Deep Learning seems good at approximating things or iterating on things in coordination with humans. Or if you are doing things that do not need an explanation for how the magic box came to its results, and does not need guarantees of correctness, it is a great tool.
Thus, it is perfect for human assistance tools, lossy compression, brainstorming, art, games, and so on.
It is also great for things where the answers are verifiable, like protein folding or various other kinds of problems where all that matters is that you find the best-fitting solution and don't care how you arrived at it.
But it seems particularly ill-suited for something like fully self-driving cars because in that case there would be no human feedback, and the decision-making process matters. If you cannot defend in court why the system performed the actions that it did, then it is a non-starter.
Even if you personally are comfortable with the probability that nothing bad will happen, the regulators and courts likely won't be. And if you don't really know why it is making decisions, then it could start making bad ones at any time with almost unlimited severity.
Which is why it is the wrong technology for any life-or-death matter.
Deep learning will chug along just fine. Musk and friends may have been too optimistic, but they weren’t too pessimistic – they knew that deep learning was only the first step in the not so long journey towards real AI.
If you want to get a sense of how long that journey might be, ask yourself: how long did it take from Zuse to Snapchat filters? We will see machine intelligence within our lifetime will defeat humans in all fields.
This does not mean that the AI singularity will arrive tomorrow. But we will soon be in a position to understand what the future will be like, and have a good idea of whether it will be pleasant or unpleasant.
2 comments
[ 3.5 ms ] story [ 20.0 ms ] threadThus, it is perfect for human assistance tools, lossy compression, brainstorming, art, games, and so on.
It is also great for things where the answers are verifiable, like protein folding or various other kinds of problems where all that matters is that you find the best-fitting solution and don't care how you arrived at it.
But it seems particularly ill-suited for something like fully self-driving cars because in that case there would be no human feedback, and the decision-making process matters. If you cannot defend in court why the system performed the actions that it did, then it is a non-starter.
Even if you personally are comfortable with the probability that nothing bad will happen, the regulators and courts likely won't be. And if you don't really know why it is making decisions, then it could start making bad ones at any time with almost unlimited severity.
Which is why it is the wrong technology for any life-or-death matter.
If you want to get a sense of how long that journey might be, ask yourself: how long did it take from Zuse to Snapchat filters? We will see machine intelligence within our lifetime will defeat humans in all fields.
This does not mean that the AI singularity will arrive tomorrow. But we will soon be in a position to understand what the future will be like, and have a good idea of whether it will be pleasant or unpleasant.