3
$\begingroup$

It seems to me that the first AGIs ought to be able to perform the same sort and variety of tasks as people, with the most computationally strenuous tasks taking an amount of time compared to how long a person would take. If this is the case, and people have yet to develop basic AGI (meaning it's a difficult task), should we be concerned if AGI is developed? It would seem to me that any fears about a newly developed AGI, in this case, should be the same as fears about a newborn child.

$\endgroup$
3
$\begingroup$

There are basically two worries:

If we create an AGI that is a slightly better AGI-programmer than its creators, it might be able to improve its own source code to become even more intelligent. Which would enable it to improve its source code even more etc. Such a selfimproving seed AI might very quickly become superintelligent.

The other scenario is that intelligence is such a complicated algorithmic task, that when we finally crack it, there will be a significant hardware overhang. So the "intelligence algorithm" would be human level on 2030 hardware, but we figure it out in 2050. In that case we would immediately have superintelligent AI without ever creating human level AI. This scenario is especially likely because development often requires a lot of test runs to tweak parameters and try out different ideas.

$\endgroup$
  • 1
    $\begingroup$ In reference to the first option: why can't we make now a piece of relatively simple software that is able to improve its own code, so it would spiral out from current constraint of waiting for people to create machine slightly better than themselves? $\endgroup$ – Alex Feb 23 '18 at 15:05
  • $\begingroup$ Please direct me to a full question regarding this if there is one. Are there ideas out there (say, in Bostrom's "Superintelligence") that detail how to kill the AGI before it kills us, supposing we can keep it contained to, for instance, one machine that we've kept detached from the internet? $\endgroup$ – frank Jul 7 at 4:23
0
$\begingroup$

This depends on the definition(s) of AGI and ASI. Both are currently ill-defined. Most researchers in AGI follow their own definition of AGI.

At least one researcher believes that there is no such thing as ASI. This is because the basic principles of said AGI always stay the same. It may be learning processes, the core logic(s) and the control logic (reasoning systems are divided into control systems and logic systems, the control system(s) decide which derivations are fruitful).

ASI may be defined as a search for any combination of these (just a subset which come to my mind):

  • search for better algorithms
    • heuristics
    • better contemporary (NN) architectures
  • learning mechanisms
  • solving techniques
  • higher subjective beauty
    • better compression of knowledge
  • better subsystems
    • NN and in general architectures
    • better embedded AGI's
  • faster solving capabilities of known problems
  • ...

There are limitations to any sort of (recursive) self improvement however. Examples of these are * the score of AlphaGo and AlphaGo-Zero plateaus after a long enough training period * supercompilation of a supercompiled program yields no improved program after a few iterations * ... Note here that these are examples about weak-AI and may not apply to AGI - but it is very likely in my opinion.


So the level of worry depends on the plausible (or followed) definition of AGI and the assumptions of the mechanisms an AGI may employ.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.