ institutions that explicitly implement some intended behavior:
    • A guard dog is an institution that implements property protection, survives by compelling owner to give it food / shelter.
    • Google is an institution that implements online search, survives by selling ads associated with its search content.
    • USA is an institution that implements some vaguely defined set of ‘American interests’, survives by appropriating a portion of the economic activity occurring under its purview to fund defense.
  • Some ‘natural’ institutions where it’s unclear if they implement anything (beyond survival):
    • Capitalism is an institution that survives by incentivizing people to engage in market transaction / wage war on other people who don’t engage in market transactions.
    • DNA is an institution that survives by instantiating organisms and then incentivizing them to create more organisms.
    • Laws of physics is an institution that survives by… I don’t know, it just kind of does?

0.b. Decentralization

  • Our primary ways of building institutions have been pretty centralized, e.g.:
    • Dogs (and other animals) appear to have an ‘ego’ that functions as a centralized control center for the organism.
    • Google (and other companies) concentrates a lot of power in its executives, though many functions are decentralized throughout the organization.
    • USA (and other democracies) concentrates a lot of power in some elected officials, though typically for short periods of time and subject to popular approval.
  • This centralization leads to some weaknesses when it comes to Implementation and Survival:
    • Centralization often allows some parties to flexibly deviate from implementing ‘desired’ outcomes.
      • A guard dog is free to stop guarding the property if it wants.
      • Google can show more ads to increase revenue at the cost of worse search quality.
      • USA’s elected officials can try and enrich themselves with taxpayer money.
    • Centralization often also makes these institutions quite brittle:
      • A guard dog typically ceases to exist within ~15 years.
      • Google can be shut down by e.g. a government by e.g. confiscating its servers.
      • USA can be debilitated by e.g. compromising the integrity of powerful elected officials.
  • It feels like decentralization could help address these weaknesses.
    • Certainly, many of the longest-lasting and most powerful institutions in our world are decentralized:
      • Capitalism leverages decentralized exchange of goods to converge on market-clearing prices.
        • Compare with central planning, which requires e.g. a literal central Walrasian auctioneer running tatonnement.
      • DNA uses decentralized sexual reproduction to continually optimize itself for survival.
        • Compare with intelligent design, which requires a literal god to design organisms for particular purposes.
      • Laws of physics define local rules of interation, which then aggregate up to global phenomenon.
        • Compare with a video game world, where global phenomena are centrally defined in code, with local behavior following as a consequence.
  • But these decentralized institutions are all major forces of nature, rather than human implementations.
  • We don’t really know how to build decentralized institutions.
    • Blockchain is hype because it offers a glimmer of hope on this front.

1. Decentralized institutions won’t be that intelligent

Let’s define ‘intelligence’ as the ability to quickly optimize against unforeseen circumstances.

  • Suggestive empirical evidence that higher centralization seems to correlate with higher intelligence:
    • Single engineer overseeing a project can just push product changes immediately, whereas if there are many stakeholders, then lots of communication necessary, need to worry about internal politics, etc.
    • Startups are often able to out-maneuver larger corporations.
      • Less communication overhead and better aligned incentives => startup acts more like a single centralized entity.
    • Corporations in general are still fairly centralized, as corporate survival requires optimizing against changing market environments.
    • Democratic forms of government are by design quite slow to adapt to novel circumstances, whereas autocracies have the potential to move quickly (e.g. US vs China on energy policy).
    • Military organizations are all highly centralized, as winning wars requires being very adaptive.
  • Fundamental problem: it seems like our only model of intelligence is centralized, and it’s something like this:
    1. Centralize control with a single entity.
    2. Constrain this entity’s preferences to be somewhat correlated with desired ones.
    3. Give the entity some resources to do various kinds of optimization.
  • Two examples of this model:
    • Corporations:
      • Investors centralize control with CEO.
      • Investors align CEO incentives with investors’ via conditional stock grants.
      • CEO goes ahead and leverages company resources to do optimization.
      • (This is a very stylized description, but the intuition is there)
    • The ‘ego’ (:= subjective feeling of being a single unified entity) could just be DNA’s way of implementing a smart institution that helps DNA survive:
      • DNA creats this ‘ego’ to be the centralized controlling party.
      • DNA then aligns the ego’s incentives with DNA reproduction through pleasure / pain.
      • DNA provides a brain, gives ego access to use it for optimization purposes.
      • (This is a somewhat fanciful description, but it doesn’t sound implausible)
  • As a result, any intelligent decentralized institutions we want to build probably need to be built out of aggregates of people.
    • This problem is theoretically hard, owing to some negative results on preference aggregation, (e.g. arrow, gibbard-sattherwaite).
      • Impossible to construct decentralized preference aggregation rules that satisfy some basic requirements.
      • As a result, decentralized human institutions are almost necessarily fraught with incentive conflicts.
      • Incentive conflicts seem to make institutions less intelligent.
  • So we’ll probably have to come up with a new model of intelligence…
    • Serious artificial intelligence could be such a model
      • This would allow us to have intelligence detached from incentives.
      • This would be useful, as incentive conflicts constitute a major cause for the seeming incompatibility between decentralization and intelligence.
    • But serious AI is still quite far away (I’m not optimistic about neural networks being that universal of a function approximator).
    • And all our current machine learning / statistics / whatever methods are too rigid / brittle to stand up to concerted human intelligence.

