The Next Abstraction(substack.com)
substack.com
The Next Abstraction
https://substack.com/inbox/post/164096497
20 comments
The funny thing is that, on a purely technical level, LLMs are more likely to do a better job at replacing upper management than replacing developers. If companies really want to save money, they should let AI replace the CEO.
Management output can be vague and shift arbitrarily from day to day. It would make sense LLM slop fits in better there. We'll circle back to this action item.
But CEOs are putting a lot of money in AI and the books need to be balanced somehow.
I guess all vaguely bait-level articles will be about AI now.
Well, so much questionable here. First of all abstraction. Everything is an abstraction to programmers. It’s not, please. The author even was kind enough to contrast it with an old school bona fide abstraction. Garbage collection eliminates memory unsafety. You just don’t have to worry about it. That’s a real abstraction. It takes power away and streamlines the whole experience since you don’t have to worry about certain variables any more, they are just gone. What does AI do? It’s leverage. It might help you do things whatever factor of times faster that you insist. Completely unevenly. There is not one thing it reliably abstracts away. Please use precise words. You’re supposed to be technologists/technicians.
Then there’s the old looking to the past in order to lecture about what is hyped as completely unprecedented space-age technology. Okay to be honest this isn’t inconsistent if you merely think that AI is a great technology but not a revolution, not even a “silver bullet”. But anyway I see no reason to slavishly look to the past. The past is in fact tiny. WWII ended one person’s lifetime ago. How much oil was in the Earth 150 years ago? How much now? How reliably could you say that we could just expand economically forever 80 years ago? With climate change and whatnot, how is that looking right now?
Why not be a little cynical and pragmatic and consider that everything might not play out exactly like they did in a person’s lifetime kind of timespan. The worst could happen. What’s a billionaire with both a robot workforce and a robot army? That’s you getting discarded like the useful idiot that you are, or were. Just someone who dutifully built the whole world up for a little wage so that it could all be taken away.
Maybe you think economists are smart because they have quips against “finite piece of pie” so-called fallacies. Maybe you think that the best programmers are the ones who hustle along to the next paradigm, well those are after all the real go-getters, the ones who just get on with business. I think those are tunnel-visioned specialist fools.
Dig yourself into your specialist niche, aspire to be the hacker among hackers. Revel in embodying the values that only other members of your professional/hobbyist group respect. Meanwhile ignore the sharks of the world circling around you and get taken advantage of without any recourse or even notification.
Well, so much questionable here. First of all abstraction. Everything is an abstraction to programmers. It’s not, please. The author even was kind enough to contrast it with an old school bona fide abstraction. Garbage collection eliminates memory unsafety. You just don’t have to worry about it. That’s a real abstraction. It takes power away and streamlines the whole experience since you don’t have to worry about certain variables any more, they are just gone. What does AI do? It’s leverage. It might help you do things whatever factor of times faster that you insist. Completely unevenly. There is not one thing it reliably abstracts away. Please use precise words. You’re supposed to be technologists/technicians.
Then there’s the old looking to the past in order to lecture about what is hyped as completely unprecedented space-age technology. Okay to be honest this isn’t inconsistent if you merely think that AI is a great technology but not a revolution, not even a “silver bullet”. But anyway I see no reason to slavishly look to the past. The past is in fact tiny. WWII ended one person’s lifetime ago. How much oil was in the Earth 150 years ago? How much now? How reliably could you say that we could just expand economically forever 80 years ago? With climate change and whatnot, how is that looking right now?
Why not be a little cynical and pragmatic and consider that everything might not play out exactly like they did in a person’s lifetime kind of timespan. The worst could happen. What’s a billionaire with both a robot workforce and a robot army? That’s you getting discarded like the useful idiot that you are, or were. Just someone who dutifully built the whole world up for a little wage so that it could all be taken away.
Maybe you think economists are smart because they have quips against “finite piece of pie” so-called fallacies. Maybe you think that the best programmers are the ones who hustle along to the next paradigm, well those are after all the real go-getters, the ones who just get on with business. I think those are tunnel-visioned specialist fools.
Dig yourself into your specialist niche, aspire to be the hacker among hackers. Revel in embodying the values that only other members of your professional/hobbyist group respect. Meanwhile ignore the sharks of the world circling around you and get taken advantage of without any recourse or even notification.
Odd take, maybe the Java story made sense in say 2012 where everything would default to Java and you were building everything from a different abstraction. If you look at the past 15 years, the landscape has split into lower abstractions of C++ (Rust) for performance critical systems and some of the higher level abstractions split even to a higher level between Python, Go and others. Java is mostly missing in the AI/ML tidal wave. There is no neat - everyone moving to a higher abstraction - story.
> AI isn’t replacing us. It’s relieving us - of repetition, boilerplate, tedium.
