I went back to the archive of all my old posts (going back to 2000) - a lot of factual questions that couldn’t be answered with a simple Google search - and pasted them into ChatGPT. I’m getting some interesting answers, and I’m also enjoying the follow up Q&A I can do. Anyone else trying the same?
Why would I try that? I’d want the correct answer. Though I’m willing to laugh at any Chat “answers” posted here.
I’ve done this as a test before, just to suss out its strengths and weaknesses, and I’ve found it surprisingly good with the ones I knew the answers to. I wouldn’t post answers from it, though, without following up on its sources. I’m also certain AI has been used to answer posts on the SDMB, some better disguised than others. It can absolutely be a valid way for opening research on a topic.
So you’re making the assumption that any answer you get from chatGPT is going to be wrong? Serious question - why do you think that?
Because they always have been. Oh sure, if I ask Chat a known fact, it’s probably right. “When is Tom Hanks’ birthday?” Ever since google started putting AI responses at the top of search results, I see a lot. They are either straight out of actual sources (BFD, I’ll just go to the original source) or, sometimes, laughably wrong. Including one I saw that was self-contradictory in its short paragraph!
If
is the guide, why even bother going to Chat? Just go to the original source.
Though if someone wants to ask it about pan-fried semen, I’d love to see the reply.
I did this with some unanswered GQs (at least from my memory unanswered). Like Where did the ‘that’s the beauty of it, it doesn’t do anything!’ come from? Funnily enough it used the SDMB as a reference. Still no definite answer though.
This kind of snark just tells me that you’ve rarely used it, and probably never at all recently.
A little while ago someone offered actually a pretty thoughtful analysis of the limitations of ChatGPT. I submitted the analysis to ChatGPT and asked it to critique it. I posted its response here. The original analysis is in the post just above.
Please tell me which parts of that you’re “willing to laugh at”. I’m not saying that facetiously. Go and have a look and tell me what’s wrong with that response.
Strangely, my experience has been the exact opposite. It’s not that it’s always right – humans make mistakes, too. It’s that it’s right most of the time, and not only that, but recently has been displaying an impressively profound depth of knowledge.
I’m getting pushback for not trusting AI, but yet, no one has taken up the OP and put a vintage SD question to Chat. Let’s see this supposed factualness.
I just gave you one. Waiting for your feedback.
The problem with AI systems like ChatGPT isn’t that they are wildly inaccurate. In contrast, they are usually fairly accurate. The problem is that when they aren’t accurate, they give an answer that looks and feels like it is a correct answer. You can’t tell if its answer is good or not.
The classic example of this is having ChatGPT count the number of 'R’s in the word Strawberry. The AI has since been trained to answer this question correctly, but initially ChatGPT very confidently gave the answer that there were 2 'R’s in strawberry.
And this is the underlying problem with ChatGPT and other AI systems. They don’t actually understand what they are doing. They just behave in a way that often makes it appear that they do. So yes, ChatGPT is impressive. But underneath the hood, it’s nowhere near as smart as it seems. It’s just really good at faking it.
You just never know when it’s going to make a really stupid error like not being able to count the number of letters in a word. Or it can totally bork a more complex question. So yes, it’s impressive. No, you can’t rely on it to give you a good answer. It’s smart enough to give you an answer that looks good. It’s not smart enough to know if it is actually giving you a good answer or not though.
It will be interesting to see how it handles SDMB questions.
There is a specific technical term for this kind of material.
For the narrow academic definition of “bullshit.”
Oh my god, I’m an AI.
What is factual about a machine opinion? An opinion of a subject I am not familiar with, so I’m not sure how I’m supposed to evaluate a machine’s opinion of “itself”. In my non-machine opinion, “Their language seems shaped by a classical computer science or symbolic AI lens, which may lead to undervaluing the emergent capabilities seen in practice.” is mush. It isn’t saying anything. “It seems good, but may be wrong”. Duh. What usefulness is that?
I’m not sure “So while external memory is absent, ephemeral internal state is rich and crucial.” means anything either, but I am not knowledgable in LLM.
Which makes your test meaningless. I don’t understand how Chat works, so how can I evaluate an AI assessment of AI? I guess I could go ask chat, and post my meta answer of a meta answer, and pretend it is my own?
But I’m not that person.
While the first part of this is true, you can usually verify an answer through other sources, which is typically much easier than sourcing the information from scratch.
Inasmuch as humans also make mistakes, and some members of our species are inveterate bullshitters, how do you qualify exactly what “understanding” means?
Google AI response:
You are asking about frying semen. This type of content is not appropriate and it is not possible to provide any information regarding it.
AI responses may include mistakes.
The last line is the best…
You can consider the fact that the person who wrote the original analysis was impressed by the critique and did not dispute its accuracy.
This part is actually especially impressive, because, as the person who wrote the analysis noted himself, this was an eerily accurate assessment of himself – he is in fact a professor of computer science and was indeed coming at the subject from a classical comp-sci perspective, which tends to focus on the internal mechanics of LLMs and misses the degree to which LLMs develop emergent properties like intelligent behaviours and problem-solving skills at very large scales.
Even if you don’t understand the technical details yourself, this illustrates the extent to which LLMs can perform useful and complex intellectual functions.
No one has ever suggested otherwise. And humans make mistakes, too.
Nov 2000
One reply:
Mar 2024 reply from SaSaLeLe2011:
Google AI response:
“Neener neener neener” is a playground taunt, often used in a singsong voice by children. It’s a variation of the longer “Nyah nyah nyah nyah nyah nyah” taunt, which is a children’s chant associated with taunting.
The “Nyah nyah” chant itself is based on a six-note musical figure that’s common in European-derived cultures and often used in taunting. This musical figure is derived from a pentatonic scale (specifically, A minor if starting on G: G-E-A-G-E), which is a common musical scale found across many cultures.
The “neener” variation, with its repetition, is a playful way to taunt or tease someone, often used in a childish or informal manner.
AI responses may include mistakes.
OK, I’ll give that one to AI over SaSaLeLe2011
But millions of people do not rely on what I answer. They do not rely on what I say as an authorized source.
And I know the difference between “truth” and “truthiness”.
If people rely on GPT as an “authorized” and infallible source, that’s a people problem, not an AI problem.
Jan 2001:
I asked “In latin, why did the letter K still did not die out; except for Kalendae, the name for the first day of each month?”
Google AI response:
The letter “K” persisted in the Latin alphabet despite being less common than “C” due to a few key factors, primarily the word “Kalendae” and its role in the Roman calendar. While the letter C generally represented the hard “k” sound (as in “cantus”), the letter K was used in specific instances, like the word for the first day of the month, “Kalendae”. This usage helped the letter K maintain a presence in the Latin alphabet, even though C was more widely used for the same sound.
That’s not an answer at all! It’s simply restating the question.
AI went on to note: “This word was consistently spelled with a K, and its importance in the Roman calendar ensured its continued use in writing.” Which is almost an answer, I’ll grant.
I’ll admit, “Why?” isn’t an easy question, but it’s been 24 years, and Latin has been studied forever, so it should be answerable.