[rozetta]

Hi all!

I'm in the process of writing some fiction and am researching AGI in order to get the most accurate information possible for my books. Time and time again, I've seen reasoning for why we don't yet have AGI being based on processor speeds. I'm wondering why this is a factor, and would like to know what your thoughts are on this.

My reasoning is as follows. Firstly, I am assuming that we're not making robots or anything that needs to interact with physical reality in real-time. I'm wondering why we can't create an AGI that has few inputs (text, visual, audio) using mostly language processing, image processing, perhaps some visual processing and audio processing in order to learn by conversing with human tutors, reading the web and reading it's own code in order to learn and improve.

If you think about Stephen Hawking, since his illness set in, he has been confined to a very limited amount of options for interaction with the world, and yet, he is still brilliant and is able to push the boundaries of his field, despite lacking access to physical abilities that most of us have.

My thoughts are that an AGI should not necessarily need to react to inputs in real-time if processing speeds are not available yet to do this. As long as it is real AGI and able to think on a human level, it will be able to self-improve and will ultimately benefit from advances in computing speed as they become available. Perhaps a system could be built that is distributed over many systems?

Is real-time physical interaction so key to building AGI?

Answers would be much appreciated!

[djb]

The brain is parallel, computers are serial. tens of billions of cells in parallel would required massive processor speeds if done serially. This is why most mainframes are now parallel and even basic PC's are "dual core" and more. True AI will likely require lots of processors acting in a group. My theory is that you will need 200-500 thousand basic processors acting in parallel to produce a true AI.

[Ultron]

To feed your imagination - consider this:


Computer processor pins and some internal contacts (as well as others on the motherboard) are made out of gold, since gold is the best conducting element. It is better then silver, which is better then the classic copper. So, higher conductivity = faster processing speeds. Graphene is even better then gold and this is why research in that field is so heavily funded - but we don't use it because it is... well, expensive.


Now imagine those metal contacts (along with cables etc.) to be replaced by something radically new, but also antique - something, organic... Biological machines used to be considered and predicted as the next (Fifth) era of computing, before all that quantum nonsense came out. This is interesting for AGIs and robots... An interesting scenario would be a non-AI (stupid robot) to turn into an intelligent one after some spontaneous mutation in his biological (bio-neural?) circuitry.


If you are hungry for more technical details and already satisfied with your own imagination, then I suggest keeping some key points in mind. I already pointed out the general hardware dependencies (conductivity), but it is also worth noting that the more transistors there are in a processor (and also, the smaller they are) increases processing speeds. Cores and 'threads' help this machine carry out parallel tasks (more at the same time). Maybe an interesting idea here is having a single-molecule transistor and call it a 'quantum transistor'... Yeah, that would be insane - and plausible, actually!


Another key point is software. The program of the system dictates the efficiency - there is always a better and faster way to solve a problem. The more efficient a program is (the less tasks it carries out for the same task) the faster the machine solves/executes it.


I hope this was helpful to you. Good luck with your fiction and we'd enjoy reading something and appreciate if you would one day share it on this forum.

[ivan.moony]

Hello rozetta

My answer will not be so inspirational, but I think it is close to reality. We (humanity) didn't built an AGI yet because we are still stupid as a specie and no raw computer power or speed can not help us to run away from that fact. Simply, generally we luck inventiveness, but who knows, maybe somewhere in the world, in some forgotten country, in some dark basement a wonder kid is just now bombastically pressing "enter" key on his old Pentium to run the very first AI program that would change our lives for ever... I'm looking forward to that moment 

Good luck with art, I've always loved it.

[Korrelan]

Hi

The essence of intelligence has very little to do with brain/ processor size or speed.

Many small animals exhibit extraordinary intelligence.  Look at Corvids, or Psittaciformes; there are many studies and videos on YouTube showing what they can accomplish.  Ant swarm intelligence is another example.  Even Slime Mold can solve the ‘salesman problem’.

If you think about Stephen Hawking

He grew up as a normal active human, the experiences were laid down while is cortex was maturing.  Although he can’t move now; his motor regions don’t degrade because he can still see other people moving.  I think if we see someone running, it fires the same cortex patterns as when we ourselves run.

Is real-time physical interaction so key to building AGI?

Yes. Certain types of knowledge/ skills require a personal understanding/ experience of gravity, inertia etc.  I think an AGI needs physical experience to get a complete grasp (no pun) on all types of knowledge/ skills.

@Ultron -

The program of the system dictates the efficiency

I agree.  The speed of the hardware/ software does dictate the speed of the simulation, but if the system is created correctly if will self optimize internal data processing just like we do.

An analogy; If the sequence 1,2,3,4,5 always =6 through experience; then 1,2 =6 as a prediction.  If 1,2,8,4,5=7 then the best optimization/ prediction would be 1,2,3=6 and 1,2,8=7. 

[Don Patrick]

I believe the idea was that a human brain does X amount of calculations per second, therefore an AGI should do as much. Personally I don't think this is necessary, but low speeds can be a drawback. Like the old supercomputers, a slow AGI may take minutes to compute a problem that has already changed by the time the AGI figured out a solution.
Despite that, IBM did connect a ridiculous amount of servers and processors to match human speed in anwering Jeopardy questions even with current technology.

Some believe that real-time physical interaction and full human sensory systems are necessary to overcome what they call the "symbol grounding problem", which is that a computer can not "understand" the meaning of the alphabetic symbols "c-o-w" if it only read these symbols instead of having real physical interaction with a cow. Personally I think that applies to a lot of city folk, so I support the reasoning laid out in the first post that none of this is a prerequisite to intelligence, but a convenience.

