[sltEvas]

ML will become a library. It has about as much to do with programming as a compiler. You don't need to know what it does, you just need to know how to make it do things. The problem with ML currently is that nobody really knows how to do things and that you have a million parameters that need tuning and most algorithms need continuous improvement and fine tuning to the use case. There is nothing "mainstream" about ML at this point, except that everyone wants to use it.

In maybe a decade, it might be found in standard libraries of programming languages and on top of things like `Math.abs`, we will have `ML.textToSpeech("Hello world")`, or `ML.isCat(image)`, etc. However, the problem I see with that is that no matter how far we wind the clock forward, we will only be able to put the most simplistic use cases into a library. `ML.isCat()` could be one of those, since most humans will be able to image categorization, it stands to reason that you could put this into a library. However, most industry application involved highly customized ML algorithms that are optimized for a very specific use-case. So there will always be a need for a research team in big companies at least. Maybe smaller companies will try to build their stuff by chaining libraries together.

[virgilp]

There's never going to be a `ML.isCat(image)`, just like there isn't a `Math.solveProblem(hypothesis)`. Yes you do have `Math.abs` and you're going to have stuff `model.fit()` and `layers.dense()` - but something like `ML.isCat` is too specific to be used in a library

[sillysaurusx]

Disagree. In the future, that'll be `npm install ml-cat` followed by `MLCat = require('ml-cat'); MLCat.isCat(image)`

It might not be npm, but something like that is probably inevitable.

The reason it seems so unlikely is because the tooling isn't there yet. No one even agrees how ML code should look, let alone how libs should be distributed to end users. But I saw the transformation for JS in 2008.

[virgilp]

the range of problems you can solve with ML/AI is simply too wide for there to be fully-canned solutions for everything. Sure, there will be canned solutions for _some_ things - maybe even for cat detection, because it's fun so why not.

But, a library that uses AI to optimize the production of your business' flux capacitors? Ain't gonna happen, you need to build that yourself. To have a library/product that solves problems using AI, you need a "language" to describe the problem (like you can e.g. use SQL to describe any data query you may have). But describing problems is notoriously hard - accurately & precisely describing the problem is very often just as hard as solving it.

[sillysaurusx]

Mm, it's a bit like arguing that "the range of text editor customization is simply too wide for there to be fully-canned solutions for everything." Meanwhile, elisp wiki go brr.

I think ML solutions will increasingly take the form of an elisp script rather than a python library, but it'll take a little while to get there.

[virgilp]

> it's a bit like arguing that "the range of text editor customization is simply too wide for there to be fully-canned solutions for everything."

But the the range of editor customization really isn't that wide. That's exactly what I'm arguing, that ML/AI is more like "math" than like "editor customization".

[csande17]

Apple begs to differ: https://developer.apple.com/documentation/vision/vnanimalide...

[fhaodhdms]

Fwiw macs have had an equivalent functionality for both text to speech and speech to text for at least 17 years to my memory. The quality is poor compared to today's server-driven approaches, of course, but the functionality has been there if you're willing to articulate yourself clearly.