This is both upsetting and sad.
What’s sad is that I spend about $200/month on professional therapy, which is on the low end. Not everyone has those resources. So I understand where they’re coming from.
What’s upsetting is that this user TRUSTED ChatGPT to follow through on the advice without critically challenging it.
Even more upsetting is this user admitted to their mistake. I guarantee you that there are thousands like OP who wasn’t brave enough to admit it, and are probably to this day, still using ChatGPT as a support system.
Source: https://www.reddit.com/r/ChatGPT/comments/1k1st3q/i_feel_so_betrayed_a_warning/
People really misunderstand what LLMs (Large Language Models) are. That last word is key: they’re models. They take in reams of text from all across the web and make a model of what a conversation looks like (or what code looks like, etc.). When you ask it a question, it gives you a response that looks right based on what it took in.
Looking at how they do with math questions makes it click for a lot of people. You can ask an LLM for a mathematical proof, and it will give you one. If the equation you asked it about is commonly found online, it might be right because its database/model has that exact thing multiple times, so it can just regurgitate it. But if not, it’s going to give you a proof that looks like the right kind of thing, but it’s very unlikely to be correct. It doesn’t understand math - it doesn’t understand anything - it just uses it’s model to give you something that looks like the right kind of response.
If you take the above paragraph and replace the math stuff with therapy stuff, it’s exactly the same (except therapy is less exacting than math, so it’s less clear that the answers are wrong).
Oh and since they don’t actually understand anything (they’re just software), they don’t know if something is a joke unless it’s labeled as one. So when a redditor made a joke about using glue in pizza sauce to help it stick to the pizza, and that comment got a giant amount of upvotes, the LLMs took that to mean that’s a useful thing to incorporate into responses about making pizza, which is why that viral response happened.
I have a compsci background and I’ve been following language models since the days of the original GPT and BERT. Still, the weird and distinct behavior of LLMs hasn’t really clicked for me until recently when I really thought about what “model” meant, as you described. It’s simulating what a conversation with another person might look like structurally, and it can do so with impressive detail. But there is no depth to it, so logic and fact-checking are completely foreign concepts in this realm.
When looking at it this way, it also suddenly becomes very clear why people frustratedly telling LLMs things such as “that didn’t work, fix it” is so unproductive and meaningless: what would follow that kind of prompt in a human-to-human conversation? Structurally, an answer that looks very similar! Therefore the LLM will once more produce a structurally similar answer, but there is literally no reason why it would be any more “correct” than the prior output.
That’s right, you have it exactly. When the prompt is that the prior output is wrong, the program is supposed to apologize and reprocess with a different output, but it uses the same process.
LLMs are our attempts to teach computers how to speak, like how we would a baby.
They can string together sentences, but the cognitive ability just isn’t there yet.
You wouldn’t have a baby as your therapist either. But you can use it as creative input to inspire new ideas for yourself.
We’d be better off talking about AI if no one used words like intelligence, cognition, think, understand, know, learn, etc. They don’t do any of those things.
And unless I’m quite mistaken, any specific correction one might possibly contribute to the LLM project in question (i.e. software hallucination), is generally-speaking, roundly & enthusiastically embraced and even celebrated by the LLM, then immediately and completely ignored.
I.e., they’re not programmed to listen to our feedback in a meaningful, educational way, only to keep munching on the databases their doggie-daddies have sicced them upon.
EDIT: that cynicism / critique aside, ChatGPT in particular has been hugely useful in my language-learning, and there’s no question to me that it’s improved a lot, just across the last few months. FWIW
Right, because “listen” and “educational” don’t apply to a software application like this. It has a model based on processing a truly huge amount of text. Your lone correctional prompt might tweak that model, but only slightly.
And sure, as a tool, LLMs can be very useful. I managed a software engineering organization for an aerospace company for a lot of years, and I made a number of constraints about how the LLM could be used (the company had one inside the firewall, so there weren’t IP issues), but I for sure encouraged it to be used. Essentially I was concerned about our software engineers using it in any way where they counted on it to be correct, because it often wouldn’t be. But it was great for things like suggesting test cases to test a piece of code.