[Ultron]

I believe the idea was that a human brain does X amount of calculations per second, therefore an AGI should do as much. Personally I don't think this is necessary, but low speeds can be a drawback.

Do we even have a raw number for this? At least an estimate?

Some believe that real-time physical interaction and full human sensory systems are necessary to overcome what they call the "symbol grounding problem"...

Sorry to brake it to you, but 'real-time' in this case translates to fast (not even 'very fast'). It takes up quite a few milliseconds (or microseconds - whatever) for environmental information to reach the sensors and cause them to react. They later send this information to cells from the peripheral nervous system which take quite some time to relay it to the brain (central nervous system).

[ivan.moony]

I believe the idea was that a human brain does X amount of calculations per second, therefore an AGI should do as much. Personally I don't think this is necessary, but low speeds can be a drawback.

Do we even have a raw number for this? At least an estimate?

You can measure it if you want. How many numbers can you multiply or add in one minute?

But operations, I think, may be as complex as we can imagine. If we are inventive enough, we can imagine and define an operation in our mind that does dozens of other operations in just few steps. And if we use it often enough, we can reduce the time to less than a second. I think we are not that much fast, it is all about optimization.

[Ultron]

You can measure it if you want. How many numbers can you multiply or add in one minute?

Umm, no... I was talking overall processing capacity of our mind. You cannot consciously 'measure' this - our CNS controls many parts of our body subconsciously - you cannot release adrenaline at will (for example), but your mind does 'calculate' the right time and amount to be released. Our conscience is merely one part (or 'thread') of our mind - this is maybe why some people say we use only a few percent of our minds and why different research projects contradict each-other.

[Don Patrick]

I'm sure the scientists know the estimated number of computations. I don't recall it since it doesn't interest me. You can look it up.
http://www.extremetech.com/extreme/163051-simulating-1-second-of-human-brain-activity-takes-82944-processors

[Freddy]

Nice little article, thanks.

[Ultron]

Sorry Don but I think that our methods of scanning neural activity are somewhat insufficient - we don't understand all those signals. One could make an estimate depending on the power consumption - which is absurd since many more variables need to be taken into account.


But if the number is really somewhere around 83 thousand, then I suppose we will have mobile devices that can mimic the brain in some, umm... 30 years? 40 tops.


Edit: Forgot to mention that the structure hierarchy of a processor is insanely different from that of the brain, to say the least. And note that we only have theories regarding that structure.

[rozetta]

Thanks for all of your answers so far. It seems there isn't even a definitive opinion on the answer to this question, which I find refreshing. I posed it originally after looking into what the guys at OpenCOG were up to.

In my mind, one of the ways the physical world problem might be overcome is to pre-populate an AGI's mind with enough "common sense" facts with high truth values that it would be able to understand physical world concepts without necessarily having experienced them.

To my mind, hooking up an AGI to a robot would require the robot to be able to process experiences at the speed humans can, which is ultimately limiting us with the need to wait for processor speeds to arrive at where we'd need them (basically in about 10 years time).

However, building a "thinking machine" that has limited input (direct, visual, audio) and carefully teaching it the concepts needed to start understanding and reading through something like wikipedia would allow a fundamental knowledge base to be created prior to the technology catching up. With this in mind, if one could create an AGI that is focused more on mastering things like theoretical knowledge and the ability to understand and improve its own code, it might go on to create its own internal simulations of the real world. Or it might go on to self-improve to the point that increased processing power is no longer needed. Robotics could be hooked up later.

I understand that many concepts that are prevalent in everyday literature would be incomprehensible to an intelligence without a physical embodiment. However, even with one, many concepts such as smell, taste, touch, pain, temperature and many emotions would still be foreign to it.

[Ultron]

In my mind, one of the ways the physical world problem might be overcome is to pre-populate an AGI's mind with enough "common sense" facts with high truth values that it would be able to understand physical world concepts without necessarily having experienced them.

The less information / knowledge you hardcode, the better. Yes, evolution / breeding passes on information to the offspring / new generation which is essential to it's survival - this can be an instinctive response to falling, ability to recognize 'bad' smell and 'good' food etc. In the field of A.I. I would refer to this as 'Simulated Instinct'. But remember that the first unit in a generation (the beginning of evolution) had no such knowledge... It would be a much more powerful experiment if one were to witness the evolution of an A.I. from the complete beginning.

To my mind, hooking up an AGI to a robot would require the robot to be able to process experiences at the speed humans can, which is ultimately limiting us with the need to wait for processor speeds to arrive at where we'd need them (basically in about 10 years time).

Why? AGI does not have to mean human intelligence - why not experiment with mimicking the intelligence of a dog? Why not create a 'dumber' creature? Researchers have succeeded in making a complete simulation of a worm's neural network (300 units), so why know it is possible.

I have long theorized that an experiment such as that would result in the creation of a new species - it does not matter if it is organic or inorganic, as long as it is intelligent (able to sense and derive output from the information gathered).

I understand that many concepts that are prevalent in everyday literature would be incomprehensible to an intelligence without a physical embodiment. However, even with one, many concepts such as smell, taste, touch, pain, temperature and many emotions would still be foreign to it.

Totally, just like we can't imagine what infrared light would look like - find a topic named 'imagining new colors'.