> But. And here's where the comparison to cloud comes in; the details of that evolution seem a bit fuzzy.
Maybe I have rose-tinted glasses on, but cloud computing was never "fuzzy" the way LLMs are. Cloud offerings were (and even moreso now, are) platforms. At the time the concept of a technical platform was very well understood with plenty of prior art. .NET is an example that leaps to mind. The trade-off was you give up control and submit to vendor lock-in, but the platform abstracts away small details so you can focus on your business. In short, cloud wasn't a huge leap, conceptually.
With LLMs, conversely, there isn't really much you can point to and say "this is a natural progression of ____". It's an entirely new thing, with entirely new problems.
But as someone who lived through the introduction of cloud computing, there were a lot of widely-held assumptions that didn't really play out. Heck, you see a lot of people here who are regularly up in arms about the assumption that of course cloud must be cheaper. Or look at The Big Switch by Nick Carr which recounts the narrative about how centralized power generation took the place of individual factories generating their own power--the irony being that solar in particular has decentralized power generation to a significant degree.
The platform/infra distinction applies here. Remember that pre-cloud there was a short lived category called "application service providers" that would offer some app for you to use. Salesforce used to aggressively advertise "no more software!" and over on the IT side of the house people like Citrix were simulating multi-user NT with GUI terminal services. Plus the whole conversation around TCO and how thin clients might be the way forward.
All that was solving a (local) infrastructure problem. Cloud moved the infra to someone else's data center but it did not take long before they started allowing anyone to consume the back end services as a platform. (My pet theory is that S3 is the most industry-changing software written in the last 30 years...)
Today, people are locked into the platform. I can't move from X to Y because my business is welded to AWS S3, Lambda or Azure Kubernetes or whatever.
ASPs essentially got reimagined as SaaS by, most visibly, Salesforce early on. They weren't the only example at the time but they were the ones on stage at every remotely relevant conference.
And, yes, there were hopes for portability early on even though it was clear that the cloud providers weren't much interested and it became harder as services became more complex. There are some glimmers of portability--such as using my former employer's OpenShift for Kubernetes--but single-pane-of-glass management and API translations between cloud providers never really panned out.
I worked with IBM at one time, who were convinced that they could build a real time workload arbitrage system, whereby loads could instantly move from one cloud to another based on spot price rates offered by cloud providers. Nice idea that quickly ran into the reality of platform stickiness (and sweet reserved-instance pricing).
Yeah. In my prior analyst stint, there was a lot of automagical cloudbursting talk (and migration between clouds quickly based on pricing). In addition to the platform stickiness you mention, there was the matter of, as Scotty would have said: You canna change the laws of physics.
What does not feel too risky to predict, though, are some general directions:
a) the era of "GPU"-style computing is here to stay. During the long era of exponential CPU speedups the architectures of vectorized computing were very niche (HPC). Going forward its clear there are potentially various economically viable "mass-market" applications of linear algebra. This may even change the economics building of silicon chips from the ground up. Which brings us to the other main point,
b) the era of algorithmic computing is also just starting. Right now there is an almost maniacal obsession with LLM's. Its not an entirely useless hype as it is trailblazing a path where much else can follow. But conceptually its just one little corner in the vast space of data processing algorithms.
While the general direction of travel seems reasonably established (for now), the details of what comes to pass depend a lot both on the aforementioned economics and the governance around the use of algorithms. Thus far the tech industry had a free pass. Its unlikely that this will continue.
Maybe I'm slow but I'm failing to read the part where AI reminds the author of specifically cloud computing. The general premise seems to be "lot of early promises never panned out"... but then you can say that about pretty much any fads or exciting trends.
It wasn't so much lots of early promises never panned out as that the dynamics of how the technology ends up being different than how people thought it would be. I admit I skipped over the specifics of cloud computing because it was already a fairly long post just focused on AI. (The general premise as you paraphrased it and the point I was trying to make are somewhat related but they are different.)
I don't like the comparison.
> But. And here's where the comparison to cloud comes in; the details of that evolution seem a bit fuzzy.
