Hacker News Reader

Comment by epicureanideal | Hacker News Reader

Right. Most people simply are not in the habit of expressing themselves with enough precision, and might even be missing some essential training.

Whenever precision is required, we end up with methodologies like technical writing, or specialized language like legalese, to provide the necessary precision.

If we didn't have "programming languages", we'd end up with some sort of "computerese" analogous to legalese, and it would still take years to become proficient in using it, and there would still be either people or AI who would translate natural language into "computerese" (but we'd likely prefer human oversight over the AI anyway, given that a human that needs an AI to convert their language into something more precise is likely not capable of judging whether the output is precisely what they meant).

View on HN

> Most people simply are not in the habit of expressing themselves with enough precision, and might even be missing some essential training.

I agree with this, I don't think we'll be programming in natural languages until the computer is intelligent enough to pick out unintended ambiguity and ask clarifying questions. What I'm unsure about is whether or not the computer will be that intelligent in 50 years. Whenever it is that intelligent, I expect us to transition quickly.

A problem is: How many clarifying questions do we want the computer to ask? How many much judgement do we want the computer to exercise?

There are going to be many tasks for which we don't want any independent judgement to be exercised by the machine. And for those tasks, we're going to still want a highly structured language.

And if the computer needs to ask the person a bunch of clarifying questions because they're providing ambiguous instructions, do we really trust that person's judgement about handling the unanticipated edge cases?

That's been a large part of my job at times - working with product stakeholders who give a high level requirements overview and then trying to figure out all the edge cases they didn't specify up front, so that my team could properly design and estimate the work to build the features.

So: yes, most companies already don't require people designing things to be able to fully spec every edge case in advance, without any dialogue.

Thats different. You're the one designing the system -- Exercising judgement about which mechanical rules to implement.

The question is: can your job be automated in 50 years? Maybe it can. But the computer has a long way to go. I don't think we have enough evidence to give any estimate of when (if) that will happen.

Ah, the good ol' Star Trek TNG computer. Even in the show people were constantly annoyed with it.

Exactly! I wish I could communicate with my wife in code, she seems extremely able to find ambiguous natural language expressions that tip me the wrong way with respect to her intent. I still love her regardless.

I think we want a precise computer language, but with a code completion/suggestion good enough to make transitioning from natural language trivial.

Agree. Most of program languages are context free. Human language is mostly context dependent. The auto completion and auto suggestion are the tools to close the gaps of the user experience.

Moreover, human communication is continuous and conversational. Programming is not. Most of existing code editors are not designed to have a conversation. Jupyter notebook is close but not there yet. I bet with a conversational agent style code editor plus a good auto suggestion and auto completion feature, we don't need to invent technology to use natural language to communicate with a machine. We just need a well designed formal language with precise and concise syntax.

Some people write news conforming to Google translation's performance just to make sure the story can be auto translated to many other languages. Most of these stories have pretty normalized vocabulary. This is how much human can adapt to the new world.

One day when the machine intelligence surpasses human intelligence, we will all speak in Python or whatever the most popular among machines.

There's the opposite problem too! If you over-specify, you may be denying the machine a chance to optimise.

I think that's what he's getting at using the word requirements in this quote - as in, you're not specifying the process, only the outcome. By definition it's not possible to over-specify the outcome, it just is what it is, and there is no theoretical reason that should limit the ability of the machine to optimise getting to that outcome.

In a sense, that's one definition of language power - to what extent can you specify that what without having to specify the how, assuming that the machine can execute the how much more optimally than you can most or all of the time.

> By definition it's not possible to over-specify the outcome, it just is what it is, and there is no theoretical reason that should limit the ability of the machine to optimise getting to that outcome.

I'm not sure what you mean by the outcome, but you can absolutely over-specify the output you want, and it's a relatively common problem. For example, you could say 'I want a system that prints lines with a saffron-based pigment', when all you wanted was an intense yellow-ish color, and don't care about the exact pigment being used. Or, you could say 'I want a program that let's a user choose red or blue or yellow and then paints a line in that color', when you actually mean the user can choose any color.

I've seen many problems of over-specification when talking high-level details with less technical people. They tend to give details to help explain what they mean, without caring about the specific details; or they use metonimy expecting everyone understands what they mean.

Ah yes you're right, I completely missed that and that's true. I was thinking about it more from the point of view of not mixing specifying output with implicit instructions on how to reach it (for example, don't say you want to order an array and then pick the last number, just say you want to pick the greatest number from the array) - but that's only assuming you actually are precise about what outcome you want and don't ask for anything irrelevant, which of course people obviously will do from time to time.

One question might be, in a natural language "conversation" with the database system, could one iteratively refine the results to the needed precision, and then perhaps bookmark the query. It's still an open question if people end up with enough precision, but I think that's an easier and different standard than equivocating a single SQL query to a single natural language query.

Reply on news.ycombinator.com