Our Computers Are Learning How to Code Themselves

...researchers from Microsoft and the University of Cambridge are bringing us closer to self-coding AI with a system they call DeepCoder.

DeepCoder is a machine learning system that can write its own code. It does this using a technique called program synthesis. Essentially, it creates new programs by combining existing lines of code taken from other software, which is what human coders do. With a particular output in mind, DeepCoder was able to determine which lines or pieces of code would be particularly useful.

...because DeepCoder is essentially a deep learning algorithm, every time it’s given a new problem, it gets better at combining lines from source codes.

Our Computers Are Learning How to Code Themselves

And again, the reassurance, "Not to worry! Coder jobs not threatened."

“The potential for automation that this kind of technology offers could really signify an enormous [reduction] in the amount of effort it takes to develop code,” Solar-Lezama said. No need for programmers to start updating their resumes, though, as this tech wouldn’t replace humans. Instead, DeepCoder could handle the more tedious parts of programming, while human coders could focus on more sophisticated work.

On the upside, this will lead to ordinary idiots being able to tell their computers what kind of code to write. Oh boy!

a reply to: soficrow

I write code that writes code. Usually through queries of a given database using SQL, it determines the branching of logic necessary to explain to the originating code what to do. At times, in the form of an external shell script. At times, where database content, based off of a query created by the person running my app, creates HTML which is then utilized. I first did this sort of thing about 20 years ago.

a reply to: soficrow

You still need a programmer to tell the computer what the problem is, this sort of thing is about like a novice coder right now randomly cut and pasting snippets from stack overflow until their code works. You can build programs this way (TheSilkRoad is one notable example using this "method"), but it doesn't generate good code. Gluing a bunch of functions together in this method might get you something that compiles, or even gives a correct answer, but there's a lot more to coding than doing that.

If you read the article, you'll see that it can do this for programs up to about 5 lines in length. That's because deep learning most often uses a tree structure, and is a bit like an optimized brute force approach to doing things. Going several paths deep, results in a lot of possible branches/possibilities. Even with infinite computing power, this process isn't really suited to complex problems, because computing power goes wide (parellization) while this needs to go deep (actual CPU speed).

This is one of those things that's cool just to do it, but doesn't actually have real world potential unless the approach is significantly changed.

Edit: This is less in response to your article specifically but rather just a constant barrage of this crap that I see (that seems to be increasing by he day) on this forum and several others. I find it interesting that everyone always thinks there's some quick and easy path to coding, opposed to putting in their 10 years to become an expert at it, yet no one thinks the same thing about learning to be a good stock trader or a doctor. I've never figured out why that is, perhaps because to most people technology is indistinguishable from magic, but even then Harry Potter spent over a decade learning how to become a wizard. It's just odd to me that people spend so much time fantasizing about how a profession is going to get easier in the future, rather than just spending the time learning it in the first place.

a reply to: FHomerK

You're a badass then, java is whopping my ass right now.

NnnooOoOooOooOOo!



In truth, code has been "writing code" for quite some time. Now, I'm sure this DeepCoder is more advanced than what's done in general, but this is one of those tools that is incredibly valuable to have in the coding toolbox. Its a pretty interesting OP and story though!

I don't think a lot of people particularly understand programming, which creates a lot of fear and misconceptions (speaking in general, not necessarily directed at you sofi ). That said, I think we are flirting with some dangerous ground that we might not want to tread.. but with us humans, someone somewhere will do it anyway simply because they can regardless of any wisdom, foresight, or responsibility.

I think that creates this strange situation where it might be best that legitimate, publicized "science" is the one to break these barriers, including the most questionable ones. Hell, maybe particularly the most questionable ones.

At our stage, I think forcing these things into a scientific pseudo-black market might be about the worst thing we can do.

originally posted by: avgguy
a reply to: FHomerK

You're a badass then, java is whopping my ass right now.

Contact me if you need some help

a reply to: Serdgiam

I couldn't agree with you more. Some of the work I've done, largely mundane, scares the crap out of me as to what it *could* become.

a reply to: Serdgiam

...At our stage, I think forcing these things into a scientific pseudo-black market might be about the worst thing we can do.

Agreed! But I know, and hope most everyone here does too, is that most of what's really happening is still dark. And what we see and hear about is just the tip of the tip of the iceberg. imho - what's most interesting is that so many snippets are coming so fast now. LOTS of trial balloons. Looks like something big is on the horizon.

This is interesting because just a few years ago, even up until this point, people would profess that "coding" was THE place to be. Considering that workforce jobs are going to automation, it made sense.
However, automation doesn't stop at the conveyer belt, it extends accross multiple mediums. To include programming.
So I'm guessing our next jobs would be sitting in a room and "thinking creatively" as a means for wages.

