←
Ethan BdM2,936 posts
Ethan BdM avatar

Ethan BdM

@ethanbdm

Ethan Bueno de Mesquita. Dean @HarrisPolicy. Studying tech, conflict, accountability. "A tragic waste of intellectual potential." - R2

◷ Joined June 2009▧ 119 Photos and videos
Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 17

From my work w/ Baliga & Wolitzky: The new possibility that autonomous AI agents are responsible for cyberattacks makes attacking more attractive to human adversaries. More hard-to-attribute attacks from one source encourage attacks by all other sources. bpb-us-w2.wpmucdn.com/voices…

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 17

I would love to see more of these folks use AI tools to help them sketch simple models of their ideas. They are going to find that way fewer of them hold together than they believe. And, perhaps, that would lead them to a bit more rigor and epistemic humility. 8/8

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 17

They are doing so in environments where they receive less critical or skeptical feedback from people who have thought and read deeply on the questions than would be optimal. They often do so with big megaphones and real consequences. 7/

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 17

We are in an age of smart people making sweeping claims about big philosophical, evolutionary, economic, social, and political questions. Things like the social and economic effects of AI, moral patienthood, intentionality and evolution, etc. 6/

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 17

Most of the time, in my experience, you'll find you can't get to a model that tells your story in a way you like. And most of the time that is because your story wasn't coherent. That is a great. The search for truth involves believing more true and fewer false things. 5/

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 17

But now you must read it with a severely critical eye. Does it really capture the story you were trying to tell. Really? No it doesn't. Point out the obviously places and have the AI try again. 4/

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 17

AI tools, which are good at crafting and solving a variety of kinds of models, provide some opportunity. But it requires discipline. Share your idea with the AI and ask for a simple model. It will give you something that looks coherent and it will solve it correctly. 3/

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 17

This has always posed a pedagogical challenge. Because you have to be reasonably good at writing and solving models to see this. Otherwise, it is reasonable to believe your story isn't incoherent, you just aren't good enough at the craft to write a model of it. 2/

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 17

One of the most powerful lessons of writing formal models is that the vast majority of what feel like really good ideas turn out to be incoherent--i.e., you cannot write a satisfactory model in which the story that struck you as convincing holds together. 1/

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 2

These plots are, however, a nice illustration of the classic results on the so-called "instability" of multidimensional spatial voting due to McKelvey and Schofield. 9/9

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 2

But Arrow's Theorem can't be "applied" to any specific case because it is a theorem whose point is that no single non-dictatorial aggregation rule can satisfy all the axioms (normative criteria) across every possible preference profile. 8/

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 2

But it certainly doesn't say there is no aggregation rule that works for that profile. There are plenty. For instance, a scoring method like Borda count yields a perfectly sensible three-way tie in the example above. (Of course, Borda violates IIA on the unrestricted domain.) 7/

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 2

The Condorcet cycle illustrates why we can rule out one aggregation rule (pairwise majority), since the preference profile above is one of the profiles included in the Unrestricted Domain. 6/

Ethan BdM avatar
Ethan BdM@ethanbdm·Nov 2

It says no non-dictatorial rule satisfies Pareto, IIA, and complete and transitive social preferences on every possible profile of individual preferences (Unrestricted Domain). 5/