That's absurd. It's a physical impossibility to bring that much power online that quickly. And the cost to get even close would make AI more expensive than just hiring knowledge workers to do the same tasks.
And it's all predicated on a tower of wobbly or broken assumptions -- chief among them that increasing the size of these models yields better performance.
We're going to look back on this era and wonder why anybody took any of the outrageous claims of tech CEOs seriously.
I'm assuming you disagree that larger models are better? Can you expand on what indicates that AI will hit a wall in scaling given the evidence of the last 9 years of scaling transformers (or other models)? Where on the plot does the line go from exponential to flat?
(Incidentally, the line of skill improvement isn't "exponential". It's been incremental in improvements per generation, but generations have been coming thick and fast of late, and have grown in parameter count exponentially since 2017.)
Speaking more broadly, LLMs don't have to "hit a wall" in scaling to become uneconomical. If incremental improvement continues to come at exponential cost, however, then the fundamental value argument falls apart.
Setting all that aside, even presuming that model performance scales linearly with dimensionality, there are just fundamental limits to the size of the training corpuses. Quality training data is not unbounded and infinite. Given the same size corpus of training data, there's a hard theoretical limit to how much meaning and inference a model can wring out of it.
And then there are other issues with the whole business model. For one thing, it's predicated on continual full scale retraining to achieve even modest gains in skill and relevancy. Topical and targeted learning requires a full retraining. Etc cetera.
I think that the next generation of AI will lean more heavily on RL to be useful beyond a few months. I also think that the energy requirements of a particular technology are a good proxy to whether it's got a realistic future.
The curve flattened out years ago. The exponential was going from GPT-2 to GPT-4 (or thereabouts). After that, it was painfully obvious to anyone observing without a vested interest in believing otherwise that the progress had slowed.
Now, it's not just that progress has slowed: it's that the exponential has reversed. In order to get marginal gains, they have to throw exponentially more hardware at the training.
China begs to differ.
Whether or not you believe we will reach it in a fee years, we are certainly wayy closer today than we were even two years ago.
The possibility of genuine AGI obliterates all the financial or energy related worries, they pale in comparison to the ultimate impact of such a technology.
However, yes, if you believe AGI is not possible or won’t arrive in the coming decade then all the data center buildup seems foolish.
Next, you have to have a clear path to reaching it.
Then, you have to have the resources to actually walk that path.
Only with all three of those can you make any credible claim that AGI is near.
As it stands, we have none of them—and the lack of the second is the most damning. It's very, very clear at this point that just scaling up the existing LLMs is not going to reach some critical mass and result in AGI, like the serendipitous sapience of Mycroft in The Moon Is A Harsh Mistress.
Given that, any path to AGI necessarily includes some new breakthrough on it (or more than one). And by their essential nature, breakthroughs are not something you can predict or schedule. Indeed, you cannot even be guaranteed that they will ever happen. (It is likely, assuming that it is physically possible to build AGI, that we will figure out how at some point...but not guaranteed.)
Gotta love this argument. Top it off by saying anyone skeptical is a fool, because of course.
If we had access to AGI today, we'd just find novel and interesting ways to ignore it, enslave it, gimp it, and/or bias it.
The last major advancement was probably GPT3, af least if we are talking about the LLM companies, the ones involved in the current data center boom.
After that was we experienced were marginal improvements of the same technology. Yes, the current models are better than what OpenAI put out at the time of ChatGPT 3, but none of it was revolutionary (and the gains have been less and less perceptible in newer versions).
We might be as far from AGI as we were in 2022. I think we are multiple revolutions in technology away from it.
https://catenaa.com/markets/cryptocurrencies/hut-8-builds-fl...
And they'll also do "high-performance computing."
Yet, I think Sun's early 2000's vision "the network is the computer" is finally coming and these data centers will all end up becoming multi-use. Want access to apps running with 128GB of memory? Fine.. it'll just be on a thin-client with a data-center powering it (and everything else it does.)
It's not a bad model. As I've mentioned previously, on the client-side I think will be a new era of all-in-one modular SBCs (medium clients.) These can become thin-clients for really beefy applications too that don't have to be "local-first" and can thus be "cloud enabled."
It'd also be interesting to see crypto become more dynamic. Like making it super easy to issue a token for say an upcoming event, or better yet, a new invention looking for early adopters and supporters like Rodin Coils. The big data centers on the backend can make it secure. Just speculating. So the "big iron" compute won't ever be wasted, just repurposed dynamically.
All these mad-scientist inventions will come from unemployed geniuses and tin-foil hatters, some of whom may actually be right. Let's see if they can find a way to vastly speed up radioactive decay with lasers, but, letting the bankers be fine with it all.
The rest of the economy is dead. Oracle is dead without OpenAI. Remember that unlike the dotcom, none of these companies are public. So when it pops, you’ll see private credit and PE funds implode, which could bring down banks with unhedged exposure. The headlines talk about JP Morgan (which likely has the risk managed), but regional banks got into that nature in the last couple of years in a big way.
It’s also why SpaceX wants to be included into index funds as soon as possible after they go public. I recall the rules may be revised to support this, meaning everyone who has money saved in those funds will automatically be tied to the fate of SpaceX.
And no bubble can last forever.
And the longer they go, and the larger they get, the worse the fallout is when they finally pop.
Let me put it this way, IF AI is a bubble then I'd like it to go bust asap instead of dragging along and going public and then us discovering BS/creative accounting revenue in S1 filings. by then it would be much worse. Right now VCs and PE firms will absorb it all.
The thing with dot-com was that there was actual public market corruption & euphoria. that caused the bust painful for everybody. RN its bigtech & PE how has heavy cash reserver and margins to bun through. I'd much rather have them take it then average 401k.
Just like now, the financial cup game is insane. Commiting money to a company that plans on doing a thing if they can get another company to do a thing and then that company is leveraging those cumulative possibilities in its own wager. The speculation is out of control.
Who’s going to pay me otherwise, becoming the chief security officer aboard the Altman-Musktani vessel USCSS Shiba Inu?
You’re all going to eat rats on a stick while I pull the charred meat off some hostile tech CEOs’ neurodivergent talent for lunch I roasted to perfect crispiness with my boring company flamethrower arm attachment.
So please all keep paying the magic word generator companies, so we can replace most of you and your miserably inefficient human production cycles, including eight hours a day not working because you lie around unconsciously, to become better human livestock, eh, I mean valued "Human Resources".
/s
So yes, AI is a bubble, but this bubble has generated value, it’s not at all like 2008.
You're also ignoring the cost of purchasing and amortizing dedicated hardware in your local model example.
It's not an apples-to-apples comparison.