I remember back in the 80s I had friends who enjoyed coding in assembly and felt that using higher-level languages was "cheating" - isn't this just a continuation of that?
What we've seen as we transitioned to higher and higher level languages (e.g., machine code → macro assembly → C → Java → Python) on unimaginably more powerful machines (and clusters of machines) is that we took on more complex applications and got much more work done faster. The complexity we manage shifts from the language and optimizing for machine constraints (speed, memory, etc.) to the application domain and optimizing for broader constraints (profit, user happiness, etc.).
I think LLMs also revive hope that natural languages (e.g., English) are the future of software development (COBOL's dream finally be realized!). But a core problem with that has always been that natural languages are too ambiguous. To the extent we're just writing prompts and the models are the implementers, I suspect we'll come up with more precise "prompt languages". At that point, it's just the next generation of even higher level languages.
So, I think you're right that we'll spend more of our time thinking like product managers. But also more of our time thinking about higher level, hard, technical problems (e.g., how do we use math to build a system that dynamically optimizes itself for whatever metric we care about?). I don't think these are new trends, but continuing (maybe accelerating?) ones.
It’s likely that a near-future AI system can suggest suitable math and implement it in an algorithm for the problem the user wants solved. An expert who understands it might be able to critique and ask for a better solution, but many users could be satisfied with it.
Professionals who can deliver added value are those who understand the user better than the user themselves.
When these optimization systems (I'm referring to mathematical optimization here) are unleashed, they will crush many metrics that are not a part of their objective function and/or constraints. Want to optimize this quarter's revenue and don't have time to put in a constraint around user happiness? Revenue might be awesome this quarter, but gone in a year because the users are gone.
The system I worked on kept our company in business through the pandemic by automatically adapting to frequently changing market conditions. But we had to quickly add constraints (within hours of the first US stay-at-home orders) to prevent gouging our customers. We had gouging prevention in before, but it suddenly changed in both shape and magnitude - increasing prices significantly in certain areas and making them free in others.
AI is trained on the past, but there was no precedent for such a system in a pandemic. Or in this decade's wars, or under new regulations, etc. What we call AI today does not use reason. So it's left to humans to figure out how to adapt in new situations. But if AI is creating a black-box optimization system, the human operators will not know what to do or how to do it. And if the system isn't constructed in a mathematically sound way, it won't even be possible to constrain it without significant negative implications.
Gains from such systems are also heavily resistant to measurement, which we need to do if we want to know if they are breaking our business. This is because such systems typically involve feedback loops that invalidate the assumption of independence between cohorts in A/B tests. That means advanced experiment designs must be found that are often custom for every use case. So, maybe in addition to thinking more like product managers, engineers will need to be thinking more like data scientists.
This is all just in the area where I have some expertise. I imagine there are many other such areas. Some of which we haven't even found yet because we've been stuck doing the drudgery that AI can actually help with. [cue the song Code Monkey]
Yes, unless models are being live fine-tuned, but generally yes.
>What we call AI today does not use reason
I don’t think this is correct- I think it’s more accurate to say it reasons on its priors rather than from first principles.
For the most part, I agree with the rest of your post.
The increase in productivity, we can all agree on, but a non-negligible portion of HN users would say that each one of those new languages made programming progressively less fun.
For instance I think the step from machine code to macro assembler is bigger than the step from a macro assembler to C (although still substantial), but the step from C to anything higher level is essentially negligible compared to the massive jump from machine code to a 'low level high level' language like C.
For instance, say that C had namespaces, and a solid package system with a global repo of packages like Python, C# and Java have.
Then you'd be able to throw together things pretty easily.
Things easily cobbled together with Python often aren't attributable to Python the language per se, but rather Python, the language and its neat packages.
This made me laugh out loud. Python is not a step up from Java in my opinion. Python is more of a step up from BASIC. It's a different evolutionary path. Like LISP.
Millions, perhaps billions of times more lines of code will be written, and automated programming will be taken for granted as just how computers work.
Painstakingly writing static source code will be seen the same way as we see doing hundreds of pages of tedious calculations using paper, pencil, and a slide rule. Why would you do that, when the computer can design and develop such a program hundreds of times in the blink of an eye to arrive at the optimal human interface for your particular needs at the moment?
It'll be a tremendous boon in every other technical field, such as science and engineering. It'll also make computers so much more useful and accessible for regular people. However, programming as we know it will fade into irrelevance.
This change might take 50 years, but that's where I believe we're headed.
An AI can only do what it is taught to do. Sure, it can offer unique insights from time to time, but I doubt it will get to the point where it can craft entirely new paradigms and ways of building software.
(Humans will still have to set the goals and objectives, unless we unleash an ASI and render even that moot.)
