Sent to me by a friend, don’t judge the misspelling of “strait” lol.
I was so confused for a moment. I was like, “What the hell does heterosexuality have to do with a noise?”
Yeah, I did the same. I love how it broke the AI even harder though.
A probable answer. That’s not a sensible question so a ridiculous answer is expected.
Strait is misspelled. Both straits and sounds are bodies of water so it’s a very sensible question. You might also ask what the difference between a cove and a bight is.
Didn’t make the connection. Very difficult for transformers since they do not listen to the words. They also don’t read the letters. So this is a ‘don’t use an ai for something it fundamentally cannot do’ example.
An error in a question should either result in correcting the question or indicating that the question doesn’t make sense.
Calling “straight” and “sound” homophones is a pure demonstration of the LLM’s ignorance. Maybe it got fooled by “straight” and “strait” being homophones and some how crossed wires, but that’s actually the point. It is ignorant, despite how “intelligent” it might sound.
I think you’re holding a fundamental misunderstanding of what today’s LLMs are.
An error in a question should either result in correcting the question
LLMs don’t have the ability to reason what you may have meant. The most they can do, if they are exposed to the right training, is understand something like “people that have used words or patterns similar to what you are using now meant X, Y, or Z, and of those the highest probability with the words you chose would be X.” This is exactly what it did.
or indicating that the question doesn’t make sense.
This would require the holy grail of AI which doesn’t exist yet: Artificial General Intelligence (AGI)
AGI is the ability to reason that humans (and some animals) can. None of today’s LLMs (Grok, Claude, ChatGPT, etc) are AGI. They are all the much more limited ANI (Artificial Narrow Intelligence). ANI can only work with whatever training data you give it, and even giant LLMs today are only a tiny fraction of what a process would need to have AGI. None of our current technology can take data we have today and build an AGI model. As the models scale the limits of LLMs start to fracture and fall apart.
I think you’re holding a fundamental misunderstanding of what today’s LLMs are.
I think you have severe misunderstanding of what this community is.
I…assumed it was a community to point where AI should would, but doesn’t. In the example we have here its not a flaw of the LLM, instead what is being asked of it is beyond its limits.
I don’t make fun of my screwdriver because its horrible and hammering in nails. If that’s what this community is for, then the mistake is mine to post in here. My apologies.
I…assumed it was a community to point where AI should would, but doesn’t.
…and that’s what’s happening in this case. You’re acting like it’s completely impossible for an LLM to go down a path where it handles that the question contained a misspelling because it isn’t AGI. In fact, to be useful an LLM should hand this better. It certainly shouldn’t start making up weird unrelated connections.
Also, it’s not impossible, and I guarantee that some LLMs would give a more appropriate answer. But this particular LLM couldn’t handle it, and went completely off the rails. Why are we not allowed to make fun of that? Why are you defending it from ridicule?
I don’t make fun of my screwdriver because its horrible and hammering in nails.
Holy strawman. We aren’t asking the LLM to be a different tool. The LLM is supposed to handle language, and a simple misspelling of a homophone caused it to misunderstand the question completely and sent it down a path of calling completely different words “homophones”. Yeah I wouldn’t make fun of my screwdriver for not being able to hammer nails, but I would be pretty annoyed if it constantly slipped due to slight imperfections in how screws were manufactured.
Maybe the LLM successfully predicted that this is a homophone issue, but screwed up correcting the word and then explained the non corrected word. Didn’t even occur to me. Fun.
10 years ago all of the search engines would have returned a site explaining the right answer, which I know because they always returned the right results even with mispellings.
Not only did it misunderstand the question, the answer was gibberish.‘Straight’ and ‘a sound’ are NOT homophones. ‘Strait’ and ‘straight’ are homophones.
Google still does? What is your problem. Both googles ai as well as their results are perfectly correct.
They might not listen to words but they can rhyme and compose songs just fine so they must have some sort of statistical correlation for the sounds of words being related.
Yeah. But there being boneappletea involved is not always expected. Barely any human would have performed better.
I disagree, I immediately knew what they were asking.
I expected some gay joke.
I thought I was looking into a mirror
