Hacker News Reader

Comment by silvio | Hacker News Reader

With more research along these lines, we may have a day when a programming job interview will consist on interacting with some codebase for 10 minutes, and the model will spit out our normalized scores on syntax comprehension, algorithmic thinking, library familiarity, etc.

View on HN

I wonder if when we get that far the computer is able to write the code for us anyways and programmers will be redundancies. Its just a matter of if, not when. Understanding human cognition well enough to evaluate it is a good indicator that the singularity is near.

Computers already write code for us, they just need to be told what we want them to write. The question is if we can design a programming environment where the computer can understand 'natural' human specifications and compile that, instead of needing to have a trained human compile the natural specifications into source code, which the computer can compile into a program.

Sure we could. This is just like Siri, you tell it what you want, she asks for clarifications, a conversation you have with the computer until...bam, you get what you want.

I would think we could do much of this already if we tried, but there are a lot of things on my list to do until I get to that.

Your optimism on this seems unfounded in my opinion.

When there are no domain restrictions, computers have not proven very capable of comprehension. Look even at the comparatively simple domain of handwriting recognition.

More importantly, you entirely miss the problem that many times the human doesn't know exactly what he wants until he sees it.

As a software engineer, I have no fear my job will be replaced by computers talking to product managers.

That's why the program should be expressed as a conversation! We don't know what we want so we should start vague, get feedback, provide clarifications, and so on. The conversation with the computer must be a two-way thing!

That's brilliant. Not in terms of some AI working out what a silver haired old grandmother meant but in terms of talking to clients - they could imagine talking to a computer and I am just providing that feedback.

Daily or weekly, feedback nice and quick

The thing is, we probably don't need hard AI yet to do some of this. Yes, it must be a dialogue system, but we have those today. We already saw the movement in this direction with Hugo Liu's concept.net work, but for some reason no one has followed up yet. We are getting to the point with speech rec/understanding technology that someone is bound to try again soon.

In general, we don't need Hard AI for much anything, and usually don't want it. Any specific problem you want solved will require some specific solution method or algorithm rather than the whole and entire "make a computer hold an adequate water-cooler conversation" thing.

I think it's less worthwhile to argue this than to try to build it.

That's kind of how TeX and METAFONT have worked since the 1970s. But maybe you have something else in mind.

The problem is that most people are simply incapable of giving correct specifications.

In a lecture about this topic (how to create specifications that can be turned into formal correct code) we were given the following simple example of an incorrect specification:

"Everybody loves my baby. But baby loves nobody, but me."

If you formalize this, you can simply conclude that the person that says that is equal to the baby - which clearly is not, what you intended.

And this is a very, very simple specification. Specifications of real software are magnitudes more complicated.

That is why the computer must hold a two way conversation with the programmer. How often do we just type out a program from our head to the computer anyways? Most of us write the program, debug it, change it because the result wasn't right, or we didn't really understand what we wanted, or someone else decided the requirements changed, or whatever...

A large chunk of the job of developing software is getting the stakeholders to understand the problem they want the software to solve. The rest of it is just typing which is the trivial part.

Yeah, right. And all the technical books on design patterns, functional programming, algorithms, etc. are out there just to teach programmers how to type faster ...

That was the hype surrounding Fortran when it came out in 1959: it would eliminate the need for programmers, because you could just punch the formulas and it would translate them into a program automatically. And in a sense it did, and its successors did to an even greater extent. Today you have "nonprogrammers" who can build, in a day, in OpenOffice Calc, computations that would have taken weeks to get written by programmers in the 1950s (when they could have been done at all.)

It's partly a matter of giving the objectives to the computer at progressively higher levels, and partly a matter of improving user interfaces so that you can tell what the thing is doing and what it's going to do.

These predictions always fail to account for C++.

That is, having computers write code depends on the code being in a form about which computers can formally reason. If you can't prove anything about your code because the language is that bad, good luck getting a computer to write code.

/explain_the_joke.jpg

has there been any work on this? it seems relatively straightforward to score a git branch in certain time scales

EDIT should have RTFA before commenting, gaze tracking stuff is very cool

It does sound representative of the bulk of the daily work.

Reply on news.ycombinator.com