[timmytokyo]

>There isn't a testable definition of GI [...]

This to me is the fundamental issue in discussions and debates about LLMs. Despite assertions by some psychologists (who themselves are practitioners of perhaps the fuzziest of "sciences"), intelligence is an entirely nebulous concept. Everyone means something different when they use the word. I can think of no better illustration of the problem than the authors of the "Sparks of AGI" paper resorting to a definition of intelligence presented in the Wall Street Journal of all places. That the WSJ definition was part of an editorial defending the Bell Curve is just the cherry on top.

[janalsncm]

Do you know what their definition was by any chance?

And yes, a cursory glance at the Wikipedia page for intelligence shows there’s no one agreed upon definition of intelligence.

A more useful framing is to say we’re not creating “intelligence” per se but automating tasks. GPT4 is an automated writer. Stable diffusion is an automated image creator. Alpha Go was an automated Go player. Google search automates the work of a reference librarian.

With that in mind, it’s immediately obvious how much of a waste of time it is to argue whether ChatGPT is “intelligent” or not. Who cares. What we are doing is automating all of the things which brains used to do.

[TheOtherHobbes]

One problem is that academic CS-researcher intelligence is completely different to average human intelligence.

Maybe 5% of the population can learn how to solve partial differential equations.

Virtually all of the population can manage extended family-related conversations over Christmas. Even when drunk.

Human intelligence is mostly social, and mostly not scientific. The average human is incredibly bad at model building and self-correcting prediction. What actually happens is that humans have developed a kind of collective cultural exoskeleton which protects - more or less - from the consequences of poor choices.

But it doesn't take much for that to stop working. Covid denial and climate change denial are just two examples.

The cost if living in this space is having to learn a lot of heavily scripted cues. There's a long list of acceptable and unacceptable behaviours and social registers in different social situations. It varies by culture. But generally humans can navigate this space without thinking too hard about it.

Academic intelligence is completely different. There's long been a joke that an AI researcher's ideal intelligent system is another AI researcher, with typical AI researcher interests - math, puzzles, abstract language models, music in an engineering way, and so on.

Current LLMs are the first cross-over product which shows signs of moving into the first space from the second.

You can imagine a future system which uses facial and gait profiling to read emotions, and links a tokenised language model with a tokenised model of various transitions through emotional and social states. Personal background will be missing, and that's not hard to invent.

And now you have something that mimics a large part of social intelligence.

Only it has the potential to do it better than humans do.