> My concerns about obsolescence have shifted toward curiosity about what remains to be built. The accidental complexity of coding is plummeting, but the essential complexity remains. The abstraction is rising again, to tame problems we haven't yet named.
what if AI is better at tackling essential complexity too?
The essential complexity isn't solvable by computer systems. That was the point Fred Brooks was making.
You can reduce it by process re-engineering, by changing the requirements, by managing expectations. But not by programming.
If we get an LLM to manage the rest of the organisation, then conceivably we could get it to reduce the essential complexity of the programming task. But that's putting the cart before the horse - getting an LLM to rearrange the organisation processes so that it has less complexity to deal with when coding seems like a bad deal.
And complexity is one of the things we're still not seeing much improvement in LLMs managing. The common experience from people using LLM coding agents is that simple systems become easy, but complex systems will still cause problems with LLM usage. LLMs are not coping well with complexity. That may change, of course, but that's the situation now.
StrongDM is doing it. In fact, their Attractor agentic loop, which generates, tests, and deploys code written as specs, has been released—as a spec, not code. Their installation instructions are pretty much "feed this into your LLM". They are building out not only complete applications, but test harnesses for those applications that clone popular web apps like Slack and JIRA, with no humans in the loop beyond writing the initial spec and giving final approval to deploy.
We're witnessing a "horses to automobile" moment in software development. Programming, as a professional discipline, is going to be over in a year or two on the outside. We're getting the "end of software engineering in six months" before we're getting a real "year of the Linux desktop". Or GTA VI.
> With the price of computation so high, that inefficiency was like lighting money on fire. The small group of contributors capable of producing efficient and correct code considered themselves exceedingly clever, and scoffed at the idea that they could be replaced.
There will always be someone ready to drag down prices of computation low enough so that it is then democratized for all, some may disagree but that would eventually be local inference as computer hardware gets better with clever software algorithms.
In this AI story, you can take a guess who are the "The Priesthood" of the 2020s are.
> You still have to know what you want the computer to do, and that can be very hard. While not everyone wrote computer programs, the number of computers in the world exploded.
One can say, the number of AI agents will explode and surpass humans on the internet in the next few years, and reading the code and understanding what it does when generated from an AI will be even more important than writing it.
So you do not get horrific issues like this [0] since now the comments in the code are now consumed by the LLM and due to their inherent probabilistic and unpredictable nature, different LLMs produce different code and cannot guarrantee that it is correct other than a team of expert humans.
We'll see if you're ready to read (and fix) an abundance of lots of AI slop and messy architectures built by vibe-coders as maintainance costs and security risks skyrocket.
It's funny, but I think the accidental complexity is through the roof. It's skyrocketing.
Nothing about cajoling a model to write what you want it to is essential complexity in software dev.
In addition when you do a lot of building with no theory you tend you make lots and lots of new non-essential complexity.
Devtools are no exception. There was already lots of nonessential complexity in them and in the model era is that gone? ...no don't worry it's all still there. We built all the shiny new layers right on top of all the old decaying layers, like putting lipstick on a pig.
More like building a house in a bottomless swamp. The moment you finish it starts sinking, so you build a new floor, but it just makes it heavier, and speed up the sinking speed.
Bro it's fine everyone's doing it get to the program. We all just need to use multiple agents with memory and skills and orchestrate them, connect them to ticketing systems with workload federation and let them create lambdas to push artifacts from a CI that has audit logs and snyk scanning, let them spin up a few kubernetes clusters per commit, then write the test suites with headless chrome and simulated agents that run A/B testing with multiple backups, regional HA, SSO, vertical and horizontal autoscaling, otel agents that rewrite what they collect based on other agentic processes that also run via lambdas that monitor datadog and splunk and sentry, automated PRs and red teaming. If you don't think about all of that even when you sleep do you even care about the customer?
I think a reasonable and sensible goal is for us to not mix the accidental and the essential. If we let AI handle what's accidental (as in not central to the solution of the essential problem) developers can focus on the essential only. The current threat is that both types become intertwingled in a code-base, sometimes irreparably.
Fortran made that distinction clear. The compiler handled the accidental complexity of converting instructions to code, but never really obscured the boundary.
Take VB as an example from wayback. For the purposes of presenting a simple data-entry dialog, it removed the accidental complexity of dealing with Windows' message loop and resource files etc., which was painful. The essential complexity was in what the system did with the data. I suppose that the AI steering that needs to happen is to direct the essential down the essential path, and the accidental down the accidental path, and let a dev handle the former and the agent handle the latter (after all, it's accidental).
But, that'll take judgement - deciding in which camp each artifact exists and how it's managed. It might be a whole field of study, but it won't be new.
Fortran is all about symbolic programming. There is no probability in the internal workings of Fortran compiler. Almost any person can learn rules and count on them.
LLMs are all about probabilistic programming. While they are harnessed by a lot of symbolic processing (tokens as simple example), the core is probabilistic. No hard rules can be learned.
