I don't understand what this word (AI) means anymore.
Everyone seems to be doing it, except me. It has something to do with "machine learning" and "big data" and it's done using neural nets, which have been around since 1940s.
Somehow I feel like you can use the term AI to any algorithm that takes some input data, does some processing and produces output data. After all, that's what it is.
Because this term is so ambiguous, I will try and translate it into more boring terms:
AI which builds AI means algorithms which produce other algorithms ?
In other words, a program which can write other programs, right ?
I think the term AI has been reduced to "anything you can make a computer do, without writing the explicit instructions for it". So what would have been described as a Bayesian process (or Bayesian algorithm) ten years ago, is now called AI. Every tool produced with some genetic programming can now be classified as AI, even if the resulting program is completely static after deployment.
It's just hype. Just like every dynamic javascript component got rebranded as Ajax ten years ago.
Funny thing is that this trend is the exact opposite of what has been the case in the past decades: things stopped being called AI as soon as they became broadly useful!
In my view, that's because scientists and programmers alike took AI seriously. No one was willing to claim "AI" as it was a holy grail, considered as-yet unattainable. What has changed since then is not really our capability, but our willingness to dilute the term for marketing gain.
'artificial intelligence' seems a reasonable descriptive term for such things to me. I don't think it's inaccurate.
It's artificial, so it doesn't have to work in the same way as human intelligence, and it doesn't carry any inherent claim about the level of intelligence, so ought to be able to cover stuff that's no way near the power of the best of human intelligence.
The word "AI" has always had a vague meaning. Ever since the 50's when people called simple logic systems AI. Everything from chatbots to video game bots have been called AI.
The problem is that "AI is whatever computers can't do yet". Hard looking problems in computer science all get called "AI". So once they are solved, they continue using the name. But solved problems obviously can't be "intelligent".
The word you are looking for is "AGI". But there is no such thing as AGI yet, so no real projects or research can honestly use the word.
Anyway more accurate terminology would be "machine learning". That's what word the researchers themselves use. Only media calls it "AI" so much.
In this case it would be machine learning models that set the parameters for other machine learning models. There's nothing in the article about "AIs building AIs", although it does automate some of the work of machine learning practitioners. I've always said that machine learning is one of the fields in a unique position to automate itself away.
> and it's done using neural nets, which have been around since 1940s.
Except now people know how to train multilayered ones faster, also have different topologies (like LSTM networks)
> to any algorithm that takes some input data, does some processing and produces output data. After all, that's what it is
Yes, and hell is just a hot place.
Machine Learning is where you can infer something (usually classify or extrapolate) about future data based on present data/examples and this inference has had no human input determining it.
> In other words, a program which can write other programs, right ?
Not really, it's not writing it from zero, more like tuning it
But calling machine learning algorithms just algorithms is not helpful at all. I had to work through a few books and courses to get a decent understanding of ML despite the fact that I had a CS degree and a good understanding of discrete algorithms. In short, the things that are put on top of the technology stack do not reduce to the layers below in a useful way.
i would say AI is today synonmous with algorithm that are automatically tuned given new data, and don't require changing their source code to perform better.
A lot of these idea came out of Harvard HIPS lab under Ryan Adams. Their work is open source at https://github.com/HIPS, specifically Spearmint and Autograd. His company, Whetlab, was mentioned in article as acquired by Twitter. Facebook is taking the idea further as are all the major ML players.
Is flow just another handy library for constructing DNN graphs? (like tensorflow) or is it actually a program that is creating DNN graphs intelligently?
I suppose the latter makes a lot of sense, actually. Some of the tweaking and retesting I do to my tf models seems like it could be automated.
It seems like the real challenge of AI isn't so much constructing the models but feeding it training data.
Facebook started out as a non-algorithmic company, AOL – geocities – friendster – myspace – Facebook – who's next? They have a lot of catch up to do compared to Google, which started as a core algorithmic company in a branch of AI called Search.
Everyone seems to be doing it, except me. It has something to do with "machine learning" and "big data" and it's done using neural nets, which have been around since 1940s.
Somehow I feel like you can use the term AI to any algorithm that takes some input data, does some processing and produces output data. After all, that's what it is.
Because this term is so ambiguous, I will try and translate it into more boring terms:
AI which builds AI means algorithms which produce other algorithms ?
In other words, a program which can write other programs, right ?