Of course, code written by human programmers on the lower end of the skill spectrum sometimes has similar problems...
I mean Ken Thompson put a back door into the C compiler no one ever found. Can you imagine what an AI could be capable of?
Maybe we'll start to evolve as a species to avoid that, but AI will be used to ensure we don't, optimising far faster than we can evolve to keep our attention
[0] https://bigthink.com/neuropsych/social-media-profound-boredo...
If I stop making progress on my personal projects, sinking my free time into games or online interaction is very unsatisfying.
Yeah, the same time the singularity happens, and then your smallest problem will be eons bigger than your job.
But LLMs can’t solve a sudoku, so I wouldn’t be too afraid.
It’s literally part of its training data. The same way it knows how to solve leetcode, etc.
I also don't know what it means to completely remove humans from all work. Who is deciding what we want done? What we want to investigate or build? The machines are just gong to make all work-related decisions for us? I don't believe that. It would cease being our society at that point.
Which brings up the heart of the matter. Why are we trying to replace ourselves? It's our civilization, automation are just tools we use to be more productive. It should make our lives better, not remove us from the equation.
My guess is the real answer is it will make some people obscenely rich, and give some governments a significant technical advantage over others.
There's already so many layers that essentially no one knows them all at even a basic level, let alone expert. A few more layers and no one in the field will even know of all the layers.
Not in the near near future. Do you know anything about nursing? The field will require some hard changes for robots to replace nurses, and the robots will need licenses
For example, when moving up to C from assembler, the task of the "programmer" remains invariant and the language tool affords broader accessibility to the profession since not everyone likes to flip bytes. There is no subdivision of overall task of "coding a software product".
With AI coders, there is task specialization, and, as pointed out, what's left on the table is the least appetizing of the software tasks: being a patch monkey.
This is the issue.
"Automatic programming always has been a euphemism for programming with a higher level language than was then available to the programmer. Research in automatic programming is simply research in the implementation of higher-level languages.
Of course automatic programming is feasible. We have known for years that we can implement higher-level programming languages. The only real question was the efficiency of the resulting programs. Usually, if the input 'specification' is not a description of an algorithm, the resulting program is woefully inefficient. I do not believe that the use of nonalgorithmic specifications as a programming language will prove practical for systems with limited computer capacity and hard real-time deadlines. When the input specification is a description of an algorithm, writing the specification is really writing a program. There will be no substantial change from out present capacity.
The use of improved languages has led to a reduction in the amount of detail that a programmer must handle and hence to an improvement in reliability. However, extant programming languages, while far from perfect, are not that bad. Unless we move to nonalgorithmic specifications as an input to those systems, I do not expect a drastic improvement to result from this research.
On the other hand, our experience in writing nonalgorithmic specifications has shown that people make mistakes in writing them just as they do in writing algorithms."
Programming with AI, so far, tries to specify something precise, algorithms, in a less precise language than what we have.
If AI programming can find a better way to express the problems we're trying to solve, then yes, it could work. It would become a matter of "how well the compiler works". The current proposals, with AI and prompting, is to use natural language as the notation. That's not better than what we have.
It's the difference between Euclid and modern notation, with AI programming being like Euclidean notation and current programming languages being the modern notation:
"if a first magnitude and a third are equal multiples of a second and a fourth, and a fifth and a sixth are equal multiples of the second and fourth, then the first magnitude and fifth, being added together, and the third and sixth, being added together, will also be equal multiples of the second and the fourth, respectively."
a(x + y) = ax + by
You can't make something simpler by making it more complex.
The distance isn’t the same between them, but each one is more abstracted than the next.
Natural language can be ambiguous and ill defined. Because the compiler is smarter. Just like you don’t have to manage memory in Python, except it abstracts a lot more.
The fact is that this very instant you can compile from natural language.
Python is closer to C (third generation programming language). Excel is a higher level example. It still takes someone who knows how to use Excel to do anything meaningful.
I think this will weed out the people doing tech purely for the sake of tech and will bring more creative minds who see the technology as a tool to achieve a goal.
I personally wouldn't have enjoyed being that kind of programmer as it was a tedious and very slow process, where the creativity of the developer was rather low as the complexities of development would not allow for just anyone to be part of it (my own assumption).
Today we have IDEs, autocomplete, quick visual feedback (inspectors, advanced debuggers, etc.) which allow people who enjoy creating and seeing the results of their work as opposed to purely be typing code for someone else.
So, I don't get why people jump straight to thinking that adding yet another efficiency tool would destroy everything. To me it seems to make developing simpler applications something which doesn't require a computer science degree, that's all.
The average attorney became much more productive after the introduction of the word processor.
I think your assumption is incorrect. I remember programming using punched cards and low-level languages, and the amount of creativity involved was no less than is involved now.