Comment on We need to start calling it Simulater Intelligence (SI): here's why:
9bananas@feddit.org 1 day agowhen it can come up with a solution it hasn’t seen before.
that’s the threshold.
that’s the threshold for creative problem solving, which isn’t all there is to intelligence, but i think it’s fair to say it’s the most crucial part for a machine intelligence.
lendra@feddit.org 1 day ago
It can come up with a brand new sentence that hasn’t been written before. Does that count?
Maybe you mean a solution to a textbook math/physics problem, it most likely would be able to solve that too with tool use.
Or maybe you mean solving something like the Riemann Hypothesis?
9bananas@feddit.org 1 day ago
no, none of those are what i mean, that’s way too specific to be useful.
a system exhibits intelligence when it can use existing insights to build entirely new insights.
a popular example is that no current “AI” can extrapolate from basic mathematical stipulations to more advanced ones.
(there’s tons of example you could put here, but this is the one i like)
here’s the example:
teach an LLM/DNN/etc. basic addition, subtraction, multiplication, and division.
give it some arbitrary, but large, number of problems to solve.
it will eventually encounter a division that isn’t possible, but is not a divide-by-zero (which should be covered by the rules it was given).
then it will either:
…but what it will definitely NEVER do, is simply create a placeholder for that operation and give it a name: square root (or whatever ot calls it, that part isn’t important).
it simply can’t, because that would be a new insight, and that’s something these systems aren’t capable of.
a human (or a lot of them) would encounter these impossible divisions and eventually see a pattern in them and draw the proper conclusion: that this is a new bit of math that was just discovered! with new rules, and new applications!
even if it takes a hundred years and scores of them, humans will always, eventually, figure it out.
…but what we currently call “artificial intelligence” will simply never understand that. the machine won’t do that, no matter how many machines you throw at the problem.
because it’s not a matter of quantity, but of quality.
and that qualitative difference is intelligence!
(note: solving this particular math problem is a first step. it’s unlikely that it will immediately lead to an AGI, but it is an excellent proof-of-concept)
this is also why LLMs aren’t really getting any better; it’s a structural problem that can’t be solved with bigger data sets.
it’s a fundamental design flaw we haven’t yet solved.
current "AI"s are probably a part of the solution, but they are, definitely, not THE solution.
we’ve come closer to an AI, but we’re not there.