The fifth epoch of distributed computing
cloud.google.com
cloud.google.com
Eugh, this article is full of MBA-speak, not something I'd expect from a highly technical Google Fellow.
https://www.salesforce.com/eu/blog/industry-4-0-where-are-we....
Mobile has extended these walls with mafia racketeering app stores and absolutely atrociously designed gambling-addict games, information stealing, and outright fraud.
The dream of the internet died completely about five years ago.
This overpaid google shill isn't worth listening to. Google had about a decade of plausable deniability on "do no evil". They stopped the charade, and we know what they are. Like sociopaths always do, they tell you and you should listen.
AI is an even more intrusive, dangerous penetration and control of our lives, and a massive power grab by these tools of the very very very elite few.
The guy is right: the internet WAS about access to the evolving body of human knowledge, WAS about sci-fi level capabilities and conveniences, and WAS a miracle.
WAS.
AI is the rocket fuel that accelerates the vector from the momentary point of near-idealism to the dystopian corporate control we are currently mostly in, that as a DOUBLE BONUS serves as fine grained total information awareness for all the state actors of the world, and likely a worldwide destruction of the last semblances of democracy.
Sometimes they change the order, but it's always the same claims, and it's always the 'unique' circumstances of (insert date here) that make everything different this time. I wonder why they bother?
I wrote that before the AI revolution, and I honestly don't know what's going to happen. The future could go one of two ways:
1. AI just becomes a smarter compiler. We programmers still "write" code, though maybe at a higher level. AI helps us deal with complexity, but we're still the ones designing UX, architecture, etc.
2. AI becomes the front-end layer for everything, so that users only interact with an AI and that AI carries out tasks as needed. All other software is just an implementation detail (the way we software people think about hardware).
Most likely we will start with #1 and evolve into #2.
Engineers need to control the behavior of their software very precisely. The fantasy of firing all your engineers and just letting product folks write something extremely high-level like "Write me an app for reviewing movies" and letting a language model fill in all the details seems like a nonstarter.
First, the details it imagines likely won't be what you actually want. So they'll have to start getting more and more precise. So precise in fact that it's back to "programming" instead of something like actual natural language.
Second, this seems unlikely to provide stability over time. Trying to evolve your FE, BE and storage to remain compatible between one prompt and the next or one language model and the next likely will involve specifying those details fully as well.
Natural language just isn't precise enough for this. There is way too much inference involved. This situation is fundamentally different from abstractions that allow one group of engineers to encapsulate implementation details within a system.
Imagine you're a non-technical person and you communicate with a programmer only via email. You could still get your product built.
Is it really hard to imagine than an LLM could (someday) be on the other side of the email?
As for stability, I think that's something that LLMs have an advantage in. Imagine the user and the LLM create a set of regression tests. The LLM can patiently run the tests on every change.
The goal isn't perfection (LLM reads my mind and instantly creates a program). The goal is better/faster/cheaper than dealing with a human programmer.
Believe me, I know LLMs today are not there yet. The question is how long will they take to get there? My guess is 10-20 years, and I'm probably a pessimist.
i see what you mean. but I wonder if this is because we have difficulty imagining a totally possible thing. i mean, surely at one point it must have seemed like less machine code engineers, and more script monkeys writing something extremely high level, was a nonstarter. or perhaps firing system software engineers and just letting FE devs write something extremely high level must have seemed like a nonstarter at one point.
but here we are.
> natural language just isn't precise enough for this
i often make the mistake of thinking business needs something more precise than what it is actually asking for (or what it truly needs). of course precise dollars and cents truly matter in some situations, but in other situations a crud app or report has a great deal of flexibility in its requirements. not willing to die on this hill, but i wonder if precise software and precise computer modelling is because the tools require it to be precise? and if so, a less precise interface tool may well lead to less precise inputs that still meet the underlying business requirements.
A waypoint, not an end goal, because surely someone will derive sellable business value from combining things in a world where AI is the near-universal front layer.
This line of thought has never occurred to me before. The idea that complexity will be so unmanageable that we’ll need to rewrite the computing stack in a more easily debuggable, functional way. Modern computing seems simple: The kernel and your application on top of it, but the existence of solutions like Antithesis suggests something different.
Not being able to hold a TCP connection open is an implementation choice of the OS developer (and some choices at the network level too). At WhatsApp, we would find Nokia Symbian devices with 45 day old TCP connections to our servers anytime we looked. Nokia S40 isn't a smart phone platform, but it could hold long connections too. Old versions of Android could do it, although old versions of Android had trouble switching between wifi and cellular. You can probably do it on current Android if you turn off all the Doze or whatever stuff (which isn't always easy to turn off).
There's an efficiency argument for having a single long TCP connection for platform push, rather than one per app that needs it, of course. But in that case, platform push better be reliable.
This line of thought has never occurred to me before. The idea that complexity will be so unmanageable that we’ll need to rewrite the computing stack in a more easily debuggable, functional way. Modern computing seems simple: The kernel and your application on top of it, but the existence of solutions like Antithesis suggests something different.
They call that point out, basically, but I had no idea it'd be a 100x improvement.
> While we cannot predict the breakthroughs that will be delivered in this fifth epoch of computing, we do know that each previous epoch has been characterized by a factor of 100x improvement in scale, efficiency, and cost-performance, all while improving security and reliability.
I'm worried about 5th gen epoch engineers. On one side there is a massive amount of smart rust programmers who are super young in all the rust communities just as an example and on the other a lot of people who don't know anything beyond a gui operating system and have never owned a computer and do everything by phone.
I've been happily teaching linux quite a bit lately.
I think actual Rust programers maybe in reasonable numbers but RIIR evangelists would be in massive numbers.
In the future the only thing you the engineer will do with a computer is type into a text area, hit the submit button, and be returned variations on strings of text that seem to say something important but upon careful considering do not mean anything at all.
We can only hope the next epoch will bring about some kind of enlightenment that will herald a new era of blog posts that convey information across the Internet.
Our current LLM's excel at nondeterministic problems (style transfer, summarization, etc), not at deterministic ones.
There is going to be an astounding gap between those in the know and those who are not.
Assuming some form of ubiquity for AI tooling people like him are going to build amazing things.
It this the 5th generation of computing, Japanese were talking about in the 90's ? It took some time... /s
https://en.wikipedia.org/wiki/Fifth_Generation_Computer_Syst...
Long memories I guess :-)