The good news is that the AI bubble/boom looks like it could be similar to the C19th railway boom.... there was a significant bust, but the boom left a widespread & transformational infrastructure in place.

The bad news is that the AI bubble/boom looks like it could be similar to the C19th railway boom.... when sentiment & economic stagnation hit, shares halved in value very quickly causing massive economic disruption.

History doesn't repeat, but it *does* rhyme!

#AI #FinancialCrisis
h/t FT

@ChrisMayLA6 the AI bubble leaves us little of direct value. AI GPUs with a short working life and no other use, some big buildings and a few unsustainable powerplant hacks.
The stuff from before it has benefitted a bit but blowing astronomical amounts of admittedly mostly imaginary capital to make screen readers and the like work a bit better isn't exactly big returns

@ChrisMayLA6 how so? Railway infrastructure depreciates over multiple decades (UK transport being still heavily reliant on Victorian decisions). GPUs obsolete in a year or two at best (though I suppose applying pressure to get local power reliability improved for data centers might stick a bit longer in places like the US).

@BenHourahine

Fair comment; the Q. is, as data centres are the key infrastructure being invested in, is whether their utility might last longer?

@ChrisMayLA6 @BenHourahine I doubt that data centres built for AI and full of GPU with very limited lifetimes and mostly powered by polluting gas turbines will have any consequential use.
They may well become the modern equivalent of coal mining's slag heaps.
Current attempts to repurpose cryptocurrency mining data centres to AI illustrate the problem of finding new uses for data centres.

https://digitalmarketreports.com/news/25353/transition-from-crypto-mining-to-ai-data-centers-faces-significant-challenges/

@marjolica @BenHourahine

fair comment - but I suppose if one was looking for partial parallels with rail one might not the difficulty rail had with the arrival of the car?

@ChrisMayLA6 @marjolica @BenHourahine more I fear canals after railways. Lots of totally useless assets with no significant new use cases. LLMs right now have two probably viable use cases which are porn and fraud, and half of the porn uses will probably turn out to be really bad ones

@etchedpixels @ChrisMayLA6 @BenHourahine canals and rail for commerical use coexisted for a long time. And some canals were actually converted to use for railway lines. Commerical use in the UK almost stopped after the WW2, but some of that was because we were the first to build canals and initially only built narrow canals, on the continent they started later and build wide canals that are still in use .
These days canals in the UK are used for leisure and they are still an important part of the UK's water distribution network - I was involved in modelling that at one time.

@marjolica @etchedpixels @BenHourahine

Yes, the canal network is a classic example of first mover disadvantage - latter versions as you say were better future-proofed

@ChrisMayLA6 @marjolica @BenHourahine and amusingly our railway tracks are too narrow and we spaced them too closely and made the tunnels and bridges too low. Hence no double deck trains, no trucks on trains, and awkward limits on vehicle lengths and widths we are still fighting

@etchedpixels @ChrisMayLA6 @marjolica @BenHourahine LLMs are ... well... pretty good LMs :-). So you can use them for speech-to-text, for example. Translation is another "easy" use case. Whether it is worth the money invested, I'm not sure, but we should be able to get practical, voice-controlled computers, for example.