And, for what it worth, "Real programmers don't use Pascal" [1] was not written about assembler programmers, it was written about Fortran programmers, a new Priesthood.
To the point on Jevons Paradox, the number of people/developers joining GitHub had been accelerating as of the last Octoverse report. Related: "In 2023, GitHub crossed 100 million developers after nearly three years of growth from 50 million to 100 million. But the past year alone has rewritten that curve with our fastest absolute growth yet. Today, more than 180 million developers build on GitHub."
Not a fan of looking at history for cases that look like the could be a step change - a new paradigm. For that it seems safer to extrapolate out from recent experiences. Normally that’s a bad idea but if you’re in uncharted territory it’s the only reference point
LLM coding isn't a new level of abstraction. Abstractions are (semi-)reliable ways to manage complexity by creating building blocks that represent complex behavior, that are useful for reasoning about outcomes.
Because model output can vary widely from invocation to invocation, let alone model to model, prompts aren't reliable abstractions. You can't send someone all of the prompts for a vibecoded program and know they will get a binary with generally the same behavior. An effective programmer in the LLM age won't be saving mental energy by reasoning about the prompts, they will be fiddling with the prompts, crossing their fingers that it produces workable code, then going back to reasoning about the code to ensure it meets their specification.
What I think the discipline is going to find after the dust settles is that traditional computer code is the "easiest" way to reason about computer behavior. It requires some learning curve, yes, but it remains the highest level of real "abstraction", with LLMs being more of a slot machine for saving the typing or some boilerplate.
Going from programming language to LLM is not the same kind of abstraction as going from assembler to Fortran or Algol.
With programming languages, there is a transparent homomorphism between the code you write and what the machine actually executes. Programmers use this property of programming languages to execute considerable control over the computational process they're evolving while still potentially at a high level of abstraction. With LLMs, the mapping between your input and the executable output is opaque and nondeterministic. This drives people like me batty.
"A student travelled to the East to hear Sussman's teachings. Sussman was expounding on low-level development in Lisp, when the student interrupted him. 'Master Sussman,' said he, 'is not Lisp a high-level language, indeed perhaps the highest-level of all programming languages?' Sussman replied, 'Once there was a man who was walking along the beach when he spotted an eagle. Brother eagle, said he, how impossibly distant is the sky! The eagle said nothing, and flew away.' Thus the student was enlightened."
The story is in some sense true: I did meet Gerald Sussman and was in some sense enlightened by him. Another hacker was talking about working in a "low-level language" like Lisp, and I corrected him telling him that Lisp was in fact very high level. He said "Uh... I need Jerry to explain this to you. Jerry? Can you come here a minute?" "Jerry" was Gerald Sussman, who proceeded to explain to me that Lisp was a virtual machine, one which he implemented in silico for his Ph.D. thesis:
Thus I was enlightened. If Lisp is a virtual machine, so are all programming languages. And even a buck JavaScript kiddie fresh out of boot camp, working in React, is working in machine code for the JavaScript+browser+React VM. An abstraction is a point of view, and the programmer working in a "high level" programming language is really just working in the same medium as machine code: computation itself. But with a point of view that offers more convenience.
LLM work is different. LLMs are enormously complicated algorithms that give probabilistic interpretations of loose, informal human languages. So instructing a computer through the filter of an LLM is inherently probabilistic, not to mention damn near inscrutable.
This is why LLMs are being met with more resistance even than compilers were. They're not the same thing. Compilers scaled the work, which remained essentially the same. LLMs are changing it.
18 comments
[ 3.3 ms ] story [ 41.8 ms ] threadwhat if AI is better at tackling essential complexity too?
You can reduce it by process re-engineering, by changing the requirements, by managing expectations. But not by programming.
If we get an LLM to manage the rest of the organisation, then conceivably we could get it to reduce the essential complexity of the programming task. But that's putting the cart before the horse - getting an LLM to rearrange the organisation processes so that it has less complexity to deal with when coding seems like a bad deal.
And complexity is one of the things we're still not seeing much improvement in LLMs managing. The common experience from people using LLM coding agents is that simple systems become easy, but complex systems will still cause problems with LLM usage. LLMs are not coping well with complexity. That may change, of course, but that's the situation now.
Ah, a work of fiction.
We're witnessing a "horses to automobile" moment in software development. Programming, as a professional discipline, is going to be over in a year or two on the outside. We're getting the "end of software engineering in six months" before we're getting a real "year of the Linux desktop". Or GTA VI.
There will always be someone ready to drag down prices of computation low enough so that it is then democratized for all, some may disagree but that would eventually be local inference as computer hardware gets better with clever software algorithms.
In this AI story, you can take a guess who are the "The Priesthood" of the 2020s are.
> You still have to know what you want the computer to do, and that can be very hard. While not everyone wrote computer programs, the number of computers in the world exploded.
One can say, the number of AI agents will explode and surpass humans on the internet in the next few years, and reading the code and understanding what it does when generated from an AI will be even more important than writing it.
