I apologize for my late reply, but I was a bit busy the last weeks. I will be able to be more responsive the next weeks. Back to the issue:
Yes, at some point some actual physical experiments need to be done. I would not be surprised, however, if 1,000,000 AI computational fluid dynamics scientists with a single physical lab at their disposal greatly outperformed the current fluids research community. It's hard to be sure either way though. As per whether AI would give you something brute force wouldn't: algorithmic breakthroughs can be worth more than decades of Moore's law in terms of performance. New numerical algorithms have been known to yield speedups of 3 of even 6 orders of magnitude. Those speedups are something brute force processing power alone won't buy you. Having a million AI numerical analysts chugging away looking for new algorithms, however...
And I wouldn’t be surprised if a SI would not incredibly outperform the whole fluid dynamics community. Well, that’s not totally right. You are right that with the broad availability of computers huge progress in numerical methods were achieved as well (I would even say that the progress comes into the same order of magnitude as the progress in microelectronics). So if you compare the way they did their calculations in the 70s with todays framwork, yes I agree. But appearently the progress wasn’t as steady as the progress of processing power, which even today continues to improve according to Moore’s law. Of course I can not exclude that AI will find ways or a new foundations of math current researchers don’t even think about. But this comparison means that we just could kind of compensate for the lack of AI by increasing processing power. Not very honorable for the former… (and not comparable to genomics).
I would believe that as well. And I would also believe that experimental physics has a higher marginal return than theoretical physics, precisely because it seems we have hit a wall in terms of armchair reasoning ala Einstein. Nonetheless, AI would greatly change the calculation. Currently it may be that the cost of funding an experimental researcher plus his lab is more worthwhile than the small cost of funding a theoretical researcher. With AI, however, we have to ask whether the experimental lab is more worthwhile than the number of theoretical researchers that can be run on silicon for an equal cost. This may be thousands or millions of researchers. And those researchers can all be copies of the *best researchers*. At some point they'll need to test some theories, but it seems likely that they could be incredibly more productive than today's research community with even less physical experiments to test their hypotheses.
I would disagree with this interpretation (though primarily speaking for optics/laser). I admit that in a number of areas there is a lack of (correct/validated) theoretical models, e.g. surface-field interactions. But this is not primarily because physicists can not connect the dots in their empirical data; at least judging from the current understanding about the gaps in our knowledge. Hypothetically speaking I can of course not exclude that it nonetheless might be so – but just because we can not exclude it doesn’t mean that it has to be the case either, so it is somewhat pointless to debate this or even base SI revolution predictions on it. I take it more that in many cases the available empirical data is insufficient to develop good theoretical descriptions (given that some kinds of lasers and radiation sources such as free electron lasers are available since just a couple of years this is not too suprising). Other research objects as the amazing Bose-Einstein-Condensates are full of predictions where the issue is currently to validate those empirically and more funding is required for experimental confirmations. In case of the (STED) super-resolution microscopy the researchers even had to promote their ideas for years before actually getting funding to build it. In short, I do not see a broad stagnation caused by not connecting the dots of currently available empirical data, but more the need to test the existing theoretical concepts or to do the first time experiments you need to ground an extended theory on.
Concerning the “worthwhile” I agree that funding bodies are too fixed on requiring short term benefits out of projects; but this applies equally to experimental and theoretical projects.
New fabrication techniques are very experimental, I agree. I was referring to chip design. Placing billions of transistors in a way so as to maximize performance is a difficult problem, but one which (I think) can largely be done on a theoretical basis alone. Look at field programmable gate arrays (FPGAs) for instance. If you program an FPGA to implement your particular piece of software you can get *Huge* performance boosts. Yet few people do it because it's *hard* to program FPGAs. This bottleneck would be removed with hordes of AI FPGA programmers, yielding a huge speedup in software across the board. Designing a chip from scratch to implement a piece of software gets you even greater speedups, but the design costs are even more prohibitive. A horde of AI chip designers would yield even bigger speedups.
I am not very familiar with FPGAs; judging from a quick google search it seems that there are considerable tradeoffs compared to ASICs, even if you can program them perfectly. Anyway, in principle I agree that SI will speed up things, but in the end it I think we also have to define what you consider to be a technological singularity and if those “traditional” processing power speed ups satisfy it or if need a new kind of physics.
Yes, incredibly expensive to *design*. The hardware may be expensive too, but this is usually because economy of scales hasn't kicked in. That is, Asimo isn't put together by a robot like a memory chip or a car is. But that's because it currently isn't economical to design and program the robots to build Asimo. With hordes of AI robotic researchers/engineers that need not be the case. Indeed, we already have sufficiently sophisticated robotics to construct 'universal' construction robots, by which I mean robots capable of filling essentially any role currently filled by low-wage labor. Those robots could be mass produced at cheap prices and installed in factories and farms all over the world.... except we don't have software good enough to make any use of them right now. With today's technology we can make a robotic arm with 30 degrees of freedom, but writing a program to use that arm in an intelligent and robust way is beyond us. Robotics is not limited by hardware, it's limited by software. AI would fix that.
