> And every time some points all of this out, they are "denier", "left behind" and all the other everything-ai fanatics can come up with. As for the imminent crash - I'm all for it. There's a whole generation that is incapable of thinking for themselves without an LLM telling them what to do or worse still - do it for them. And for much of the tech world, that is becoming the single point of failure.

I feel like you can still be an AI maximalist while thinking the numbers simply don't add up.

The problem as I see it is that this is a zero sum game. Nobody has a moat. Even if AI eats all white collar work, where's the profit for these entities? They become commodities and they will compete with each other (and open weight models) only on price, not to mention the workers who lose their jobs to this stuff aren't going to be contributing to the economy at all.

Maybe the "left behind" rhetoric stems from the underlying anxiety of this whole thing imploding and taking the economy with it.

IMHO I think the blast radius is implicitly contained. The major investments are in data centres, and the current investment cases impute near zero residual value after five years. Meaning that current valuations already assume "catastrophic" declines in equipment valuation. This is unusually clear-eyed and sober investment calculus in the tech space.

Further, unlike during the dotcom crisis, most of this spending is not driven by debt. It's mostly funded by the large companies which are producing enormous revenue and profit to pay for this. It's an order of magnitude different.

The major question mark on these valuations is the revenue assumptions, which can be reasonably criticised. A bear case here is that revenue growth is not as aggressive as projected, and valuations steadily decline over time. I don't see a likely scenario where the entire sector collapses. There is no apparent cascade failure mechanism. Of course, these mechanisms aren't always immediately clear prior to crashes.

We must remember that all of the models coming out of China are presumed to be distilled frontier models. Meaning a) they will always be x days/weeks/months behind the frontier models, b) they will never be quite as good, c) inference will generally be constrained by compute capacity (especially as the frontier studios have an incentive to capitalise on their moat), meaning Chinese studios will always be at a disadvantage.

The real wildcard here is self-improvements. It looks like we're already in the singularity, meaning a large proportion of LLM development is already done by LLMs. The development cycle on these might be months now, but it will be weeks soon. Days within a year, then hours, minutes, seconds, and milliseconds. It's impossible to predict what this curve looks like.

You'll be able to buy used H100s. One careful owner. Never overclocked etc.

They'll be knackered though, unlike a good Herman Miller chair. :-)

We might even be able to afford RAM again!