3 ms·

I would like to see some intelligence challenges that do not depend on the identity of the test taker. By that I mean that the results would be accepted as equa

I would like to see some intelligence challenges that do not depend on the identity of the test taker. By that I mean that the results would be accepted as equally valid for people, non-human animals, and machines. If DeepMind developed a system that could ace Raven's Progressive Matrices questions, I don't think that we would say "machines are now as smart as human geniuses." Turing's Imitation Game still seems like a valid challenge to me but is perhaps too stringent. Someone who is not literate in English would fail to convince the questioner even if otherwise intelligent. "General" human intelligence appears to have a very complicated and particular shape. If there is such a thing as universal intelligence (in the sense that there are universal constants), the boundaries of human intelligence in this space look like a gerrymandered voting district more than a symmetrical geometric object.

username90 6y ago
The real Turing test is "Get a job as a remote developer and keep it for a year while getting paid". An AI that can reliably do that is what I'd consider general intelligence. If it can't learn how to interact with all the API's and programs you need to complete your tasks such as email, chat, code etc then it is not a general intelligence.
- philipkglass 6y ago
  That seems like a strong enough test, but it also seems too strong in that 98% of gainfully employed adults could not pass it. I'd like to see a test that correlates well with our colloquial notions of human intelligence but that also generalizes to non-humans (including machines). Maybe those are fundamentally incompatible goals. Maybe the outcome of AI research is finding that colloquial notions of intelligence resemble "real" universal intelligence only as closely as the intuitive physics of throwing and catching a ball resembles the unification of quantum mechanics and general relativity.

username90 6y ago

The real Turing test is "Get a job as a remote developer and keep it for a year while getting paid". An AI that can reliably do that is what I'd consider general intelligence. If it can't learn how to interact with all the API's and programs you need to complete your tasks such as email, chat, code etc then it is not a general intelligence.

philipkglass 6y ago
That seems like a strong enough test, but it also seems too strong in that 98% of gainfully employed adults could not pass it. I'd like to see a test that correlates well with our colloquial notions of human intelligence but that also generalizes to non-humans (including machines). Maybe those are fundamentally incompatible goals. Maybe the outcome of AI research is finding that colloquial notions of intelligence resemble "real" universal intelligence only as closely as the intuitive physics of throwing and catching a ball resembles the unification of quantum mechanics and general relativity.