Programming won't be automated, or it already has been

That's because the spec is the code. The product is the binary.

Human "specs" are more analogous to statements of intent.

Ever higher levels of abstraction will come about but generating the "spec" will continue to be the programmers job.

A compiler that could ask the right questions when the spec is ambiguous would be able to replace most of what a programmer does.

> Past experiences don't predict the future accurately. Just because we couldn't fly before 1900s didn't predict we never will. We already seen birds doing it but it took a while to reach a breakthrough.

OK, but by this logic literally any outcome is as likely as any other. Not much basis for prognostication there.

> An AGI system should be able to write computer code based on vague specification. If it's intelligent it will write back clarifying questions.

What makes you think this approach is faster than just writing down the specification in your head? Is it because you assume the machine can guess important design decisions that you leave out?

To me, what you describe sounds like what every programmer does: enter code and look at what the compiler says. There's nothing qualitatively different from what you describe and a sufficiently powerful type system.

You have to somehow communicate a specification to a computer; this is programming. Languages might become more powerful, but you always want to be able to point to the place in the specification that implements your important feature, and not have some computer hopefully deduce it.

If you have some extremely important implementation note, would you rather specify it in exact terms via code, or give vague hints about it and depend on it being deduced by some AI compiler? What if you have hundreds of important implementation notes? It all ends up being code.

> Artificial general intelligence (AGI) is not an impossible goal, but we are simply not there. But saying we never will needs a great evidence which we don't have afaik.

"AGI is not an impossible goal" also has no evidence afaik.

Theoretically there isn't any reason true human-like intelligence has to imply human-like consciousness.

written by robots, you think?

As a platform, they are big on encouraging people who wouldn't otherwise write to do so, so you get a lot of articles that are by new writers or who lack a presence elsewhere. This is a good thing in general as we need more diversity in people writing and sharing their experiences. It does mean you need to take this context on board as a reader, though.

That's still a far cry from being able to voice vague requirements and have a computer develop a program from them.

Programming was in its infancy 50 years ago. Things tend to change drastically the first few decades but then the rate of (significant) change slows down as something matures. Take aviation for example. Think of how rapidly aviation technology advanced between the Wright brothers first flight and 50 years later. Now think of how aviation has changed over the last 50 years - sure there's been some advances, but nothing like the changes in the previous 50 or the first 50.

Programming will certainly continue to make advancements, but I wouldn't expect the advancements of the next 50 to rival the magnitude of those of the last 50.

How about python? It is as easy as it gets.

I worked for a software company where the CEO believed he could eliminate the programmers if he could get them to build enough "components" that could be "glued" together. He had a large team of architects and programmers working on this for multiple years. They literally had a GUI that was like a circuit board where you could drag and drop these components and connect them via a wires.

After multiple years they couldn't build a basic clock out of these components, let alone any kind of program. The CEO eventually fired all the programmers on the team because he felt they were undermining his brilliant idea and were refusing to get it done.

Sometimes a good idea! Just probably not what they meant.

We may be getting closer, but it is not exactly guaranteed that our progress will not reach any wall preventing it from going further.

I mean, you aren't really engaging any of the arguments.

Why is that "obvious"? What really made computers revolutionary was the opposite development.

As if they would have a formally defined contract in the first place. We can't even get those for the vast majority of software we write by hand!

Assuming neural nets could solve every problem, who writes the formal contracts and validates the proofs? I would call the person writing the contract the "software engineer" and the prover the "QA engineer", and we're back to the status quo.

I do think machine learning will become more mainstream, like the saying "Google uses Bayesian filtering the way Microsoft uses the if statement".

Even so, if statements tend to be written by humans, compartmentalised into functions/modules/methods/etc. and plugged together to form whole systems. Microsoft software (presumably) isn't one huge `if/elif/elif/...` in a single `while` loop, even though that would suffice in theory.

Likewise, I imagine Google's software isn't one giant formula, conditioned on inputs and sampled for outputs. Although many of their user-facing applications do contain a black box of this form (e.g. for translation, text-to-speech, etc.), they're only used in very narrow, well-defined tasks, and are surrounded by a huge amount of "normal" code.

Your comment is being dismissed (and downvoted) pretty thoroughly; but I think there's a grain of truth to it. As has been said by others, we don't get complete specs as it is, and our industry has moved towards "release often and fix the bugs or mistaken assumptions." The only real difference between the output of a NN and a human is the inference of the context to drive those assumptions.

Seeding a NN would be a simple listing of invariants: Headers are at the top of the screen, footers at the bottom, there exists a name field in the page and the datastore; the name field on the page corresponds to the name field in the datastore. Once those invariants are defined, the intermediate product is released to the customers and they directly drive further development of the NN by simply saying "this is not the expected behavior" in some easy drop-down.

I think the biggest hurdle will be that a human's assumptions are going to be more palatable than a NN's assumptions. But once a number of those assumptions are codified (or a statistical preference is determined), this weakness may be effectively mitigated.

This "crowd-sourced" definition of correctness has already been proven - in chatbots. People have been able to coerce chat bots into providing their desired output without touching a line of code.

Cleanly connecting NNs to contracts / rules is indeed a hard problem, which many people would like to see solved. Please see my post about that [1], and the corresponding HN comment thread [2].

[1] http://blog.foretellix.com/2017/07/06/where-machine-learning...

[2] https://news.ycombinator.com/item?id=14717692