[paulbjensen]

I think the scepticism around whether the AI firms can provide the return on investment for their capital expenditure plan is completely understandable.

It reminds me of a mini-story within Michael Lewis' Flash Boys - there was a massive project to drill through the solid granite of Pennsylvania mountains in order to lay a new fiber-optic cable so that High-Frequency-trading firms could use it to get their data quicker between Chicago and New York. It was done at huge cost, I think around $300m. The Fiber-optic cable could transmit data between the locations in around 13.1ms - 13.5ms.

However, an alternative option of Microwave towers was setup and installed to transmit data between the same locations. The Microwave Tower could do the same, but around 8.5ms-9ms, ~30% faster. And it didn't cost anywhere near as much.

I worry that not only are those current capital expenditure plans wildly unaffordable, but also that the risk of an innovation rendering all that infrastructure obsolete can't be ruled out. It's a massive gamble.

[adverbly]

> risk of an innovation rendering all that infrastructure obsolete can't be ruled out. It's a massive gamble.

It's not just a gamble on technology - it's a gamble on the company too!

You own a company's stock - not a technology's!

In the dot com era, it was the realization that the wrong companies were picked that popped the bubble. Turns out pets.com was never gonna be the billion dollar business. It was that book store Amazon instead. And it was the Google search engine not the other one. And browser A not browser B... And so on... When the expected winner shifts away from one company towards a position where people want to sit on cash and wait and see, you suddenly get a massive rush to the exits and poor pets.com needs to throw in the towel.

We're gonna find out in the next couple of years who the pets.coms are.

[iso1631]

The thing which really amuses me about that time is the Futurama Episode "300 big boys" from April 2003

Hermes: You invest this penny like you wanted.

Dwight: Thanks, Dad. I'm gonna take this and buy five shares of Amazon.com.

Hermes: A risk-taker? That's my boy!

In today's shares Amazon peaked at $5/share in December 1999 (adjusting for splits etc) but by early 2003 had collapsed to about $1.

Certainly enough to joke about it as a cartoon writer.

Today of course it's $250 a share.

[jvanderbot]

What would the equivalent be in AI? A dirt cheap inference chip for consumer hardware? 2T sized networks on a cell phone so we don't need data centers?

Your analogy required no technological innovation, just a good look at solution space. Where's the microwave tower for AI that everyone is missing?

[lumost]

A breakthrough in grokking/regularization, quantized or sparse training methods, quantum optimization, or more efficient architecture could reduce infrastructure needs. Although any of these could simply advance the parrot frontier without solving the prisoner’s dillema driving infrastructure spend at ai firms.

[zipy124]

More efficient architecture/training regimes would be a big one.

[1vuio0pswjnm7]

One difference is the HFT firms were certain that increased speed would translate to revenues that exceeded the project's expenditures

The "AI" firms have no such certainty that "AI" translates to revenues that will exceed their expenditures

All they have is endless speculation