It’s definitely in the nature of the current-gen systems. It’s possible to imagine other ways to train chatbots that wouldn’t favor bullshitting in the same way, though. The way they currently are isn’t a way they necessarily have to be.
Anyway, it’s very clearly a case where it would have been better to cut out the middleman and just share the prompt, which was my point in sharing it.
It’s possible to imagine other ways to train chatbots that wouldn’t favor bullshitting in the same way, though
How? The way they work is that they predict the next word (token). They don’t know anything. So, when someone says “what’s the book that features an X doing Y and has a Z that looks like a Q” the process is basically “if this were a forum post, what’s the next word that’s likely to appear, what’s the one after that, what’s the one after that”.
If the scene being described is something from Harry Potter that has been discussed endlessly on the web, it’s likely to predict next words that end up being helpful. If the scenario being described has never been described before, there’s no way for it to say “aha, this book doesn’t exist!” Instead it just invents a plausible conversation between two people about a fictional book.
Now, you could design a filter that sits in front of the LLM that makes things better. It could be designed to recognize quotes, and if it spots a quote, it can “search the web” (or really look through a big index) to see if those words in that sequence have actually been seen before. If not, it could reject the entire response as being a “hallucination”.
But, fundamentally, this isn’t a problem you can fix through training, because of how LLMs work.
You can’t fix it with a pure LLM, no. But very few of the major models are pure LLMs anymore. They’ve got LLMs built in to them as subsystems, but they’ve got control systems layered on top of those to adjust their behaviors. And those don’t have to be trained purely with next word prediction.
The problem with that is that, it seems like it will become an unwieldy piece of spaghetti code, even assuming the code wasn’t written by LLMs.
An LLM is a crazily complex thing but at least it’s text in, text out, one input, one output.
As soon as you start bolting more and more things onto it to deal with special conditions, you have to maintain all those things, and understand how they interact.
I’m not talking about bolting things together at random. I’m talking about things like mixture of experts or chain of reasoning, which apply structures to coax more rational behavior out of the systems. You do have to generate new bodies of training data for them, but there are ways of doing that algorithmically. This allows them to move away from the strict “regurgitate the Internet” models of the previous versions. There isn’t, currently, this kind of structure to drive the systems towards an honest assessment of their own knowledge and limits, but it’s basically inevitable that such a system will be developed, because it’s a valuable kind of behavior to be able to favor.
It’s definitely in the nature of the current-gen systems. It’s possible to imagine other ways to train chatbots that wouldn’t favor bullshitting in the same way, though. The way they currently are isn’t a way they necessarily have to be.
Anyway, it’s very clearly a case where it would have been better to cut out the middleman and just share the prompt, which was my point in sharing it.
How? The way they work is that they predict the next word (token). They don’t know anything. So, when someone says “what’s the book that features an X doing Y and has a Z that looks like a Q” the process is basically “if this were a forum post, what’s the next word that’s likely to appear, what’s the one after that, what’s the one after that”.
If the scene being described is something from Harry Potter that has been discussed endlessly on the web, it’s likely to predict next words that end up being helpful. If the scenario being described has never been described before, there’s no way for it to say “aha, this book doesn’t exist!” Instead it just invents a plausible conversation between two people about a fictional book.
Now, you could design a filter that sits in front of the LLM that makes things better. It could be designed to recognize quotes, and if it spots a quote, it can “search the web” (or really look through a big index) to see if those words in that sequence have actually been seen before. If not, it could reject the entire response as being a “hallucination”.
But, fundamentally, this isn’t a problem you can fix through training, because of how LLMs work.
You can’t fix it with a pure LLM, no. But very few of the major models are pure LLMs anymore. They’ve got LLMs built in to them as subsystems, but they’ve got control systems layered on top of those to adjust their behaviors. And those don’t have to be trained purely with next word prediction.
The problem with that is that, it seems like it will become an unwieldy piece of spaghetti code, even assuming the code wasn’t written by LLMs.
An LLM is a crazily complex thing but at least it’s text in, text out, one input, one output.
As soon as you start bolting more and more things onto it to deal with special conditions, you have to maintain all those things, and understand how they interact.
I’m not talking about bolting things together at random. I’m talking about things like mixture of experts or chain of reasoning, which apply structures to coax more rational behavior out of the systems. You do have to generate new bodies of training data for them, but there are ways of doing that algorithmically. This allows them to move away from the strict “regurgitate the Internet” models of the previous versions. There isn’t, currently, this kind of structure to drive the systems towards an honest assessment of their own knowledge and limits, but it’s basically inevitable that such a system will be developed, because it’s a valuable kind of behavior to be able to favor.