r/m’lady
2 Likes
Here’s another example of us being poorly equipped to measure AI.
I love Doug Hofstadter, and I love the Economist. If anyone can tell us about AI, it’s Hofstadtler, right?
But he is dead wrong here. And this is probably the worst article I’ve ever read in the Economist.
I would call gpt-3’s answers not just clueless but cluelessly clueless, meaning that gpt-3 has no idea that it has no idea about what it is saying.
With a prompt that clearly indicates that the task includes distinguishing sense from nonsense, GPT-3 is able to reliably distinguish Hofstadter and Bender’s nonsense questions from sensical ones.
That’s not to say that GPT-3 is super smart, it’s not that smart. But it’s easy to underestimate its brains by prompting it to act stupid.
Or (coincidentally?), Beetlejuice?
I understood that to be significant because it supposedly shows that gtp3 does not understand the content of its words. It is really just a fancy regression, with many normalized parameters, that makes predictions of new words based on collections of words from the past.
I have two difficulties imagining that we are close to sentient, self aware ai. The first is that human beings, our one example, don’t look at all like a deep neural net. We have “subsystems” that do, such as our ability to recognize faces. But ultimately we choose between a wide variety of strategies to predict and shape both our environments, and our bodies. When we use language, we are relating this to possible, embodied actions we might take, which gives it a deep meaning to us. And we are deeply self aware,
My second difficulty is that we do not understand, at all, how our brains give us self consciousness. Presumably it doesn’t depend entirely on the material of our brains. But it might depend on a kind of spatial, causal continuity that is not necessarily achieved by some kinds of computation.
To me, it seems extremely neglectful to say “it has no sense that its random babbling is random babbling” when it plainly knows that some things are nonsense.
Certainly, it’s interesting that it you need to prompt it, but that’s not the same conclusion.
I agree that it is just a fancy regression. But at this point, it’s hard to tell whether a fancy regression can have intelligence.
I agree with the rest of what you are saying here. But I think it’s possible we will soon have AIs that are not “conscious” and not “sentient” but are intelligent, understanding, and creative.
I also think this deserves a caveat, because we spend a lot of time thinking about things that are poetic, absurd, philosophical, religious, mathematical, or fictional.
I do think regression represents a kind of intelligence.
But i don’t think of it as a being with intelligence. And the kind of intelligence it represents in this case does not depend at all on the content of the words. I’m sure that understanding was important for building the model, and so the structure of the model contains a kind of echo of it that is apparent in its predictions/responses, but this is not the same as the model making decisions based on the meanings of the words:
But what are the meanings of the words? Some words have images or feelings behind them, but many words are simply defined as a combination of other words.
I would say what we are having is an “intellectual” conversation, because none of the words we’re using refer directly to anything physical. Nothing we’re saying here is accessible to an animal or a baby. And as we talk, we don’t need to evaluate these words back into anything physical or emotional. In fact it would be difficult to do so.
Even when we do have words that refer to physical things, “city”, “rock”, “dance”, “joy”, it’s not clear to me how much physical information we store in our brains, rather than just verbal relationships. Do you need to picture a king as a guy wearing a crown, or is it enough to think of his powers and duties and so on?
If a blind person uses the word rainbow in a story, is that less intelligent? Or is it okay because they relate by metaphor?
If a mathematician writes a dissertation, with hundreds of pages of the most abstruse math you can imagine, does that lose sight of intelligence?
I agree with this to some degree. I think you are talking about this paradox of the ideal vs the actual.
I do think that actual experience has to be involved somewhere, even if increasing layers of abstraction is layered on top of it.
I’m glad we understand each other.
I agree meaning has to be grounded in something physical to be uh meaningful. But I don’t think thinking/intelligence needs be grounded in meaning.
Maybe that’s just semantics…(And somewhat moot anyway, since of course AIs can also see and hear and sing and visualize.)…What we call “thinking”, “intelligence”, “knowing”, “conscious”, “creative”, “sapient”, etc. a lot of words that we haven’t had to define well because we only care about using them to measure humans. And we just don’t have to worry about what it means to be “thinking but not feeling” or “talking but not thinking” or “intelligent but not sentient”.
Although I suppose they all come up when we are deciding whether it’s okay to kill/abort/enslave/eat something.
I think something like an ant colony or a bee hive can exhibit a kind of collective intelligence that is not anchored in the meaning of abstract concepts, and presumably does involve a centered, sentient experience like we have.
why do these frogs suck at jumping?
tl;dr
lol doesn’t matter
they’ve got armor and poison
look at these dorks
2 Likes
what is it with the frogs
frogs: doesn’t matter had sex?
1 Like
This guy just got fired
Spinlaunch is planning to launch smaller satellites into low earth orbit using kinetic energy. It would use a fraction of the fuel of rockets and cost much less. They are already launching suborbital payloads successfully, with video of that here: https://www.spinlaunch.com/suborbital
<3 I love mad scientist solutions. Though I don’t really expect to work anywhere near as advertised. I doubt the engineers think so either. Makes me want to play kerbal.
Hopefully there is some use for that thing though. Like as a moon catapult, or as one part of a launch technology, or for amazon package delivery, or for sending nukes even faster.
Oh, just got to the end of the video. The guy is trying to monetize on telling Reddit nerds that they are wrong. I love it.
Matt Parker addressed using a trebuchet to get things into orbit in one of his recent podcasts.