Andrew Little

Huh weird. Humans: adding value from 2 million BC until at least March 2026.

From some experiments playing around with using various models to check proofs of formal theory papers, pdf parsing issues are a huge problem there too. By contrast, generally works super well when inputting raw .tex files.

Final point: not that most voters are paying attention, but I don't think "we want AI systems that allow us to surveil you and build autonomous killer robots" is a winning pitch

“We will nuke your company if you don’t allow us to do X, because we would never do X” is also a very weird argument

Sort of an obvious point but one drawback to firing all your attorneys and bombing random boats in a way all experts say is illegal is that partners will no longer trust you to follow the law.

“I transformed my life by writing a CC app to automatically post complaints about CC posts”

(Of course said inertia may sometimes be a good thing!)

Never underestimate inertia generated by senior academics

next level: Kavanaugh policy (something blatantly illegal but stays in place because it is costly to reverse)

I think that at least for the medium term there will be some difference between the quality of checks authors run themselves (if at all!) and the quality of checks done by "pros" but yeah that will likely shrink over time.

Iran news pretty good for the Zeitzoff hypothesis.

Order No Option But Sabotage here:

-OUP (30% discount code: ASFLYQ6):
global.oup.com/academic/pro...

-Bookshop:
bookshop.org/p/books/no-o...

-Amazon:
www.amazon.com/dp/019779684...

Tired: replicate a paper

Wired: Get claude code to replicate 100 related papers

Inspired: Just wait for @yiqingxu.bsky.social to publish his paper replicating all empirical political science

If you aren't checking the reproducibility/robustness of your papers with AI, someone else is going to do it for you!

Definitely think everyone should be trying these tools to get better at producing/checking code but if one can do that well with their old workflows I don't think this instantly renders them a fossil.

Why do you mean by left behind in this context? I see the takeaways here are that a few people will write cool large scale replication/robustness papers, and everyone else should expect their papers will be ingested into such an exercise and be more careful with code moving forward?

Would defer to Macartan here but I think fair to say there is no core in 2d under the “standard” assumptions? Though ofc reasonable people can differ on whether the standard assumptions are the best starting place.

yeah definitely possible that both parties can end up on one side of the median on some dimensions, but the directional claim that either one moving to the median would help win is still hard to break!

Yeah to be clear I’m making a statement about properties of electoral competition models, including out of equilibrium choices. Much easier to break what most people mean by MVT (ie prediction that convergence is exact). But true that “people like candidates close to them” is typically just assumed.

Great post! technical nit to pick: chaos theorems typically just require 2 dimensions not 3+. (At least the examples I teach are 2d…)

Eg uncertainty about where the median is, bimodal ideal points, endogenous turnout (depending on how it’s modeled), non-policy considerations, more dimensions and parties generally don’t break this.

Agreed but also “moving closer to one’s best guess of what the median voter wants (all else equal) increases the probability of winning an election” survives most of the natural assumptions one might want to loosen.

From where I sit “use the tech to help write the best JMP possible” is probably a good rule of thumb.

Thanks, need to actually read the paper before saying more :)

The places I’ve worked put much more weight on “how good is the best work” (on paper and in the job talk) rather than counting pubs already, and I’ll guess we go further in that direction. Might be different elsewhere.

In any case we are definitely going to have to change what we weigh and pay attention to in hiring and promotion

I guess? Of all the potentially scammy uses of the tech this seems pretty low return.

Oh yeah I meant the value of sending slop papers with a pseudonym to journals seems…minimal?

The cost may be low but the benefit is zero?