This is obviously valuable work in general, because a lot of people don't read science fiction, blogs, dream about technology, etc. I question how valuable this is to the HN audience who spends much of their time thinking about these topics, though. I spotted nothing that stood out from a skim through this article. Maybe the full report has more, but most of what it covers seems to be things along the lines of "who will have liability when a self-driving car kills someone?"...
Any specific format/information you are interested in? We are using AI heavily on our platform to improve urban living. I would be willing to share to my best knowledge.
Its the same for me, but I suppose if you're not actively looking into buying an electric car he doesn't directly affect many people's lives. Indirectly yes, but not so much directly. Other than that space technology hasn't been big in the popular conscience for many decades now and solar is boring power utility stuff.
Until the iPod most people had never heard of Steve Jobs, or if they had he was swiftly forgotten as just another computer guy.
We don't hear much talk about quantum computing and AI, is this because general purpose QP is considered unrealistic? Another interesting aspect in the podcast is that "doctors can't be replaced because they are able to interact with a patient", yet boston dynamics have shown us some pretty interesting progress in robotics, and we have some pretty amazing imaging and recognition software. The final thing I find interesting that people often refute on is that there isn't access to a good training set on humanity to start to understand us, well, I'd suggest that the WWW is an good collection of human knowledge.
The bigger problem in replacing doctors is actually getting enough data for machine learning in the medical field because of privacy issues. For instance you can't give out head CTs because you can reconstruct how the face that belongs to it looks like.
It's mostly because (to my knowledge) we don't currently know of any ways in which quantum computing will make artificial intelligence faster or more feasible. If you read up on the relationship between quantum and classical complexity classes, and what effective quantum algorithms we've developed so far, this may be more clear.
These types of reports seem overly conservative. Unless you think there is something magic about human or animal brains (many people still do believe this sort of thing or think there is some quantum (read 'magic') feature) then you have to anticipate a strong possibility that we will be able to emulate most of human capabilities and traits with hardware.
When we get to that point, it is very likely we will be able to improve the performance of these AIs by speeding them up or interconnecting them, etc. So another likely possibility is that these AIs could be twice as smart as us. Many speculate thousands of times smarter, but I don't think there is any reason to assume that is feasible -- and twice as smart as us is significant enough.
So these AIs could become much smarter then us and because they are not constrained by biology could take any type of physical form.
I think that if you are trying to speculate cautiously then you should consider this very similar to an alien invasion. The one advantage we have is that we will be training/programming the first generations of these things, so we had better get that right. But after they become say twice as smart, normal humans probably won't have as much sway over them anymore.
I think that when you look at all of the amazing abilities from the last few years from Watson winning at Jeopardy, Deep Dream, Deep Mind with WaveNet speaking and arcade game learning, Atlas walking and picking up boxes, winning at Go, etc. -- these are all amazing and people are very serious about pursuing general AI again. So how do we know how many more fundamental breakthroughs are actually required to get to our truly general humanlike intelligence? Are we even sure that we don't have the techniques already and just need to combine them in a certain way?
So my way of thinking is that to speculate conservatively, guess that there may be two major research breakthroughs required on the same order as deep learning. If we had not seen such an increase in AGI research then that would sort of console me. But because AGI funding has increased and the belief is back that these types of goals are possible, it seems like we should not assume that it will take many decades to achieve these breakthroughs (if we actually need more breakthroughs). How do we know that one or two of the dozens (hundreds? thousands?) of literal geniuses with an adequate background who are working in this area will not come up with some new major breakthroughs in the next two years or five years or eight?
So yes my personal opinion is that we should start planning for an 'alien invasion' of superintelligent AGI which if we are being cautious could pretty much come at any time now, and certainly it might happen before 2030.
The rate of discovery certainly is very high, every week something amazing pops up, but the more I read, the more I realize just how far we still are from human level intelligence. We don't yet even have a robotic body able to fold a shirt or walk about on its own (BD robots were remote controlled), or a chatbot able to pass as human. The good thing is that there is interest and funding. Maybe we still need two or three fundamental discoveries, could take anywhere between 5 and 50 years.
I completely agree there's a ridiculous amount of work to be done for a human level intelligence -- I'd even speculate that there's little economic need for one too; most economic tasks, even high-level creative and strategic ones are much more narrowly defined, you could make do with systems that simply comply with their master's bidding in a request/response fashion, which seems antithetical to open-ended self-motivated actions of an introspective human being -- I'm not sure about your examples though, as these seem precisely where I've read about significant progress made lately:
- teleoperation is generally orthogonal to creating a robust bipedal gait, and it seems like capture point approaches by BD & whatever Shaft is doing, if one can believe their marketing videos, as well as underactuated systems like DURUS and ATRIAS (and its bird-like successor, can't find the name, C-something) are starting to look quite robust. Some work still to be done, I think but looking quite promising. And this is all just rather classical approaches; nothing machine-learning-y at all. You'd still teleoperate them to navigate and make them do a manipulative task, but not just because it can't walk otherwise; but because those are different domains and problems (like manipulation, autonomous navigation etc). Hell, ATRIAS is completely blind anyhow, all you can do is define foot clearance height and direction of movement.
