DeepMind says its new AI coding engine is as good as an average human programmer(theverge.com)
theverge.com
DeepMind says its new AI coding engine is as good as an average human programmer
https://www.theverge.com/2022/2/2/22914085/alphacode-ai-coding-program-automatic-deepmind-codeforce
21 comments
Following that, the question is, will AI improve enough to pay off its own tech debt? If it can't, everyone will have to hire 10x the number of developers.
I will say this...
If there's an AI that when given inputs can create a super efficient program that calculates the right outputs, it would be an AMAZING tool in programming. The few times I need to write such functions, a ridiculous time saving would occur and let me/other engineers focus on the test cases.
Will such a program actually solve the need to have engineers? No.
Will such a program actually assist engineers with things that a computer can actually be really good for? Absolutely. 90% of the time a solution is good enough. The problem is when we need to optimize, but that's okay, we can have humans take over and figure out if the AI did something that could be improved upon, or if the problem space needs to be changed.
Point is, its a fantastic tool. Yay. Now let's figure out how to make practical usage of it.
If there's an AI that when given inputs can create a super efficient program that calculates the right outputs, it would be an AMAZING tool in programming. The few times I need to write such functions, a ridiculous time saving would occur and let me/other engineers focus on the test cases.
Will such a program actually solve the need to have engineers? No.
Will such a program actually assist engineers with things that a computer can actually be really good for? Absolutely. 90% of the time a solution is good enough. The problem is when we need to optimize, but that's okay, we can have humans take over and figure out if the AI did something that could be improved upon, or if the problem space needs to be changed.
Point is, its a fantastic tool. Yay. Now let's figure out how to make practical usage of it.
GitHub Copilot is already available and it's surprisingly usable. Well, not in a "here's a description of a program and create everything from scratch", but it can already create simple functions by reading the comments or the function's signature.
Phew, glad I am not just average haha.
I've seen many tools that claim to save dev time or make it so you don't need any devs and they do indeed take a lot of the chore out of it, but the remaining 20% is what takes 80% of the time.
I've seen many tools that claim to save dev time or make it so you don't need any devs and they do indeed take a lot of the chore out of it, but the remaining 20% is what takes 80% of the time.
If it codes what people actually ask for, they will be very, very dissatisfied.
That's not saying much...
Blog post: https://deepmind.com/blog/article/Competitive-programming-wi...
Demonstration: https://alphacode.deepmind.com/
This is scary and amazing. What's next? AlphaCode writing itself? Super intelligence?
I'm not talking "scary" in terms of robotic overlords, I'm talking about companies possibly getting rid of programmers and the whole industry getting a lot more difficult to enter.
I can certainly envisage AlphaCode turning interface mockups into actual HTML code or HTML code generators - a ton of work goes into that and it's quite tedious and ungrateful.
Demonstration: https://alphacode.deepmind.com/
This is scary and amazing. What's next? AlphaCode writing itself? Super intelligence?
I'm not talking "scary" in terms of robotic overlords, I'm talking about companies possibly getting rid of programmers and the whole industry getting a lot more difficult to enter.
I can certainly envisage AlphaCode turning interface mockups into actual HTML code or HTML code generators - a ton of work goes into that and it's quite tedious and ungrateful.
A nightmare scenario for _every_ programmer.
It doesn't matter if you think you are doing complex programming that is not [yet] automatable: all the programmers whose job can be automated will be coming for your job.
It doesn't matter if you think you are doing complex programming that is not [yet] automatable: all the programmers whose job can be automated will be coming for your job.
Not at all. Programmers will simply become the people instructing the AI on what to code/program. Just as higher level programming languages "eliminated" or "automated" programming in machine code...
Yeah...Then we get to the debugging part.:-)
Compilers automated 99.9% of work already.
If they can automate away my job I hope they do
The average human programmer isn't very good, so I'm not sure that is the achievement that they should be shooting for.
It's a major milestone. When a computer became as good at chess as an average human player, it was only a step removed from beating Kasparov.
While it's impressive, competing on a well specified problem does not compete with a typical developer. A very small part of an average human programmers job involves well specified problems.
> AlphaCode was tested on 10 of challenges that had been tackled by 5,000 users on the Codeforces site.
> “I was sceptical [sic] because even in simple competitive problems it is often required not only to implement the algorithm, but also (and this is the most difficult part) to invent it.”
So AlphaCode can "invent" a solution after seeing 5,000 other solutions? I'm not clear, is it producing an answer to a unique question? Or is it doing the same work a single solution plus copy and paste could do?
Considering how many of us do CRUD tasks all day, every day, this might be surprisingly effective. The question is, will unique work take the place of those CRUD tasks we've all done 5,000 times?
> “I was sceptical [sic] because even in simple competitive problems it is often required not only to implement the algorithm, but also (and this is the most difficult part) to invent it.”
So AlphaCode can "invent" a solution after seeing 5,000 other solutions? I'm not clear, is it producing an answer to a unique question? Or is it doing the same work a single solution plus copy and paste could do?
Considering how many of us do CRUD tasks all day, every day, this might be surprisingly effective. The question is, will unique work take the place of those CRUD tasks we've all done 5,000 times?
This new role promises to cut costs without losing any developer velocity, in fact it increases velocity (see carefully chosen, unreproducible productivity paper on arxiv). It accomplishes this by allowing an auto-developer to guide an AI to generate code in line with industry best practices. Hype!