Read on Twitter
Recently I've been thinking about (with @Jabaluck,
@itaisher, and many others) the seeming need in economic theory for a better language for taking about values than the one we get from the "standard toolkit."
A natural economist response is: what would that even be?
1/
@itaisher, and many others) the seeming need in economic theory for a better language for taking about values than the one we get from the "standard toolkit."
A natural economist response is: what would that even be?
1/
Obviously I don't know. But in trying to answer the question I found it helpful to imagine what it would be like to be living just before a recent "scientific revolution": the invention of modern information economics.
2/
2/
Before Akerlof's lemons model etc., it's not as if people didn't recognize that
* information matters in the economy; and
* people sometimes have different beliefs.
But, they might say, no big deal. Just say people have different beliefs over outcomes X: distributions in Δ(X).
* information matters in the economy; and
* people sometimes have different beliefs.
But, they might say, no big deal. Just say people have different beliefs over outcomes X: distributions in Δ(X).
There just wasn't anything very interesting to write a paper about if your whole theory was going to be that two people might have different, arbitrary, beliefs over asset payoffs X.
The revolution was in realizing the "obvious" primitive, distributions over X, was not
4/
The revolution was in realizing the "obvious" primitive, distributions over X, was not
4/
quite the right one.
But a closely related abstraction is more interesting: we have a prior (let's say the same one) over X, but we get different signals about X and so have different beliefs when we play.
The structure afforded by "same prior, different signals" made it
5/
But a closely related abstraction is more interesting: we have a prior (let's say the same one) over X, but we get different signals about X and so have different beliefs when we play.
The structure afforded by "same prior, different signals" made it
5/
possible to ask interesting questions, e.g., "Can we have speculative trade based ONLY on asymmetries in information?" (which led to the "no trade" theorems).
It was a key bit of structure that made a boring observation - we might have different beliefs - much more generative.
It was a key bit of structure that made a boring observation - we might have different beliefs - much more generative.
Later, we saw there are interesting questions even with JUST ONE agent if you hold her prior over X fixed but vary the signal she gets: that's the information design literature.
The theme is that asking questions about a structured subset of beliefs is way more interesting
7/
The theme is that asking questions about a structured subset of beliefs is way more interesting
7/
than just recognizing that there could be different beliefs.
But, crucially, pre-revolution you could easily trick yourself into thinking there's not much to think about here, saying,
"I know there can be different beliefs! We have a primitive for that."
8/
But, crucially, pre-revolution you could easily trick yourself into thinking there's not much to think about here, saying,
"I know there can be different beliefs! We have a primitive for that."
8/
So, what about new, interesting models of values?
Well, we think we know that one good basic primitive is a preference over outcomes for each agent.
We know these can be different, and we even have some standard hopeless things to say about trying to aggregate/reconcile them.
Well, we think we know that one good basic primitive is a preference over outcomes for each agent.
We know these can be different, and we even have some standard hopeless things to say about trying to aggregate/reconcile them.
But at this point we might fall into a trap analogous to pre-revolution thinking about beliefs: yes, they can be different, we're stuck.
And we could spend quite a while there before having an idea analogous to asymmetric information models --
10/
And we could spend quite a while there before having an idea analogous to asymmetric information models --
10/
What would this look like? Well, if it were really analogous it would
(1) build on the existing primitives (in this case, preferences);
(2) even admit "more" of them in some sense -- just as in a signal structure there are "more distributions" than in a bare prior.
11/
(1) build on the existing primitives (in this case, preferences);
(2) even admit "more" of them in some sense -- just as in a signal structure there are "more distributions" than in a bare prior.
11/
But, crucially,
(3) it would introduce new structure that is formally richer/more interesting than just bare heterogeneous preferences.
It seems like there is a huge range of models that could be like this!
Coming back down to earth,
12/
(3) it would introduce new structure that is formally richer/more interesting than just bare heterogeneous preferences.
It seems like there is a huge range of models that could be like this!
Coming back down to earth,
12/
it's clear that "preferences over outcomes" isn't how people really think about ethics.
E.g., a typical person doesn't naturally care about whether most other people even exist, until those people and their interests are made salient to them in some way --
13/
E.g., a typical person doesn't naturally care about whether most other people even exist, until those people and their interests are made salient to them in some way --
13/
often as part of a category or narrative, e.g., in a charity mailer.
Should this just be some awkward failure of our model that is left for cleanup by the psychologists?
14/
Should this just be some awkward failure of our model that is left for cleanup by the psychologists?
14/
Maybe a better fundamental model with the general shape I sketched would be able to discuss such things more organically.
Again, I don't know what it is. There are candidates out there: models of identity, for instance.
But it did help me to specify some ways
15/
Again, I don't know what it is. There are candidates out there: models of identity, for instance.
But it did help me to specify some ways
15/
in which our successors could see us as having been in a "simple" conceptual trap:
having the right primitives to dismiss or trivialize many of the interesting gaps in our toolkit, but not the closely related primitives that would help us fill them.
16/16
having the right primitives to dismiss or trivialize many of the interesting gaps in our toolkit, but not the closely related primitives that would help us fill them.
16/16
