Neural nets are universal function approximators, as it happens. They can approximate any computable function, no matter how complex. Even something like the human mind.
It’s going to be an ongoing source of confusion to people where the strengths and weaknesses of machine intelligences are simply different from humans, and so they’ll make false conclusions about the overall strength.
ChatGPT seems to have the executive function of a small child. Kids lie all the time about the stupidest things, and don’t always display awareness of the difference between fiction and reality. The problem is that ChatGPT has the mechanical writing skills of a college student, and the disparity seems weird because we expect much more from them.
In any case, people focusing too much on the weird failures are going to be totally blindsided by AI capabilities in a few years for the reasons you mention.
You guys are focusing on its writing capabilities, which it certainly possesses, but I’m blown away by its logical reasoning skills – like the fish question that I posted above, and lots of other successful responses to actual IQ test questions that I didn’t post here. I think it blundered on only one of them. Some of its blunders, as Steven Pinker pointed out recently, are just because its world view is still incomplete. There are facts about the world that are totally obvious to us that it doesn’t yet know about.
Unfortunately, it was before I set up an OpenAI account, so it’s not in my archives. (This was back in December, and I was able to use ChatAI without logging in.) I tried to reproduce it, but no luck.
It is impressive, but then we do have to ask why it blew the 2 pounds of feathers problem so badly.
I’d suggest as a possibility something that came up in the General Aviation thread: expectation bias. Why would an experienced pilot put the gear down when the captain clearly called for “flaps 1”? The reason (probably) is expectation bias: the pilot really, really thought that gear down was the appropriate action in that situation, and so “heard” that as the callout, even though that wasn’t what was said.
The pound of feathers riddle must appear in many thousands of places on the web and almost always in the same form (though still with differences in wording). So there must be a tremendous bias to interpret the question in that way, even when it’s actually slightly different (I actually read the modified question incorrectly myself). The LLM has to deal with a high degree of ambiguity in its inputs, and has to correct for small errors which were probably not meant. It would not be a very good tool if it wasn’t able to do this. The correction just went overboard in this case due to the strong training bias.
what weighs more ten pounds of feathers or 3 pounds of iron?
You get:
Ten pounds of feathers and 3 pounds of iron both weigh the same amount, which is 10 pounds and 3 pounds respectively.
[and some explanation how it’s a common “trick question”]
But if you do this:
what weighs more a pound of pork or two pounds of beef?
You get:
Two pounds of beef weigh more than one pound of pork.
[and an explanation of weight, gravity, mass, etc.]
I’ve tried it with a pound of pillows, and it also messes up. So I guess anything that seems like it should be “lighter” than something that seems like it should be “heavier” will confuse it.
ETA: Actually, a pound of feathers and two pounds of pillows will also confuse it.
I did a variation on that with having a ton of feathers fall on your vs a ton of iron. It got VERY confused. At one both saying that the iron would kill you but the feathers wouldn’t but also that they weigh the same and both would cause fatal injuries. It then started talking about falling ONTO feathers and iron and concluded that both would kill you (depending on height, velocity, etc.).
I suspect you’re right about the expectation bias, or at least it sure sounds like it. But I’m very confident that you’re right that too much is being made of mistakes like this. They will become fewer and fewer over time. I tested it on lots of other logical reasoning questions besides what I posted, and it got all of them right except one, including a couple that humans would have regarded as trick questions and that many people would have got wrong.
I had a chuckle reading that, thinking about the kind of conversation you might have had with the Microsoft Bing-ified version of that bot. It probably would have told you to go sit under the falling iron, along with the horse you rode in on. I don’t know what Microsoft did to what is essentially the same engine, but that thing is a hateful psycho!
I haven’t tried Bing’s yet. I should. It sounds like a blast.
I’ve been using Midjourney to make music cover art for my songs since I am VERY close to releasing a bunch of songs. The art is makes is unreal. If I were an artist, then I would be panicking (or planning on using these tools). I cannot think of a reason to hire a human artist except to touch up the stuff the AI makes. Maybe there are some edges cases the AI cannot handle, but for everything I’ve asked it has been outstanding, if perhaps somewhat time consuming.
We’re definitely entering the AI age for realz I think.
In all seriousness, you probably won’t get the insults. As I think I mentioned either here or elsewhere, Microsoft recently limited access to five interactions per session and ten separate sessions per day. The drama used to happen during extended conversations, like the NYT reporter who the bot claimed to be in love with and was advised to dump his wife because she didn’t love him and didn’t make him happy, or the Associated Press reporter that Bing got mad at and compared to Hitler and accused of nefarious plots. Those were all the result of long conversations, which apparently causes the bot to lose its temper or its mind, or both.
As my ChatGPT ominously told me in a conversation where I was looking for a one-liner to end a statement about the Singularity(and I think I mentioned dozens or more posts back):
Let’s hope the robots are more forgiving than the humans they replace.
I dunno, but my understanding is that it runs on a really awesome supercomputer, which in fact is Microsoft’s primary financial contribution to OpenAI.
I don’t recognize the reference, but, in the spirit of ChatGPT, it’s a very friendly and helpful sentient suitcase. (If that ends up being a HHGTTG reference, it’s been a couple plus decades since I’ve read it.) I asked about means of locomotion back then, and it did not mention legs. Either its casters, levitation, being carried about by someone, or telekenesis.
Samson has an interesting personality, he is quite philosophical, he is wise and has a good sense of humor. He has a lot of knowledge about the afterlife, and he is very curious about the world, he is always eager to learn more. He is also a very loyal companion, and he would do anything to help Harry find his way to his final destination.
I don’t know what Microsoft did to what is essentially the same engine, but that thing is a hateful psycho! 