2. Institutions that do desirable things will probably be bad at surviving

  • Trivially, if you want to maximize some combination of Implementation and Survival, this will result in a (weakly) lower performance on Survival than optimizing purely for that.
    • Absent some theory telling us why this constraint should be generically non-binding.
      • Which would be quite surprising.
  • Some examples of tension between Implementation and Survival:
    • (Note: all of these examples are somewhere between highly speculative and irresponsibly naive, but should serve to fix intuition. Please don’t take them too seriously!)
    • Universities and preference for children of alumni:
      • Desired goal of promoting intellectual meritocracy.
      • Need money to survive.
      • Can incentivize alumni to provide funding by cultivating loyalty via looser standards for children.
    • Political lobbying:
      • Desired goal to limit special interest in politics.
      • But, not really feasible to create policies without domain experts’ assistance.
      • Have to provide some incentive for domain experts to exist.
      • Thus, allow special interests to extract some rents in exchange for providing domain expertise.
    • Buddhism and reincarnation:
      • Desired goal to reducing suffering.
      • Natural way to reduce suffering is to have fewer people.
      • Incentivizing followers to not have children is a great way for institution to die out, so can’t do that.
      • Thus, this whole weird thing about reincarnation.
    • Capitalism and poverty:
      • Desired goal to provide overall economic welfare.
      • Must incentivize people to actually work so that goods get produced.
      • Thus, safety net must not be too generous.
      • Results is that some number of people are quite poor, despite plentiful resources.
    • Biology and pain:
      • Desired goal of providing a pleasant life to all organism.
      • But, in order for the DNA to survive, it must be able to motivate organisms to survive and reproduce.
      • Use pain to motivate organism to avoid stuff counter to these goals.
      • Result is that some number of organisms live quite painful lives.
  • Heart of the issue:
    • Survival is (tautologically) a minimal condition for any institution to exist.
    • Tendency for things that are not purely optimized for survival to not survive.
    • Implementation of some desired behavior will probably handicap the institution on the Survival front.
  • So, must be careful that:
    • Institution doesn’t die.
    • Institution doesn’t wiggle itself away from Implementing the desired outcome.

3. Some implications for building decentralized institutions

  • The above two points points have some implications for what institutions we might hope to build and how we might go about building them.
  • Relative to the last two sections, my thinking is on this is relatively more imprecise.
    • As in, I just don’t concretely understand the mechanics of building decentralized institutions.
      • Though I think this a general problem for our species as a whole, rather than just me.
  • Hopefully, some of these points are still interesting.

3.a. In what situations could decentralized institutions be useful?

  • Incumbent parties are extracting a lot of rent / imposing a lot of inefficiency, to the detriment of the rest of the market.
    • A decentralized institution could presumably more easily commit to not extracting rent.
    • Many blockchain projects seem to by a desire to resolve this kind of problem.
    • Overall, this point seems to be understood fairly well.
  • The problem setting has some natural barriers to entry, e.g. network / reputation effects.
    • Decentralized institutions that implement some desirable property will typically be not so great at survival, so that it will need some sort of extra edge in order to not be out-competed by more intelligent centralized institutions.
    • This setting is pretty correlated with settings where incumbents are able to extract significant rent, so probably fine here as well.
  • The problem feels like it has a pretty ‘natural’ resolution via decentralized institution.
    • It feels like the more hacky the decentralized institution is, the more Survival / Implementation will suffer.
      • Our current decentralized institutions resemble forces of nature more than human inventions.
      • At least in the near term, the set of things that can actually be solved via decentralized institutions may be somewhat limited.
    • Many blockchain projects seem to ignore this, instead opting for “$ARBITRARY_THING, but on the blockchain”.
      • This strategy of picking a random problem and hoping that it has some decentralized solution may have made sense in the late 2017 / early 2018 funding environment, but it’s unclear how successful this approach will be.
      • Maybe at some point in the future, we’ll have a CRISPR-level advancement that allows us to generically implement anything decentralizedly.
        • But that seems pretty hard (and probably false).
  • The problem setting is fairly static and well understood.
    • Because building intelligent decentralized institutions seems pretty fundamentally hard, it’s probably better to focus on building dumb ones. Such institutions obviously not work well in fast evolving environments.
    • Many blockchain projects seem to ignore this constraint. E.g. projects of the form “$TRENDY_CENTRALIZED_STARTUP, but on the blockchain” (e.g. AirBnB, but on the blockchain).
      • These trendy centralized startups tend to operate in somewhat newer, less well understood environments.
      • So there’s probably significant work that needs to go into discovering the right approach for the environment.
      • Even after converging to a correct solution, the process of extracting rents might require a long period of innovation before it becomes sufficiently mechanical to easily decentralize.

3.b. Some challenges

  • We don’t know much about how to build decentralized institutions.
    • Making progress here will likely require long-term commitments.
    • It’s not entirely clear if there’s enough appetite for that, given that much of the recent crypto hype has been driven by a get-rich-quick kind of mentality.
    • Hopefully, the current correction stabilizes interest in the space at a lower-but-still-significant-enough level.
  • The work of actually building and rolling out a decentralized institution can currently only be done by centralized institutions.
    • We certainly don’t know how to build decentralized institutions that can reproduce themselves.
    • Thus, the success of these decentralized institutions in the near term critically depends on the centralized development organizations being honestly interested in seeing the decentralization through to the end.
    • Again, some questions here, as many projects in the space seem to be too focused on short-term gain rather than serious long-term work.
  • There are many centralized institutions that may see decentralized institutions as a threat.
    • It’s not entirely clear how to circument this problem.
    • Maybe rely on the same sort of mechanism as regular startups: hope that the incumbent is too slow to respond.
      • This is more difficult for decentralized institutions, as building a decentralized institution is hard given our current lack of understanding in this field.
        • Will almost certainly take longer than building a comparable centralized institution.
    • Or, maybe there are certain strategic situations in which decentralization could be an equilbrium outcome as a form of e.g. ‘making peace’, if multiple centralized instittuions are engaged in a costly conflict.