My problem with current AI is that it isn't relieving us of boilerplate, it's making boilerplate cheaper to write. And when you make something cheaper, you get more of it! We already struggle as an industry with an overabundance of bad code. A fast bad-code generator might be a step towards a higher level of abstraction... but in the meantime it has more negative externalities than positive benefits.
Then again, the author lists "Java enterprise code" as a positive, so perhaps we just have fundamentally incompatible values and aesthetics.
My problem with current AI is that it isn't relieving us of boilerplate, it's making boilerplate cheaper to write. And when you make something cheaper, you get more of it! We already struggle as an industry with an overabundance of bad code. A fast bad-code generator might be a step towards a higher level of abstraction... but in the meantime it has more negative externalities than positive benefits.
Then again, the author lists "Java enterprise code" as a positive, so perhaps we just have fundamentally incompatible values and aesthetics.
Lisp is a good example of boilerplate elimination. Simple data structure, versatile functions, you're often just one or two layer up in terms of abstractions. Most libraries are lateral instead of foundational to your use cases. Like you don't build on top of a web framework, you mesh with it.
i feel like lisp owes more to compositionality than it owes to abstraction per se, as an antidote to complexity
I would say both. The list and its children (associated list and property list) are quite versatile. Struct in a typed language gave you benefits too, but the issue is loss of flexibility even if you have protocols like Swift or implicit interface like Go. With lisp you think in terms of data instead of bothering with names and interfaces. Most lisp functions are projections or builders.
Something similar, dx wise, could be achieved with using immutable maps, lists, and sets. But most languages relies on being able to mutate blob of memories.
Something similar, dx wise, could be achieved with using immutable maps, lists, and sets. But most languages relies on being able to mutate blob of memories.
It's also making tests easier to write, though.
I don't feel much inspired by the metaphor of garbage collection and AI.
Garbage collection makes thinking about memory irrelevant 99% of the time. Time saved with AI is spent figuring out what the AI did.
The garbage collector rarely makes itself a problem. AI almost always makes itself a problem.
Developers can go years without thinking about memory if they aren't in a complex environment. AI can't go a day without screwing up.
Garbage collection is very predictable. AI isn't.
Garbage collection makes thinking about memory irrelevant 99% of the time. Time saved with AI is spent figuring out what the AI did.
The garbage collector rarely makes itself a problem. AI almost always makes itself a problem.
Developers can go years without thinking about memory if they aren't in a complex environment. AI can't go a day without screwing up.
Garbage collection is very predictable. AI isn't.
It's just better-search. You'll spend less time searching for things. Searching for "why this is broken", searching for "how can this be better", searching for "is it possible to..."
You still have agency when that's the usage pattern. You have the problem, you use the tool to find the solution. Who cares if the one it spits out isn't character-for-character what you're looking for. Seeing the wrong solution sometimes helps you identify the right one _for you_.
If we _must compare them_ in this case, I see AI as a more useful tool than garbage collection. One that is far less invasive.
You still have agency when that's the usage pattern. You have the problem, you use the tool to find the solution. Who cares if the one it spits out isn't character-for-character what you're looking for. Seeing the wrong solution sometimes helps you identify the right one _for you_.
If we _must compare them_ in this case, I see AI as a more useful tool than garbage collection. One that is far less invasive.
Yep, we have a new tool and it's making us more productive when we learn how to use it properly. The malloc/GC example is a good analogy.
One thing to note: understanding memory management as a Java programmer makes you a better programmer. Having used malloc in bare metal mode helps you understand issues like memory leaks more clearly. It gives you a model of what is going on grounded in experience.
So that knowledge and experience of the old ways isn't useless. It's very useful and makes you a better developer. This is true of many advancements in tech, including AI.
Knowing how to build software without AI makes you a better developer when building with AI.
One thing to note: understanding memory management as a Java programmer makes you a better programmer. Having used malloc in bare metal mode helps you understand issues like memory leaks more clearly. It gives you a model of what is going on grounded in experience.
So that knowledge and experience of the old ways isn't useless. It's very useful and makes you a better developer. This is true of many advancements in tech, including AI.
Knowing how to build software without AI makes you a better developer when building with AI.
The JVM and Garbage collector were good but were already present with Lisp and Smalltalk. I've been developing professionally since '94 and the first two years of my career were with C++. In '97 I had the opportunity to move to a green field project using Java. The most useful thing for young me was not the Garbage collector, it was no header files and Javadoc.
I never really liked it, but I see its influence in Swift, every day, and I do like Swift.
I think that we are at the "unlikable Java" stage of AI, right now. In a few years, we'll be seeing the next generation of tools, and they will be pretty cool.
And no, CEOs, you won't be able to fire all of your developers, and still stay in business. The developers will just have different tools at hand.