You seem to imply, that the costs of designing the digital model constitute the bulk of the price of industrial robots or whatever machine you employ for automatisation. That’s simply not the case (enlighten me if I am misguided). The reason why building automated factories for humanoid robots is not done is not the design costs for the required automation, but the low demand for mass producing HRs. It’s a relatively new product which still needs to be perfected. Also the problem is not (primarily) programming the degrees of freedom, but getting conceptual knowledge into the memory of robot workers. As seen in one of the videos, e.g. a cleaning robot does need to know/learn which kind of items it can pick up, leave on the ground and clean around it, what kind of pressure to use etc.
Its not different from humans, as we basically need the first 6, 7 years to learn those concepts, the control of our body and environmental interaction. Other than in case of humans of course we can copy the learned things for mass production. And as this is more a step in getting SI in the first place we unfortunately can not count on a non-existing SI to do this work for us. Leading me to my previous point that either way, robotics is our critical technology for overcoming current research limitations, and even SI will be much, much less useful without robotics provided (also robotics appears to be currently a much more feasible technology than SI).
Responding in particular to "If you can replace those humans and have an unlimited supply of automated systems/robots you have largely elimanted scarcity of research resources/resources in general without SI."
Automation can and does have profound effects on research, but you still need people doing the actual science and analysis. Science does not progress solely by means of doing experiments. The job of theoreticians is not to simply provide a constant stream of random hypotheses and experiments for experimentalists to perform in the hopes that such a random walk will eventually stumble upon better theories. In general, the more a theory is developed before it is experimentally tested the easier it becomes to falsify the theory. That is, theoretical work increases the usefulness of experiments. It may be that 'simple' automation will exhaust the potential for reducing bottlenecks in physical experiments (although I think AI will get us to that point much, much more quickly), but what AI buys us is an increase in theoretical capacity.
Yes, SI (I use the more general term of super intelligence), definitely increase theoretical research – that’s its sole point which I acknowledged since starting the discussion. However, my point was that even current theoretical knowledge has no problem at all in exhausting the ever increased research resources; and they don’t even need a stream of random hypothesis for that. But for an SI to get a much better idea of theory + massive resource savings even in those cases where funding is provided for experiments (and there is much research where funding is not even available due to limited research resources) this means that the whole human research community would simply overlook so much, leading to many/most experiments being not needed, and this at a broad range of fields within the sciences. This might be the case, but to use this as basis of your singularity argument this assumption needs to be justified.
Imagine two groups, both with fixed resources for physical experiments. The first group is today's research community. The second group is an AI research community, with 1000 times as many researchers, all of them of the highest caliber, each running 1000 times faster than a human. Both groups are limited by resources for experiments. The second group, however, will have vasts amounts of mental capital to spend on selecting the best experiments to perform. String theory isn't testable yet? Fine, have a few thousand researchers flesh out the mathematical theory for a few hundred years (a few months real time) and see if any testable ramifications are found. Repeat for a dozen flavors of quantum gravity. Sure, at some point you've gotta do an actual experiment, but it seems to me the second group is going to have a huge leg up in choosing good experiments.
The same issue as described before: maybe we overlook the obvious (or the not so obvious) – maybe not and LHC will bring us the empirical input for Higgs/Quantum gravity. And as there are currently no experiments underway explicitly for proving string theory this doesn’t contribute to the restriction on resources, i.e. money spend on proving string theory is money not spend on investigating other fields of research (minduploading etc.).
I still laugh every time I watch this
I cry a little inside too =O
This is not exactly a sober comment. Stick to the facts: in the last decade robotics, especially HRs made huge progress. Its like they figured out the principles (e.g. building the first airplane or transistor) and now are on for steady progress – can you claim the same for SI-research? If you look at the pipeline of progress you see that every year new models with advanced features are scheduled. Assuming slow progress in robotics is unjustified. Of course more funding would yield more progress, as always…
New HR models cost about 250.000 – 400.000 dollars, this on a prototype level without mass production. Mass production should bring down it to the level of an average car, at which point and with even further enhanced hard- and software it’ll be cheaper to use them even for blue collar jobs.
Look, the message I try to make is, people here tend to put way too much importance into SI/AI, which is no wonder machine. It makes no sense for transhumanists or singulairtaerians to hold back support for life extension research today, assuming that SI will do todays research in no time, assuming it generally doesn’t need empirics or normal Is are generally just very bad in connecting dots (and apparently assuming that they are still in good shape when SI is available).
As to the dangers of AI: it really depends whether we put AIs in charge of everything. Give it the controle over the US-nukes and it just might decide to pressure humans to serve its goals. Actually there is a relatively entertainig movie about this scenario
http://en.wikipedia...._Forbin_Project
LongeCity
