The shared information environment has collapsed from the rise of the internet and social media.

In its wake is relentless competition for attention. Simple ideas that appeal to instincts and fears are dominant.

Populism is contra-elite, an appeal to the gut. Learning is hard and we don't want to.

This is the worst case for a bubble - the frontier labs have no moat, never realize the profits to repay their loans, and bankrupt.

That would freeze the state of the art, but the inference shops would keep running them as a service. There are still decades of applications to build atop that.

The 'plain language to convince and flatter you' is a big issue. GenAI adopts the meta-syntactic tics associated with persuasive experts, and people trust that much more than the actual content being delivered.

We evaluate the messenger, not the message, and the LLM messenger is suave.

Applaud this so hard. We need papers like this to be the norm.

So much of what we do results in failure, and it is impossible to quantify how many failed studies are quietly replicated over and over because there is no incentive to write about it.

Where we decide to draw the regulatory boundaries affects the size of the black market for jailbroken models - and how accessible they will be.

People generating explicit images of their coworkers is uncomfortable and distasteful, but if the only model that does that also helps you build bombs...

In a multi-dimensional space, distance from the center follows a chi-square distribution; for even a modest number of dimensions, the number of voters close to the center approaches zero.

'Moderate' voters just aren't polarized along the hyper-partisan dimension. They still have preferences.

A fun corollary: with enough data to avoid overfitting, adding successively higher polynomial terms is effectively constructing a Taylor series - and thus fitting an arbitrary function over the range of the data.

I'd push back on this slightly. 'Deliverism' matters for highly engaged, 'base' voters, but probably matters very little for disengaged 'swing' voters.

This held for Trump's coalition as well - the campaign message was about grocery prices but the base knew project 2025 was coming.

This main effect aligns with the across-firm regularity of larger firms producing lower quality patents on average. Any quality increase in 'the good stuff' is overshadowed by increases in strategic junk patenting.

Seems to be some sort of variant on patent thickets, ala Shapiro 2000?

We highlight a fact that at least a lot of economists I talked to didn't really understand well enough, which was that in 1965, 2% of the entire US economy was public science. One in every $50 was public science. That's now about half a percent of GDP.

3/n