Here are recent rough notes from the site.
I see that Derek Thompson has reversed course from our discussion on the current AI buildout, arguing that I'm wrong: we are underbuilt for ... something. Again,I don't think AI itself is a bubble, just that we will vastly overbuild.
My response to DT:
We always underbuild GPTs (railroads, fiber, etc) until we overbuild. The demand signal he points to is largely endogenous—coding agents—not an external market clearing price.
https://x.com/DKThomp/status/2019484169915572452
Given declining quality data, AI models are becoming increasingly dependent on post-training—reinforcement learning and such. A new Deepmind paper argues, however, that current RL methods are fraught and even counterproductive, leading to behaviors opposite to what is expected.
Hybrid neural–cognitive models reveal how memory shapes human reward learning
https://deepmind.google/research/publications/94006/
This should come as no surprise, but will eventually matter: The top downloaded skill for the overhyped, fragile, and dangerous Clawbot seems to be malware. AI vastly expands attack surfaces, both knowingly and unkowingly.
https://x.com/DanielLockyer/status/2019422410018267328
“The standard way to make an action movie ... was, you usually have three set pieces, One in the first act, one in the 2nd, one in the 3rd. You spend most of your money on that one in the 3rd act. That’s your finale.
And now [Netflix says], ‘Can we get a big one in the first 5 minutes? We want people to stay. And it wouldn’t be terrible if you reiterated the plot 3 or 4 times in the dialogue because people are on their phones while they’re watching.’”
- Matt Damon on tech's consequences