Maybe I have rose-tinted glasses on, but cloud computing was never "fuzzy" the way LLMs are. Cloud offerings were (and even moreso now, are) platforms. At the time the concept of a technical platform was very well understood with plenty of prior art. .NET is an example that leaps to mind. The trade-off was you give up control and submit to vendor lock-in, but the platform abstracts away small details so you can focus on your business. In short, cloud wasn't a huge leap, conceptually.
With LLMs, conversely, there isn't really much you can point to and say "this is a natural progression of ____". It's an entirely new thing, with entirely new problems.
But as someone who lived through the introduction of cloud computing, there were a lot of widely-held assumptions that didn't really play out. Heck, you see a lot of people here who are regularly up in arms about the assumption that of course cloud must be cheaper. Or look at The Big Switch by Nick Carr which recounts the narrative about how centralized power generation took the place of individual factories generating their own power--the irony being that solar in particular has decentralized power generation to a significant degree.
The platform/infra distinction applies here. Remember that pre-cloud there was a short lived category called "application service providers" that would offer some app for you to use. Salesforce used to aggressively advertise "no more software!" and over on the IT side of the house people like Citrix were simulating multi-user NT with GUI terminal services. Plus the whole conversation around TCO and how thin clients might be the way forward.
All that was solving a (local) infrastructure problem. Cloud moved the infra to someone else's data center but it did not take long before they started allowing anyone to consume the back end services as a platform. (My pet theory is that S3 is the most industry-changing software written in the last 30 years...)
Today, people are locked into the platform. I can't move from X to Y because my business is welded to AWS S3, Lambda or Azure Kubernetes or whatever.
ASPs essentially got reimagined as SaaS by, most visibly, Salesforce early on. They weren't the only example at the time but they were the ones on stage at every remotely relevant conference.
And, yes, there were hopes for portability early on even though it was clear that the cloud providers weren't much interested and it became harder as services became more complex. There are some glimmers of portability--such as using my former employer's OpenShift for Kubernetes--but single-pane-of-glass management and API translations between cloud providers never really panned out.
I worked with IBM at one time, who were convinced that they could build a real time workload arbitrage system, whereby loads could instantly move from one cloud to another based on spot price rates offered by cloud providers. Nice idea that quickly ran into the reality of platform stickiness (and sweet reserved-instance pricing).
Yeah. In my prior analyst stint, there was a lot of automagical cloudbursting talk (and migration between clouds quickly based on pricing). In addition to the platform stickiness you mention, there was the matter of, as Scotty would have said: You canna change the laws of physics.
I mean, I think as sold to laypeople, 'cloud' was _very_ fuzzy; it was sold as a magic thing of rather vague nature which would fix everything.
"this is a natural progression of _evolution_"? Chemistry -> DNA evolution -> cultural evolution -> AI evolution?
But yeah it's a new thing.
> the details are hard to predict
What does not feel too risky to predict, though, are some general directions:
a) the era of "GPU"-style computing is here to stay. During the long era of exponential CPU speedups the architectures of vectorized computing were very niche (HPC). Going forward its clear there are potentially various economically viable "mass-market" applications of linear algebra. This may even change the economics building of silicon chips from the ground up. Which brings us to the other main point,
b) the era of algorithmic computing is also just starting. Right now there is an almost maniacal obsession with LLM's. Its not an entirely useless hype as it is trailblazing a path where much else can follow. But conceptually its just one little corner in the vast space of data processing algorithms.
While the general direction of travel seems reasonably established (for now), the details of what comes to pass depend a lot both on the aforementioned economics and the governance around the use of algorithms. Thus far the tech industry had a free pass. Its unlikely that this will continue.
I don't think I got anything from this article over and above what has already been written about AI elsewhere and in greater depth and detail.
Maybe I'm slow but I'm failing to read the part where AI reminds the author of specifically cloud computing. The general premise seems to be "lot of early promises never panned out"... but then you can say that about pretty much any fads or exciting trends.
It wasn't so much lots of early promises never panned out as that the dynamics of how the technology ends up being different than how people thought it would be. I admit I skipped over the specifics of cloud computing because it was already a fairly long post just focused on AI. (The general premise as you paraphrased it and the point I was trying to make are somewhat related but they are different.)
The author flutters around but doesn't land on a point.
Expect the unexpected, basically.
Not a useful article.