So you do not get horrific issues like this [0] since now the comments in the code are now consumed by the LLM and due to their inherent probabilistic and unpredictable nature, different LLMs produce different code and cannot guarrantee that it is correct other than a team of expert humans.
We'll see if you're ready to read (and fix) an abundance of lots of AI slop and messy architectures built by vibe-coders as maintainance costs and security risks skyrocket.
[0] https://sketch.dev/blog/our-first-outage-from-llm-written-co...
Nothing about cajoling a model to write what you want it to is essential complexity in software dev.
In addition when you do a lot of building with no theory you tend you make lots and lots of new non-essential complexity.
Devtools are no exception. There was already lots of nonessential complexity in them and in the model era is that gone? ...no don't worry it's all still there. We built all the shiny new layers right on top of all the old decaying layers, like putting lipstick on a pig.
More like building a house in a bottomless swamp. The moment you finish it starts sinking, so you build a new floor, but it just makes it heavier, and speed up the sinking speed.
Fortran made that distinction clear. The compiler handled the accidental complexity of converting instructions to code, but never really obscured the boundary.
Take VB as an example from wayback. For the purposes of presenting a simple data-entry dialog, it removed the accidental complexity of dealing with Windows' message loop and resource files etc., which was painful. The essential complexity was in what the system did with the data. I suppose that the AI steering that needs to happen is to direct the essential down the essential path, and the accidental down the accidental path, and let a dev handle the former and the agent handle the latter (after all, it's accidental).
But, that'll take judgement - deciding in which camp each artifact exists and how it's managed. It might be a whole field of study, but it won't be new.
LLMs are all about probabilistic programming. While they are harnessed by a lot of symbolic processing (tokens as simple example), the core is probabilistic. No hard rules can be learned.
And, for what it worth, "Real programmers don't use Pascal" [1] was not written about assembler programmers, it was written about Fortran programmers, a new Priesthood.
[1] https://web.archive.org/web/20120206010243/http://www.ee.rye...
Thus, what I expect is for new Priesthood to emerge - prompt writing specialists. And this is what we see, actually.
https://github.blog/news-insights/octoverse/octoverse-a-new-...
Because model output can vary widely from invocation to invocation, let alone model to model, prompts aren't reliable abstractions. You can't send someone all of the prompts for a vibecoded program and know they will get a binary with generally the same behavior. An effective programmer in the LLM age won't be saving mental energy by reasoning about the prompts, they will be fiddling with the prompts, crossing their fingers that it produces workable code, then going back to reasoning about the code to ensure it meets their specification.
What I think the discipline is going to find after the dust settles is that traditional computer code is the "easiest" way to reason about computer behavior. It requires some learning curve, yes, but it remains the highest level of real "abstraction", with LLMs being more of a slot machine for saving the typing or some boilerplate.
With programming languages, there is a transparent homomorphism between the code you write and what the machine actually executes. Programmers use this property of programming languages to execute considerable control over the computational process they're evolving while still potentially at a high level of abstraction. With LLMs, the mapping between your input and the executable output is opaque and nondeterministic. This drives people like me batty.
A while back I wrote a Lisp koan:
https://www.hackersdictionary.com/html/Some-AI-Koans.html
Mine goes like this:
"A student travelled to the East to hear Sussman's teachings. Sussman was expounding on low-level development in Lisp, when the student interrupted him. 'Master Sussman,' said he, 'is not Lisp a high-level language, indeed perhaps the highest-level of all programming languages?' Sussman replied, 'Once there was a man who was walking along the beach when he spotted an eagle. Brother eagle, said he, how impossibly distant is the sky! The eagle said nothing, and flew away.' Thus the student was enlightened."
The story is in some sense true: I did meet Gerald Sussman and was in some sense enlightened by him. Another hacker was talking about working in a "low-level language" like Lisp, and I corrected him telling him that Lisp was in fact very high level. He said "Uh... I need Jerry to explain this to you. Jerry? Can you come here a minute?" "Jerry" was Gerald Sussman, who proceeded to explain to me that Lisp was a virtual machine, one which he implemented in silico for his Ph.D. thesis:
https://dspace.mit.edu/handle/1721.1/5731
Thus I was enlightened. If Lisp is a virtual machine, so are all programming languages. And even a buck JavaScript kiddie fresh out of boot camp, working in React, is working in machine code for the JavaScript+browser+React VM. An abstraction is a point of view, and the programmer working in a "high level" programming language is really just working in the same medium as machine code: computation itself. But with a point of view that offers more convenience.
LLM work is different. LLMs are enormously complicated algorithms that give probabilistic interpretations of loose, informal human languages. So instructing a computer through the filter of an LLM is inherently probabilistic, not to mention damn near inscrutable.
This is why LLMs are being met with more resistance even than compilers were. They're not the same thing. Compilers scaled the work, which remained essentially the same. LLMs are changing it.