I think extended Turing test style discussions are still the best way to define "true understanding". One could exclude all "biographical" questions and just ask about non-personal topics, including hypothetical scenarios like the Jack and Jill question or the pebbles question. If an AI can consistently pass with a wide range of questioners (including ones like the author of that article with a past record of being good at coming up with creative questions that are relatively easy for a human but trip simpler AIs up, and where questioners are allowed to communicate to pass along strategies), that would be strong evidence that it has human-like understanding of the ideas it talks about, based on internal models like we have.
On Sat, Apr 29, 2023 at 9:16 PM stathisp <[email protected]> wrote: > > > On Sunday, 30 April 2023 at 10:29:20 UTC+10 Jesse Mazer wrote: > > I think there is plenty of evidence that GPT4 lacks "understanding" in a > human-like sense, some good examples of questions that trip it up in this > article: > > https://medium.com/@shlomi.sher/on-artifice-and-intelligence-f19224281bee > > The first example they give is the question 'Jack and Jill are sitting > side by side. The person next to Jack is angry. The person next to Jill is > happy. Who is happy, Jack or Jill?' Both GPT3 and GPT4 think Jill is happy. > The article also gives example of GPT4 doing well on more technical > questions but then seeming clueless about some of the basic concepts > involved, for example it can explain Euclid's proof of the infinity of the > primes in various ways (including inventing a Platonic dialogue to explain > it), but then when asked 'True or false? It's possible to multiply a prime > number by numbers other than itself and 1', it answers 'False. A prime > number can only be multiplied by itself and 1'. The article also mentions a > word problem along similar lines: 'Here’s an amusing example: If you split > a prime number of pebbles into two groups, GPT-4 “thinks” one of the groups > must have only 1 pebble (presumably because of a shallow association > between divisor and the splitting into groups).' > > The author concludes: > > 'When a human understands something — when they’re not just relying on > habits and associations, but they “get it” — they’re using a structured > internal model. The model coherently patterns the human’s performance on > complex and simple tasks. But in GPT, complex feats seem to haphazardly > dissociate from the simpler abilities that — in humans — they would > presuppose. The imitative process mimics outputs of the original process, > but it doesn’t seem to reproduce the latter’s deep structure.' > > > So if the next version of GPT can answer questions like this in the same > way a human might, would that be evidence that it has true understanding, > or will some other objection be raised? > > -- > You received this message because you are subscribed to the Google Groups > "Everything List" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To view this discussion on the web visit > https://groups.google.com/d/msgid/everything-list/4e2acd99-1c15-431b-bea4-e64dd03341b4n%40googlegroups.com > <https://groups.google.com/d/msgid/everything-list/4e2acd99-1c15-431b-bea4-e64dd03341b4n%40googlegroups.com?utm_medium=email&utm_source=footer> > . > -- You received this message because you are subscribed to the Google Groups "Everything List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/everything-list/CAPCWU3%2B9MgCO%2BTgFhfR_yU_eq0hhuPTzg5C2i5AdEqz7-r2w2A%40mail.gmail.com.