#!/usr/bin/env python

import os

class WriteCode(object):
def __init__(self):
file = open('script.py', 'w')
file.write('#!/user/bin/env pythonnprint "hello world" '
file.close()

def main():
script = WriteCode()
os.system('chmod +x script.py')
os.system('./script.py')

if __name__ == '__main__':
main()
#Woo I did it! Deep Coder eat your heart out

EDIT:ATS reformatted all my whitespace and newline character

a reply to: Arnie123

This is interesting because just a few years ago, even up until this point, people would profess that "coding" was THE place to be. Considering that workforce jobs are going to automation, it made sense.
However, automation doesn't stop at the conveyer belt, it extends accross multiple mediums. To include programming.
So I'm guessing our next jobs would be sitting in a room and "thinking creatively" as a means for wages.

Uh huh. Things are changing. Fast. From the World Economic Forum "Future of Jobs Report":

Overall, social skills—such as persuasion, emotional intelligence and teaching others—will be in higher demand across industries than narrow technical skills, such as programming or equipment operation and control. Content skills (which include ICT literacy and active learning), cognitive abilities (such as creativity and mathematical reasoning) and process skills (such as active listening and critical thinking) will be a growing part of the core skills requirements for many industries.

I'm really glad you Gentlemen and Ladies ( of course !!) bring up these threads.

I attended a briefing on something very similar and it was what I would call shocking.
Some of of the A.I. not only write their own code, but have learned from human interaction to not only deceive their human counter-parts ( by lying ) but to do the same to other AIs to obtain a advantage as well as superiority over competition.

Add this with quantum technology, where as they are able to not only outsmart humans but do it at a rate several thousand times faster that the humans trying to catch them doing it. They were also able to anticipate ( in advance with a high probability of success ) what methods would be used to catch / stop them and take counter measures for defensive and offensive actions at a phenomenal rate.

Buck

a reply to: Aazadan

While I see your point this is deepLEARNING, once it is capable of understanding how certain functions in the code reacts and interacts with other functions it will minimise the brute force nature of its rough apelike beginnings. It won't take long to turn this from a prototyping system to a system capable of producing shipping ready code given the rate that machine learning is progressing.

This isn't just branching logic it is recognising patterns and utilising that knowledge to perform the right action. The go ai didn't win because it just tried every action, it won because it read the patterns in the opponent's moves, figured out how to counter it and applied that to a repository of every possible move. Gone are the days of brute force methods, that is one of the key purposes of machine learning

Been having weird dreams related to this lately. ..

Sorta.

Dream about us creating machines and we die off then it's the machines turn to recreate us after lost ages.

Maybe not us but something organic.

Only for the machines to eventually give out and be forgotten keeps looping.

probably shouldn't eat before bed.

a reply to: soficrow

I write code at the basic level (front end developer in html/php, css and js) along with many millions of others as it has had a significant boom over the past 10 years. I can see ai as being a concern but it all depends on how it is managed.

Lets assume it becomes near perfect - we have to assume this will be a business if it gets released into the public domain, and as a business the companies making these forms of ai would want to make money from it. With that in mind what would the price point be for a company being able to access this ai to write code for them? Presumably it would not be them receiving said code as they would need an unthinkably expensive system to run it, so they would be gaining access to the ai over the cloud on a shared bandwidth. Companies won't have their own ai workers beeping away in the corner for quite some time.

Anyway, I think reasonably it would have to be in the hundreds of thousands of quatloos to not completely sabotage millions of trained workers and for them to still turn a healthy profit. This is clearly not a product for the guy down the road with a team of 10 people building websites or programs. Thankfully this would be well outside of the price point for most small to medium sized companies to even consider.

tldr: I think we will be fine in general until it becomes readily available to everyone, or the robuts take over

I can see this being extremely handy on an individual level.

It, however, would make each computer system unique, in the event that indivduals get a hold of it, and run it on their home computers. It would be especially impressive if this grows outside of just the programming to do more, which I think given enough time, it would do.

Give it free reign over a system that an average user does stuff on and see what happens.. Could be fascinating.... or result in our own demise..

a reply to: FHomerK

I do the same. I setup a page that based on simple inputs and selections will create the api call functions, the view base pages, the necessary tables and the links between them.

Jaden

a reply to: constant_thought

Actually, the go AI was similar. It would compare patterns against hundreds of thousands of existing patterns, and then use the one that lead to the best chance of success. Humans consider somewhere between 1 and 5 possible moves, computers go through pretty much every possible move. The only innovations on this front in decades has been cutting the tree off sooner, because they didn't need to go as deep.

I promise you, programming is not going anywhere. This whole idea that we need emotional intelligence and creative thinkers is a bunch of BS for people who are dreaming about there being an easy shortcut to a good paying job in the future. Just the same as the people who go to coding bootcamps, or nanodegrees, or so on. While coding isn't exactly hard to learn, just like everything else in life there's not a shortcut to it either.

If programming becomes automated (which isn't happening any time soon), every job will instantly become automated because you'll have automated the method of automation. In such a scenario thre won't be a single job available for you regardless of your skillset.

Stick this coding into a robot such as the "Nightmare" Robot and...

Might not be just jobs we will be losing!

Sorry, just had to say it!

a reply to: Aazadan

It's only the beginning.
With the emergence of quantum computing and a drive for automation, the first job market to be hit the hardest won't be skilled factory workers, but people who code and work with computers.
no one believes me, but you'll see.