Subject: Superb Interview on AI by Krugman
Paul Krugman posted on his SubStack an interview with Paul Kedrosky, MIT research fellow and tech investor, about all things AI, and I found it most illuminating.

For those of you who lost track of Krugman when he left the NY Times, he has an extremely active SubStack with high quality free content most days.

Here's the interview:

https://paulkrugman.substack.c...

This is free content if you subscribe, so I'm going to share some extended quotes. I do recommend subscribing.

"
[Krugman, introduction]
It’s annoying for economic analysts that two huge things are happening at the same time: a radical change in U.S. trade policy and a giant AI boom. Worse, while I think I know something about tariffs, the more I think about AI the less I believe I understand. So I talked to Paul Kedrosky, investor, tech expert and research fellow at MIT, for some enlightenment. Lots in here that I found startling.
...
Kedrosky: "We...stumbled into a kind of Saudi Arabia of data...the public internet, which suddenly became useful as training data in the context of these massive models that required huge amounts of data and improved on the basis of scaling, meaning that a 10X improvement in the amount of data you trained on led to a predictable increase in the capacity of the model to make what we would call “useful inferences.”...

the nature of large language models, that architecturally—for reasons of data set exhaustion and for reasons of declining returns for increasing investment—we’re kind of at a cul-de-sac already. We’re seeing that happen. So the notion that I can extrapolate from here towards my own private God is belied by the data itself, which shows you that we’re already seeing this sharply asymptotic decline in the rate of improvement of models...

from the standpoint of the data centers themselves, roughly 65-70% of the cost of a data center is specifically the equipment...

I sometimes say “a data center full of GPUs is like a warehouse full of bananas, that’s got a relatively short half life in terms of its usefulness.” That’s important to keep in mind. That’s what makes it different from prior CapEx spending. Moments like railroads, canals, rural electrification, take your pick, because of the nature of the perished ability of the thing that we’re investing in...

Training new models ... [uses] giant amounts [of processing power], at least 10 to 20,000 GPUs inside of one of these data centers ... I’m running the chip flat out 24 hours a day, seven days a week, which requires an immense amount of cooling, incurs a lot of thermal stress (heat-stress)... Some chips last for quite a while, but there’s a high failure rate in the first 2 to 3 years, with a mean time between failure of about two and a half years or so ... This leads to the turnover [of chips] long before generally speaking, you might turn it over just because there’s some hot new chip out..."

Krugman: "What we’re going to end up with is a bunch of burned out chips."

Kedrosky: "The notion that I built a fixed asset that’s perpetually useful is really dangerous."
"

I think there are two valuable insights in there that I didn't know: 1) that generative AI is already hard up against a decreasing rate of return due to the lack of vast new fields of data to train up on, and 2) the investment into data centers will face an immense ongoing cost replacing the chips that are worn out during the training phase.

Krugman and Kedrosky go on and talk extensively about other aspects of the AI boom, including the power consumption, it's effects on retail prices and other problems, the circular and interlocking investment/payment/demand issues fueling the bubble, the use of SPV (special purpose vehicles) to package the debt issued to build the data centers, etc ...

I'll leave off with this quote by Kedrosky:
"
Historically, the U.S. has been very good at speculative bubbles. This is one of our main core competencies here. They tend to be about real estate, or they tend to be about technology, or they tend to be about loose credit, and sometimes they even have a government role with respect to some kind of perverse incentive that was created. This is the first bubble to have all four. We’ve got a real estate component, we have a loose credit component, we have a technology component, we have a huge government component, because we’re told “we’re in an existential crisis with China, and we must win this at any cost.”
"

Lots, lots more at the link.