The "breakthrough" AI of today is deep learning on massive amounts on data applied to two areas - speech/NLP and vision. How this generalizes to a super intelligence that can take over the planet is so strange. Does a child need to look at billions of images to figure out what a chair or cat is? Will this AI figure out how to select a good business partner?
The problem is that you ask someone who is good in one field (say electric cars or theoretical physics) to opine on something like AI. The correct response is to say that you don't know anything about AI. But, the ego of being an public intellectual prevents that. So, what is the safest option to not seeming dumb - say something like we need to make sure safeguards are place to protect AI from becoming dangerous and kill everyone.
Meanwhile people who actually build these systems know that these systems not are generalizable to variety of tasks (like humans) and they are not intelligent. Best case, they augment humans in their tasks.
>Does a child need to look at billions of images to figure out what a chair or cat is?
Of course! Not exclusively images of cats or chairs, but children have absolutely seen billions of images by the time they start to exhibit discernibly human-level intelligence.
assuming you see an image every 400ms which is given blinking and activation of neural pathways a good approximation.
Billion images per that rate is equivalent to 12 years of never stopping to watch
there have been systems that learned to generalize after seeing couple of examples not thousands (digit recognition)
child can see just one animal and label it as a monkey, an algorithm could probably do the same with more algorithmic machinery, but we are still not there
>The "breakthrough" AI of today is deep learning on massive amounts on data applied to two areas - speech/NLP and vision.
What of IBM's Watson? It's one of the premier AI projects in the world, its specialty is neither of those two areas, and it utilizes deep learning on massive amounts of data.
>How this generalizes to a super intelligence that can take over the planet is so strange.
No one's really saying it does, just that recent progress in AI has accelerated such that it seems probable that even more progress is imminent.
>The correct response is to say that you don't know anything about AI.
I really tire of this constant appeal to authority, it's arrogant at best. By that same logic, Nick Bostrom shouldn't have even written his latest book due to a lack of technical knowledge.
Watson has a very superficial understanding of what is being asked. It would be able to tell you the answer based on the data it was trained on but it wouldn't really understand what the question meant or what the response meant. I'm sure you'd have a very difficult time having a conversation with Watson or using it on anything other than Jeopardy style questions.
I never claimed Watson had anything beyond a superficial understanding, nor did I claim it was somehow generally intelligent.
My sole point was that it's a state-of-the-art AI project that utilizes deep learning, and its primary specialization is neither vision nor speech/NLP.
>... or using it on anything other than Jeopardy style questions.
Given how many verticals IBM is building Watson for, they would likely disagree strongly with that statement.
It's designed to interact with humans using natural language it's designed to break down normal speech and build a cognitive process around it to get answers.
I've always thought of Watson to be an automated data analysis platform at its core. However, after doing a bit of research, you're essentially correct.
Watson is technically multi-faceted, but the majority of its focus right now is NLP. In the future, I imagine that will change as they improve upon Watson's vision and structured data support.
The problem is that you ask someone who is good in one field (say electric cars or theoretical physics) to opine on something like AI. The correct response is to say that you don't know anything about AI. But, the ego of being an public intellectual prevents that. So, what is the safest option to not seeming dumb - say something like we need to make sure safeguards are place to protect AI from becoming dangerous and kill everyone.
Meanwhile people who actually build these systems know that these systems not are generalizable to variety of tasks (like humans) and they are not intelligent. Best case, they augment humans in their tasks.