3 ms·

>It's easier to argue it's not GPT-3 that's advanced, but it's humans that are simple. Many of the arguments made by GPT sceptics are alarmingly close to the

>It's easier to argue it's not GPT-3 that's advanced, but it's humans that are simple. Many of the arguments made by GPT sceptics are alarmingly close to the intelligent design argument against evolution. If a really dumb process with countless billions of iterations can produce human beings, why couldn't a different dumb process eventually produce human-level machine intelligence? Until the curve of progress shows some signs of flattening, I'm not going to speculate about what GPT-4 or GPT-5 will or won't be able to do.

notahacker 6y ago
The natural counter argument is that [i] the dumb process which culminated in cognition and consciousness consisted of billions of years of atoms consisting of our entire world interacting with each other, some in ways more condusive to life formation than others. Obviously we're working with a much less informationally complex process overall and just hoping we've got the right ingredients. [ii] the dumb process whereby humans try to generate machine intelligence looks very different to the factors thought to have been most relevant to the evolution from microorganisms towards the human mind. It may therefore not be possible for us to generate a mind at all with the inputs we're working with, and even if there's no practical bar to a sufficiently large number of human-assembled devices running binary code achieving cognition, our insistence on selecting based on intermediate goals like performing useful transformations on text might actually locally optimise to such an extend that the possibility of cognition is impeded. [inputting human ideas of good Go games into AlphaGo made it a less effective Go-learning process than AlphaGo Zero....] (and just as sceptic arguments can resemble ID and dualism, the pro-AGI argument that intelligence selectively guiding an evolutionary process is as plausible a route to consciousness as random chance and a sufficiently large quantity of atoms and time has similar superficial parallels with arguments for 'theistic evolution', not to mention near-future curve exponentiation arguments resembling millenialism)

notahacker 6y ago

The natural counter argument is that [i] the dumb process which culminated in cognition and consciousness consisted of billions of years of atoms consisting of our entire world interacting with each other, some in ways more condusive to life formation than others. Obviously we're working with a much less informationally complex process overall and just hoping we've got the right ingredients. [ii] the dumb process whereby humans try to generate machine intelligence looks very different to the factors thought to have been most relevant to the evolution from microorganisms towards the human mind. It may therefore not be possible for us to generate a mind at all with the inputs we're working with, and even if there's no practical bar to a sufficiently large number of human-assembled devices running binary code achieving cognition, our insistence on selecting based on intermediate goals like performing useful transformations on text might actually locally optimise to such an extend that the possibility of cognition is impeded. [inputting human ideas of good Go games into AlphaGo made it a less effective Go-learning process than AlphaGo Zero....] (and just as sceptic arguments can resemble ID and dualism, the pro-AGI argument that intelligence selectively guiding an evolutionary process is as plausible a route to consciousness as random chance and a sufficiently large quantity of atoms and time has similar superficial parallels with arguments for 'theistic evolution', not to mention near-future curve exponentiation arguments resembling millenialism)