- deep reinforcement learning based visuomotor control techniques demonstrated last year on the PR2 platform and later on I think also demonstrated on a Baxter robot seem pretty dextrous at such manipulative tasks like folding shirts, and does it realtime, after some trial and error self-training. Granted, didn't see them do exactly that task, but a bunch of other manipulative exercises were demonstrated, realtime and robust to environment destractions, perturbations etc.
Admittedly this later part is still limited in that the problem needs to be fully observable at all times by the current robot -- it can't deal w say occlusions for it has no memory of what it previously saw -- and its not actually trained to be robust to disturbances and distractions -- what it demonstrated on that front is essentially just what it gets for free from the robustness of deep approaches -- so it could be improved.
I know it's been done, but not done brilliantly, just barely. Just like walking, grasping (the Google robotic grasping experiment) - they are still a long way off. Seems like progress is advancing at a slow pace in robotics, unlike deep learning which is relatively much more successful with vision, audio and behavior (reinforcement learning).
15 comments
[ 3.1 ms ] story [ 61.7 ms ] threadUntil the iPod most people had never heard of Steve Jobs, or if they had he was swiftly forgotten as just another computer guy.
When we get to that point, it is very likely we will be able to improve the performance of these AIs by speeding them up or interconnecting them, etc. So another likely possibility is that these AIs could be twice as smart as us. Many speculate thousands of times smarter, but I don't think there is any reason to assume that is feasible -- and twice as smart as us is significant enough.
So these AIs could become much smarter then us and because they are not constrained by biology could take any type of physical form.
I think that if you are trying to speculate cautiously then you should consider this very similar to an alien invasion. The one advantage we have is that we will be training/programming the first generations of these things, so we had better get that right. But after they become say twice as smart, normal humans probably won't have as much sway over them anymore.
I think that when you look at all of the amazing abilities from the last few years from Watson winning at Jeopardy, Deep Dream, Deep Mind with WaveNet speaking and arcade game learning, Atlas walking and picking up boxes, winning at Go, etc. -- these are all amazing and people are very serious about pursuing general AI again. So how do we know how many more fundamental breakthroughs are actually required to get to our truly general humanlike intelligence? Are we even sure that we don't have the techniques already and just need to combine them in a certain way?
So my way of thinking is that to speculate conservatively, guess that there may be two major research breakthroughs required on the same order as deep learning. If we had not seen such an increase in AGI research then that would sort of console me. But because AGI funding has increased and the belief is back that these types of goals are possible, it seems like we should not assume that it will take many decades to achieve these breakthroughs (if we actually need more breakthroughs). How do we know that one or two of the dozens (hundreds? thousands?) of literal geniuses with an adequate background who are working in this area will not come up with some new major breakthroughs in the next two years or five years or eight?
So yes my personal opinion is that we should start planning for an 'alien invasion' of superintelligent AGI which if we are being cautious could pretty much come at any time now, and certainly it might happen before 2030.
- teleoperation is generally orthogonal to creating a robust bipedal gait, and it seems like capture point approaches by BD & whatever Shaft is doing, if one can believe their marketing videos, as well as underactuated systems like DURUS and ATRIAS (and its bird-like successor, can't find the name, C-something) are starting to look quite robust. Some work still to be done, I think but looking quite promising. And this is all just rather classical approaches; nothing machine-learning-y at all. You'd still teleoperate them to navigate and make them do a manipulative task, but not just because it can't walk otherwise; but because those are different domains and problems (like manipulation, autonomous navigation etc). Hell, ATRIAS is completely blind anyhow, all you can do is define foot clearance height and direction of movement.
- deep reinforcement learning based visuomotor control techniques demonstrated last year on the PR2 platform and later on I think also demonstrated on a Baxter robot seem pretty dextrous at such manipulative tasks like folding shirts, and does it realtime, after some trial and error self-training. Granted, didn't see them do exactly that task, but a bunch of other manipulative exercises were demonstrated, realtime and robust to environment destractions, perturbations etc.
Admittedly this later part is still limited in that the problem needs to be fully observable at all times by the current robot -- it can't deal w say occlusions for it has no memory of what it previously saw -- and its not actually trained to be robust to disturbances and distractions -- what it demonstrated on that front is essentially just what it gets for free from the robustness of deep approaches -- so it could be improved.