Lots of examples I've seen have phrases like "see table below". Of course there's no table and it's hard to imagine how there could be.
But GPT is trained on internet content and the internet is full of terrible writing that never gets to the point. I doubt there's any way to know how much is "not actually understanding the subject matter" vs. "learning bad writing from bad writers". I'm inclined to believe the majority is the former but there's got to be a little of the latter sprinkled in.
One thing I learned was it has detailed knowledge of the world of Avatar: The Last Airbender, seemingly through fanfics. It was fun having it to teach me the lost arts of pizzabending ("form your hands into the shape of a letter 'P'" and so on, and needing to practice by juggling rubber pizzas) and beetlebending ("always remember that to beetle bend it helps to like beetles," my wise uncle suggested). Each of these tended to precipitate a narrative collapse.
The writing style was surprisingly homogeneous, and it reminded me of young adult novels. It would definitely be interesting to see it with other writing styles, beyond the occasional old poetry.
> The man walks away and starts undressing. You shrug and keep following him. Soon, you find yourself naked.
So this morning he heard about an animal, it was kind of a lion. But with bat's ears, it lives in Africa. It looks like it's a rock, but it's actually not, it's rock shaped but has tiny legs. And it's gray and hard. Its face... It doesn't really have a face. It lives up in trees where it eats bamboo and apples. It has these huge fangs like sabertooth tigers, you know?
It's glorious.
My smallest kid has a habit of telling stories about himself that actually come from whatever he heard recently, e.g. "once I was Godzilla..", or claims about things in reality that come from stories or misunderstandings all mixed up "did you know, there are three pigs, but they are not pigs, they are wolves and a hunter came and killed them but they weren't wolves they were dragons..."
It's actually very GPT-3-ish now that I think of it.
Maybe. Right now this reads like a glorified shopping list. It's coherent, but actually sounding human also requires a theory of mind.
E.g. I explain here why it's possible for written statements to be objectively insightful, informative, interesting, or funny, but objectively in a way that's relational to other information or beliefs. The implication being that statements are only going to seem subjectively funny or insightful (or whatever) to others who have that knowledge or those beliefs, which means that you can't reliably create those subjective experiences in a reader without having some sort of theory of mind for them.
I guess you can create content that's funny or insightful relative to that content itself, but that's not especially useful. It's entertaining at the time, but the experience is more like seeing a movie that you laugh a lot during but then leave and are kind of like what was the point? It's an empty experience because it wasn't transformative.
I definitely don't think it's impossible, but I also don't think it's a matter of just adding a couple more if-else statements.
https://alexkrupp.typepad.com/sensemaking/2010/06/how-writin...
I'm going to call this goalpost shifting. This article is better writing than some % of humans, theory of mind or otherwise. The AI has comfortably surpassed Timecube-level writing and is entering the pool of passes-for-a-human.
'Sounds human' is a spectrum that starts with the mentally ill and goes up to the best writers in human history.
That's completely fair. On the other hand, without a theory of mind it can't really educate or inspire people, the only thing it can do is maybe trick them about the authorship of something. But once people learn the techniques for identifying this kind of writing, it can't even do that anymore. To me this is like the front end of something, but it still needs a back end.
Don't get me wrong, it's super cool research and seems like a huge step forward, and I'm excited to see where it goes. But I also don't see this AI running a successful presidential campaign or whatever, at least within the next couple years.
And that got me thinking about what I could do with this thing, whether I should, what I wanted to try out...
So the BS random ideas were still inspiring a bit.
I wouldn't agree with that, either. How often have we heard of someone gaining useful insights by considering ideas that were misapplied or just plain wrong? Entire branches of physics have evolved that way. As far as successful presidential campaigns are concerned... well, let's not even go there.
If there's such a thing as a 'theory of mind', it applies to the reader, not the writer.
For example, I delayed in writing this comment because the cat was on my lap, and I couldn't fit the laptop and the cat both. You get that. I know you do, even if you don't own a cat, and even if you're reading this on a phone or a desktop.
GPT-3 does not understand about the cat. To GPT-3, they're just words and phrases that occur in the vicinity of each other with certain probability distributions. As a result, it can't write something and know that there's something there in your mind for it to connect to.
Cyc would handle the bit about the cat differently. It would ask "Did you mean cat-the-mammal, cat-the-Caterpillar-stock-symbol, or cat-the-ethernet-cable-classification?" It has categories, and some notion of which words fit in each category, but it still doesn't understand what's going on.
But you the human understand, because you have a lap, and you've at least seen a cat on a lap.
You really think GPT-3 never came across a comment about a cat in lap? 50% of all the pictures on the internet are cats sitting on people. GPT-3 doesn't need to understand it to echo this common knowledge.
Airplanes don't look like birds at all but they do fly.
I would like to see a GPT model where training data is weighted by credibility / authority (e.g. using Pagerank).
What if the ultimate theory of mind turns out to be that consciousness is an illusion and nothing separates us from a sufficiently sophisticated markov process.
Conscious experience would still exist (see cogito ergo sum, Chalmers, etc). If we were to be shown we're just Markov processes, that wouldn't disprove the existence of conscious experience. Just like confabulation, a misleading experience is still an experience.
What it would disprove is any sense of agency.
I certainly do, don't you? When I read a blog post and it's full of poorly-integrated buzzwords that make it seem like it was churned out by a non-English speaker being paid very poorly per word, I stop reading and move on.
I recently read a few pages of a book someone had recommended to me and stopped reading because of the writing style.
Heck, you can read a few pages of, say, a Dan Brown novel, and based on the writing style might choose not to read it, since the style tells you a lot about the kind of book it is.
I guessed it was fake before getting to the end, not from the content, but from the fact that all the sentences are roughly the same length and follow the same basic grammatical patterns. Real people purposely mix up their sentence structure in order to keep the readers engaged, whereas this wasn't doing that at all. Still very impressive though; if not for the fact that the post was about computer generated content I probably wouldn't have noticed.