Hacker News Reader

Comment by bitL | Hacker News Reader

Trends are: finance is being eaten by software, software is being eaten by machine learning. The question should be "do I study machine learning?".

View on HN

Software eats finance because it’s making traditional finance jobs obsolete.

ML does not eat software in the same sense, because ML and software engineers generally do non-overlapping work. ML involves analysis, software involves synthesis.

There may be a day when ML automates software development, but it’s nowhere close yet.

At most, ML is merely dominating current hype among software-adjacent technologies (along with block chain).

> because ML and software engineers generally do non-overlapping work

not in the mainstream right now. Check back in 10 years.

It'll be either ML or nuclear winter. Heck, maybe both.

software is being eaten by machine learning

Could you elaborate on this one?

Only genius-level people have any bright future in software these days; when you see what is being cooked in the labs of top companies, in 10 years software developers that missed the train on machine learning will be fighting for scraps in low-income jobs.

I do have that FUD, but then I check reality (jobs) and people are getting paid to do Java, C#, Cobol.. I think we have a few more decades! But you are right.

It's not a popular opinion around here, but it's coming whether we like it or not. Better be ready...

What do you mean by that? That machine learning will be able to distill vague, shifting requirements from non-technical businesspeople into functional software within 10 or so years? Or that all software development jobs will involve machine learning in some way?

Both to some extent. There will be much higher floor for software part that is not yet automatable by ML, requiring more skilled developers to do anything. And that development involves ML is happening already and in some areas is in more advanced stage.

What leads you to believe that in 10 years, we will have developed general artificial intelligence? Current machine learning algorithms are very good at deriving heuristics to solve narrow problems with well-defined, testable goal-functions - e.g. image classification, stereo cost calculation, optical character recognition - but code generation from business requirements is far too broad.

If you have any links to papers demonstrating anything close to what you are stating, feel free to post them.

Who said you need AGI to automate away a lot of what low-level software engineers are doing (majority in fact)? Automatable level will rise in time, moving more and more software engineers out of business if they don't adapt (because regular folks will be able to achieve with a piece of ML software what software engineers had to do for them previously). Advances might move in a form of a cliff, like what might happen to taxi drivers once self-driving cars are good enough and affordable. Not sure why do you think software engineering is immune to that.

How does ML help in that regard? I don't understand exactly what you're saying by "ML software". How does the regular user specify what he wants? By writing it in plain English with all its ambiguities, turning the "ML software" into a very fuzzy compiler that constantly tries to guess your intentions? Through some GUI or visual programming language, like we've had for decades but which have yet to cause a shortage of programming jobs?

We've been automating programming since the very beginning, moving from machine language to assembly to high-level languages to using libraries and frameworks (although making different libraries cooperate is sometimes just as labor-intensive as writing the functionality from scratch).

I've seen comments like his and I really don't know if I'm just ignorant or if the other guy is talking out of his ass. I see no way that machine learning as it currently exists could possibly perform the work of coders or any other human function that involves ambiguity and shifting requirements.

Depends on the level of standardization of "building blocks" in a given industry where software is needed. Most software engineers nowadays spend time mimicking what others posted on Stack Overflow and just gluing together different modules with not much thinking. There is a lot of opportunity to unleash ML there. One example - imagine you need to convert data from one supplier to another one's format. ML can identify which data columns looks similar and propose a conversion between schemes to an engineer, engineer just needs to approve it and make small changes if necessary. Algorithm then can learn from the adjustments and make them automatically in the future. As a consequence, you might not need to employ 3 people doing that, but only 1. So you won't get rid of humans completely, but boosting up one's capabilities by augmenting intelligence allows you to save cost and time. Next big thing is identifying which of these processes can be standardized and automated away in development. The pressure coming to rank-and-file developer will be substantial, given internal prototypes I've seen already.

It's coming, but ten years is unreasonably optimistic.

I firmly believe this. If you arent brilliant, you arent writing real software, you are copying what other people are already doing. Any monkey can do that.

Reply on news.ycombinator.com