4 ms·

> it seems logic is a weak foundation for intelligence What an offensive statement. Over two thousand years of fighting intuition-based judgements, and now th

> it seems logic is a weak foundation for intelligence What an offensive statement. Over two thousand years of fighting intuition-based judgements, and now that DL can merely tell a gorilla from a mouse (with 712 hours of training on 91 GPUs), it is best form of "intelligence" but Logic - a fundamental part of mathematics that helped us u.n.d.e.r.s.t.a.n.d the universe better - is a "weak foundation for intelligence?

jcelerier 8y ago
> and now that DL can merely tell a gorilla from a mouse (with 712 hours of training on 91 GPUs) is a baby able to tell a gorilla from a mouse after ~2 months of being awake ?
- astrobe_ 8y ago
  I strongly believe that what you don't want for an AI is for it to make the same mistakes and have the same weaknesses as humans. For instance, people will never accept that the AI of a car kills by accident, even thought human drivers kill a lot more etc. A system that makes life-or-death decisions but has no liability nor ethics is difficult to accept. The article talks about expectations - unrealistic expectations even. People expect from AI that it makes clever decisions like humans and "no mistakes" like machines (scare quotes for bugs, failures etc.). Now, I don't suggest that DL is useless. It is good at what it does, but contrary to what OP suggests, I don't see it as the sole or even main component of an AI system.
opless 8y ago
Logic (needs to be specified by GP, but assuming predicate logic, as we're talking prolog here!) can't hold probabilities (very well) or contradictions.
- astrobe_ 8y ago
  And you have very faint hopes to have a DL system explain to you why it thinks it is a gorilla. Each system has its weakness. Still, logic systems are a lot closer to reasoning and understanding (if not for the machine itself, at least for its operators), in other words closer to intelligence than other systems.
  - opless 8y ago
    Oh hell no! DL are plain incapable of reflection/showing its workings. I'm just asserting that the GP was probably considering a small subset of logic based systems and was incorrect in their assumptions.

jcelerier 8y ago

> and now that DL can merely tell a gorilla from a mouse (with 712 hours of training on 91 GPUs) is a baby able to tell a gorilla from a mouse after ~2 months of being awake ?

astrobe_ 8y ago
I strongly believe that what you don't want for an AI is for it to make the same mistakes and have the same weaknesses as humans. For instance, people will never accept that the AI of a car kills by accident, even thought human drivers kill a lot more etc. A system that makes life-or-death decisions but has no liability nor ethics is difficult to accept. The article talks about expectations - unrealistic expectations even. People expect from AI that it makes clever decisions like humans and "no mistakes" like machines (scare quotes for bugs, failures etc.). Now, I don't suggest that DL is useless. It is good at what it does, but contrary to what OP suggests, I don't see it as the sole or even main component of an AI system.

opless 8y ago

Logic (needs to be specified by GP, but assuming predicate logic, as we're talking prolog here!) can't hold probabilities (very well) or contradictions.

astrobe_ 8y ago
And you have very faint hopes to have a DL system explain to you why it thinks it is a gorilla. Each system has its weakness. Still, logic systems are a lot closer to reasoning and understanding (if not for the machine itself, at least for its operators), in other words closer to intelligence than other systems.
- opless 8y ago
  Oh hell no! DL are plain incapable of reflection/showing its workings. I'm just asserting that the GP was probably considering a small subset of logic based systems and was incorrect in their assumptions.