Some of the best papers of past were only 2 or 3 pages. Certainly not this one though, that's not even a paper, that's just an extended online comment, letter to the editor, journal preface.
"It's ok if it's true as I figure Professor of AI will be the last job title standing" (Patrick Henry Winston, MIT AI & The Future of Work Conference 2nd Nov 2017)
Hasn't it done so already? To a first approximation, none of us spends significant time coding in machine language; rather, we write in higher-level languages which heuristics-based artificial intelligences then compile down to machine language. Why, some of us use yet more-advanced AIs which can even perform certain safety checks for us, to ensure e.g. that we aren't using kilometres per hour when we ought to be using inches per second.
I'm not just being coy: as long as there are computers generating code, there will be a need for people to tell those computers what to generate. Those people will be coders, whatever they are called at the time.
Indeed. "Software developer" > "coder", and the difference will only become more stark as machines become more advanced, giving the developers more powerful tools.
The key ability of a good developer is finding out and understanding what needs to be done, the translation form the language of business to the language of technology. Technology has moved from low-level register juggling of 1950s to modern high-level, more and more declarative languages and frameworks. But the need to be able to map the informal requirements of business to it did not go away, and likely will not for quite some time.
At that point then, I imagine absolutely everything will have already been automated. Lawyers would certainly no longer be needed, nor would much of the judicial system. Music would be automated, hell maybe even animated movies. We might have celebrity neural nets that produce the best music/paintings/memes.
I definitely agree that coding will look different after AI has permeated most of our information systems, but fully automated coding by 2040? Really? I honestly only see mundane boilerplate type of coding tasks being automated -- though arguably that is probably what a large majority of software developers are currently doing.
I talk to people today who think that mechanical design is basically trivial because we have CAD. All these tools do is help us do much more than we could before. But at some point humans get to decide what we want to use our tools for.
We’ll just have much more technological power in the future.
Surely this is because the systems are more complex that you would have no hope designing using pens and pencils, even with the largest team in the world.
I’ve seen things like that. They’re awesome and I look forward to more cad programs having features like this. But it still takes a human to define the system requirements - we just get more powerful tools.
Before that happens the area's of code review and auditing would be most likely to happen as without that, we will just have AI that can produce buggy code much faster than before.
After all, if AI could code like a human, even the best of us make mistakes and without that avenue of teaching them to be better, by auditing and reviewing existing code, we will only create more problems than we solve.
For many others, intelligent compiler reports would be their first wish.
The paper's abstract contradicts the article title. The paper argues that AI will "allow humans to cope with the difficulty of programming different devices efficiently and easily." This is much more believable than the assertion that coders will be replaced.
> The academics are certain that there will therefore be a shift from human coders to AI-driven coders by 2040.
On the ML front, little appears to be fundamentally new in the last two decades. Not sure how the ML classification craze is going to metastasize into full AI in the next two without some major new developments. ‘Academic certainty’ isn’t going to do it.
The final paragraph gives a clue about how serious the authors themselves seem to take their claims:
Machines writing code under human direction will only further improve our ability to explore the universe, enjoy life, and stream Netflix, especially if it saves us the trouble of learning how to make extremely heterogeneous systems work together.
You go ahead and stream Netflix, I think I'll write some more code.
Programming "entities" argue about golang not having generics. Initiate global thermonuclear war thanks to security bug found in Rust library. AI finds .net of no interest. World saved by humans who fired up an ancient version of Delphi and wrote a worm.
Why is this flagged? Oakridge is a reputable national lab - just because their research is unpalatable, doesn't make it irrelevant. HN - denial is NOT the answer
42 comments
[ 0.26 ms ] story [ 98.4 ms ] threadhttps://arxiv.org/pdf/1712.00676.pdf
At least this kind of academia wont be replaced because their work is worthless, thus we have no interest into automating it.
If you discovered something amazing that takes one page to describe, why write a book about it? Even if it took 80 people in the team.
PS. exactly this kind of work (crappy paper) has been automated: http://thatsmathematics.com/mathgen/ ;-)
Seriously though, if this happens, the repercussions might be far greater than coders having been replaced by (I'd assume self-improving) AI.
I'm not just being coy: as long as there are computers generating code, there will be a need for people to tell those computers what to generate. Those people will be coders, whatever they are called at the time.
The key ability of a good developer is finding out and understanding what needs to be done, the translation form the language of business to the language of technology. Technology has moved from low-level register juggling of 1950s to modern high-level, more and more declarative languages and frameworks. But the need to be able to map the informal requirements of business to it did not go away, and likely will not for quite some time.
I definitely agree that coding will look different after AI has permeated most of our information systems, but fully automated coding by 2040? Really? I honestly only see mundane boilerplate type of coding tasks being automated -- though arguably that is probably what a large majority of software developers are currently doing.
We’ll just have much more technological power in the future.
As the capacity increases so does the complexity.
After all, if AI could code like a human, even the best of us make mistakes and without that avenue of teaching them to be better, by auditing and reviewing existing code, we will only create more problems than we solve.
For many others, intelligent compiler reports would be their first wish.
On the ML front, little appears to be fundamentally new in the last two decades. Not sure how the ML classification craze is going to metastasize into full AI in the next two without some major new developments. ‘Academic certainty’ isn’t going to do it.
Machines writing code under human direction will only further improve our ability to explore the universe, enjoy life, and stream Netflix, especially if it saves us the trouble of learning how to make extremely heterogeneous systems work together.
You go ahead and stream Netflix, I think I'll write some more code.
Plus, that human direction is essentially what programmers do now.
Get real. This is just futurology. Knowingly deceptive article trying to set what people talk about.