Historically, AI is an umbrella term for anything and everything around using machines to make decisions. Spam detection, video game adversaries, explosive detection, and the old Xbox Kinect are all different things that have used artificial intelligence for 10+ years before GenAI became practical.

If you’re interested in learning more, all this stuff falls under the umbrella of “machine learning”.

How do you differentiate AI and Generative AI?

So, back in 2005 or so when I was in college, I took a class called “Introduction to Artificial Intelligence”. In that class, among other things, I learned about the A* algorithm. If you’ve ever played a game where an NPC approaches your character while having to take into account obstacles or differences in terrain or some such, that used the A* algorithm. (For instance, escort missions in Skyrim.) It’s quite a versatile algorithm, in fact, and can be used to optimize a lot of different sorts of things.

And if I were asked to come up with two algorithms more dissimilar from each other than any other pair of algorithms, the algorithm behind LLMs and the A* algorithm would probably be a pretty good guess. The A* algorithm requires no training. LLMs do. It can be easily understood by a “person having ordinary skill in the art” (PHOSITA) of coding how the A* algorithm came up with any particular answer it gave just by tracing through the code. With LLMs, there’s no real way to figure out exactly why it produced the particular output it did. (Asking “why did it do this particular thing?” is like asking “which specific rep at the gym made you able to perform that 500lb dead lift today?”.) The A* algorithm is for a pretty specific use case: finding the “least expensive path” from A to B (potentially in the presence of obstacles or differences in traversing cost on different path segments). LLMs are aimed at anything you might want to throw at them. And yet, both are “AI” (somehow).

Now, I don’t know that I have the most precise definition of “Generative AI”, and I haven’t researched the following bullets enough to be sure about all of them, but:

• I don’t know that the term “Generative AI” is even very precisely defined. It’s probably valid to consider it more of a “marketing term” than something that actually has a direct connection to the algorithms behind it. (Like, for instance, the term “Enterprise Software”.)
• It seems like the term “Generative AI” is “new”. So at very least any technology called “AI” before the term “Generative AI” came up is not “Generative AI”.
• “Generative AI” seems to mostly be used for things that try to be as “big” and “general purpose” as possible. Stuff that tries to solve every problem, not just one niche of problems. (On that basis, I would probably generally exclude narrowly-focused SLMs from the definition of “Generative AI”.)
• In Korea, the “AI bubble” seems more focused on bipedal robots. I don’t know that I’ve heard that called “Generative AI”, though, so I’d generally tend to exclude it.
• I think it could be reasonable It strives for being as general-purpose as

But perhaps more precisely, “Generative AI” pretty much seems to refer to LLMs, Stable Diffusion, and maybe the DALL-E family of image generation algorithms. (If there’s anything else that refers to, I’m not familiar with it.) That’s the definition I’ll use for this conversation (at least until further notice). I don’t have an opinion on (or care) whether some particular upcoming technology (like LeCun’s “Joint Embedding Predictive Architecture” or whatever) will or won’t be generally considered “Generative AI” or just plain-old “AI” or “not-AI” or whatever.

Sidenote: All the articles that have gone super viral about new drug research breakthroughs from “AI”? They’re expecting most readers to think biochemists are asking ChatGPT to design new drugs. (Or at least to not think about it deeply enough to realize that nothing learned from building that kind of AI is going to benefit LLMs or Stable Diffusion algorithms.) It’s just a misleading way to drive up more AI hype.

The investment bubble may collapse on them

There are definitely voices out there saying that’s inevitable. (Ed Zitron one of the most vocal.) And I seriously hope Ed’s right. But I have to admit that the blockchain bubble hasn’t gone the way I would have expected. The bubble has popped in some sense. Almost nobody’s talking about blockchain any more. The big companies that used to support Bitcoin don’t any more. There isn’t a ridiculous proliferation of projects with shoehorned-in “blockchain” because they can’t get investment without including that term any more. It’s… popped.

But… Bitcoin’s still worth $84,000-ish? Jesus Christ. If I were making predictions, I would have thought the death knell for blockchain would have come in the form of people finally figuring out that Tether was inventing fictitious trillions of USD that didn’t exist and the Tether “stablecoin” going to zero, bringing the whole blockchain ecosystem, resting like a house of cards atop Tether, down with it. Maybe that’ll still happen? Who knows. Whatever the case, I’m glad I don’t have to think about blockchain these days or get buttonholed by my deranged coworkers about it these days.

