3 ms·

I would bet a small fortune that the ML hype will not live up to expectation on a 7 year horizon. It is true that there are cool things coming out of the field,

by tw1010 9y ago

I would bet a small fortune that the ML hype will not live up to expectation on a 7 year horizon. It is true that there are cool things coming out of the field, and it is true that there are a lot of smart people are entering the field. But the demand for researchers, using actual mathematics day-to-day, is going to be incredibly slim compared to what you'd think it'll be if you read the news about this stuff. Data science has been hyped up for close to 10 years now. And all that time were there fears that we'd all have to become statisticians in a not too far away future. But the day has not come. In fact, you could argue that there is even less reason to get dirty with pen and paper than ever, since a huge mountain of frameworks have been developed (e.g. tensorflow) precisely with the motivation that it'll be able to bring in math-afraid coders into the ML sphere.

dnautics 9y ago
I have a lot of sympathy for this view, but I think you should be a bit more specific about whose "expectation". ML has exceeded my "expectations" for the 5 year horizon now, and I'm expecting saturation of the class of current-ML-solvable problems soon; but that (barring some sort of paradigm shift that I can't see at the moment) the siloed nature of ML training will mean that actual applications will still create a strong demand for deployment solutions for about 5-7 years to go.
agitator 9y ago
Give it a year or two and you will see frameworks where the complex mathematics behind ML will be more abstracted away, the model architectures used for various applications will converge on their best uses and any software engineer will be able to train a model to do what they want and apply it to their product. There's a lot of hype right now, but you probably won't need the in-depth ML knowledge to get create products in the coming future. Right now there is still a lot of research (trial and error) on fiddling with architectures to get the best results. Personally I like to build products, and struggling with messy training code, running tests, and comparing results aren't my cup of tea.
muzani 9y ago
People have said this about apps and it's still going. I think eventually ML will become a grunt job like Android, PHP and so on.

by tw1010 9y ago

dnautics 9y ago

I have a lot of sympathy for this view, but I think you should be a bit more specific about whose "expectation". ML has exceeded my "expectations" for the 5 year horizon now, and I'm expecting saturation of the class of current-ML-solvable problems soon; but that (barring some sort of paradigm shift that I can't see at the moment) the siloed nature of ML training will mean that actual applications will still create a strong demand for deployment solutions for about 5-7 years to go.

agitator 9y ago

Give it a year or two and you will see frameworks where the complex mathematics behind ML will be more abstracted away, the model architectures used for various applications will converge on their best uses and any software engineer will be able to train a model to do what they want and apply it to their product. There's a lot of hype right now, but you probably won't need the in-depth ML knowledge to get create products in the coming future. Right now there is still a lot of research (trial and error) on fiddling with architectures to get the best results. Personally I like to build products, and struggling with messy training code, running tests, and comparing results aren't my cup of tea.

muzani 9y ago

People have said this about apps and it's still going. I think eventually ML will become a grunt job like Android, PHP and so on.