I don’t think the investment bubble in “AI” is sustainable in the least. But I fear no one can really predict how (nor, unfortunately when) exactly the unrealistic mania will resolve.

but the usage isn’t going anywhere but up.

I mean, the same was true of Beanie Babies, blockchain, subprime mortgages, fiber optics, and tulip bulbs until it wasn’t. That’s how bubbles work. The promises have to get more unrealistically, unreasonably bold all the time to make sure “line go up”. And dumbass CEOs keep falling for it because all the other CEOs are falling for it and due to FOMO.

I don’t know why you think it’s some sort of scam.

Jeez. Where to start.

I guess if I had to pick one thing, probably the biggest issue with LLMs is that getting your employees en masse to outsource their thinking and forget how to think themselves is a losing strategy. Especially when subtle hallucinations are such a problem with LLMs (which I don’t think can be overcome). The evidence that LLMs are causing more problems than they’re fixing are in, if nothing else, the massive Amazon outages that have resulted because “whoopsie doopsie the AI deleted the whole IT department and rebuilt it from scratch and the engineers all had their brains switched off from too much LLM use and didn’t prevent it ahead of time”.

But also, the finances just don’t seem like they can possibly work. The LLMs are all being sold at a ridiculous loss and burning investment capital to fund this ill-conceived experiment. To be actually profitable, OpenAI/Anthropic/etc will have to charge many times what they’re charging now. And “local” LLMs aren’t that much cheaper.

If it was such a scam, companies that have been using it for two years already would be throwing it in the trash

It’s not like we don’t have examples of that. Klarna is one that has scaled back on AI adoption. As is SalesForce with regard to its whole “AgentForce” thing.

and instead they’re buying up more capacity.

Again. That’s how bubbles work, until of course they don’t.

I have legitimate business use cases for it today

…or maybe you have a “mild” case of AI psychosis and are ignoring all the drawbacks. Plus, again, if you’re using LLMs today, remember that the pricing scheme you’re observing now is still in the “sell at a sizeable loss to get people hooked” phase and there’s no guarantee people will ever get “hooked”.

and my company can’t get hands on hardware or even available compute from the big players because it’s in such high demand.

It’s not like that problem started with AI. What business case do you have for blockchain now-a-days? And when was the last time you bought a graphics card for… you know… graphics?

But also, companies have bought ridiculous amounts of hardware that they can’t even use because the data centers don’t exist. And the “plans”/promises they make about the rate at which they’re goig to build data centers are beyond outlandish.

All that to say that the reason you can’t get your hands on software is because OpenAI/Anthropic/Microsoft/etc are hoarding billions of dollars worth of chips that they’re not even plugging in. A theory that Ed Zitron has put forth is that likely Nvidia has told their buyers that “if you don’t buy this generation’s chips, we’ll refuse to sell you the next generation chips” and so the big players in LLMs are buying way more of every generation of chips than they can use so they don’t fall behind because Nvidia is still selling newest-generation chips to their competitors. Nvidia’s just playing into the LLM companies’ bottomless well of FOMO. (Did I say it was a “bubble”?)

It used to be that most of the compute was going to training models, but that hasn’t been the case for a year now. The majority (and it’s increasing fast) is inference at this point and while it doesn’t live up to the “It will do your entire job and get you fired” it’s definitely useful in many situations.

Without hearing more specifically what “situations” you’re referring to, I can’t really comment. But any time you have a use case in mind, you should be asking yourself whether it would still be worth it at many times that cost. As well as whether you aren’t sacrificing more (more of your brain, more in terms of the quality of your output, more good will among your presumed customer base) in the long run than you’re getting in value in the short term. Furthermore, I think there’s a lot of danger of kindof losing touch with reality when using LLMs. They’re ridiculously sycophantic, and it’s not like people aren’t already going into using them with the mindset of outsourcing their thinking to the LLM. AI psychosis is obviously a really stark example, but you don’t have to think you’re the second coming, destined to wed ChatGPT who you “awakened” and is now your completely conscious soul-mate to have lost enough touch with reality to think LLMs have a use case.

I get your point, but it rests on me having said the data centers will replace all humans, so it’s either/or. It’s going to be both, even with reduction in human work force replaced by AI, there will still be people doing people things as well as AI running. And the energy costs! So if AI doesn’t consume that much energy comparably, why are some of them trying to get dedicated nuclear plants?