Hey again! First of all, thank you for continuing to engage with me in good faith and for your detailed replies. We may differ in our opinions on the topic but I’m glad that we are able to have a constructive and friendly discussion nonetheless :)
I agree with you that LLMs are bad at providing citations. Similarly they are bad at providing urls, id numbers, titles, and many other things that require high accuracy memorization. I don’t necessarily agree that this is a definite proof of their incapability to understand.
In my view, LLMs are always in an “exam mode”. That is to say, due to the way they are trained, they have to provide answers even if they don’t know them. This is similar to how students act when they are taking an exam - they make up facts not because they’re incapable of understanding the question, but because it’s more beneficial for them to provide a partially wrong answer than no answer at all.
I’m also not taking a definitive position on whether or not LLMs have capability to understand (IMO that’s pure semantics). I am pushing back against the recently widespread idea that they provably don’t. I think LLMs have some tasks that they are very capable at and some that they are not. It’s disingenuous and possibly even dangerous to downplay a powerful technology under a pretense that it doesn’t fit some very narrow and subjective definition of a word.
And this is unfortunately what I often see here, on other lemmy instances, and on reddit - people not only redefining what “understand”, “reason”, or “think” means so that generative AI falls outside of it, but then using this self-proclaimed classification to argue that they aren’t capable of something else entirely. A car doesn’t lose its ability to move if I classify it as a type of chair. A bomb doesn’t stop being dangerous if I redefine what it means to explode.
Do you think an LLM understands the idea of truth?
I don’t think it’s impossible. You can give ChatGPT a true statement, instruct it to lie to you about it, and it will do it. You can then ask it to point out which part of its statement was a lie, and it will do it. You can interrogate it in numerous ways that don’t require exact memorization of niche subjects and it will generally produce an output that, to me, is consistent with the idea that it understands what truth is.
Let me also ask you a counter question: do you think a flat-earther understands the idea of truth? After all, they will blatantly hallucinate incorrect information about the Earth’s shape and related topics. They might even tell you internally inconsistent statements or change their mind upon further questioning. And yet I don’t think this proves that they have no understanding about what truth is, they just don’t recognize some facts as true.
In my sense of “understanding” it’s actually knowing the content and context of something, being able to actually subject it to analysis and explain it accurately and completely.
This is something that sufficiently large LLMs like ChatGPT can do pretty much as well as non-expert people on a given topic. Sometimes better.
This definition is also very knowledge dependent. You can find a lot of people that would not meet this criteria, especially if the subject they’d have to explain is arbitrary and not up to them.
Can you prove otherwise?
You can ask it to write a poem or a song on some random esoteric topic. You can ask it to play DnD with you. You can instruct it to write something more concisely, or more verbosely. You can tell it to write in specific tone. You can ask follow-up questions and receive answers. This is not something that I would expect of a system fundamentally incapable of any understanding whatsoever.
But let me reverse this question. Can you prove that humans are capable of understanding? What test can you posit that every English-speaking human would pass and every LLM would fail, that would prove that LLMs are not capable of understanding while humans are?
If I were to have a discussion with a person responding to me like ChatGPT does, I would not dare suggest that they don’t understand the conversation, much less that they are incapable of understanding anything whatsoever.
What is making you believe that LLMs don’t understand the patterns? What’s your idea of “understanding” here?
As I understand it, most LLM are almost literally the Chinese rooms thought experiment.
Chinese room is not what you think it is.
Searle’s argument is that a computer program cannot ever understand anything, even if it’s a 1:1 simulation of an actual human brain with all capabilities of one. He argues that understanding and consciousness are not emergent properties of a sufficiently intelligent system, but are instead inherent properties of biological brains.
“Brain is magic” basically.
GPT3 is pretty bad at it compared to alternatives (although it’s hard to compete with calculators on that field), but if it was just repeating after the training dataset it would be way worse. From the study I’ve linked in my other comment (https://arxiv.org/pdf/2005.14165.pdf):
On addition and subtraction, GPT-3 displays strong proficiency when the number of digits is small, achieving 100% accuracy on 2 digit addition, 98.9% at 2 digit subtraction, 80.2% at 3 digit addition, and 94.2% at 3-digit subtraction. Performance decreases as the number of digits increases, but GPT-3 still achieves 25-26% accuracy on four digit operations and 9-10% accuracy on five digit operations, suggesting at least some capacity to generalize to larger numbers of digits.
To spot-check whether the model is simply memorizing specific arithmetic problems, we took the 3-digit arithmetic problems in our test set and searched for them in our training data in both the forms “<NUM1> + <NUM2> =” and “<NUM1> plus <NUM2>”. Out of 2,000 addition problems we found only 17 matches (0.8%) and out of 2,000 subtraction problems we found only 2 matches (0.1%), suggesting that only a trivial fraction of the correct answers could have been memorized. In addition, inspection of incorrect answers reveals that the model often makes mistakes such as not carrying a “1”, suggesting it is actually attempting to perform the relevant computation rather than memorizing a table.
In my comment I’ve been referencing https://arxiv.org/pdf/2005.14165.pdf, specifically section 3.9.1 where they summarize results of the arithmetic tasks.
That’s not entirely true.
LLMs are trained to predict next word given context, yes. But in order to do that, they develop internal model that minimizes error across wide range of contexts - and emergent feature of this process is that the model DOES perform more than pure compression of the training data.
For example, GPT-3 is able to calculate addition and subtraction problems that didn’t appear in the training dataset. This would suggest that the model learned how to perform addition and subtraction, likely because it was easier or more efficient than storing all of the examples from the training data separately.
This is a simple to measure example, but it’s enough to suggests that LLMs are able to extrapolate from the training data and perform more than just stitch relevant parts of the dataset together.
This is partially true and partially not. It’s true that LLMs can’t learn anything wildly novel, because they are not flexible enough for this. But they can process new information, in fact they do it all the time. You can produce conversations that no one had before, and yet LLMs like ChatGPT will respond to it appropriately. This is more than just shape matching.
In fact, there are techniques like Few-Shot Learning and Chain of Thought that rely on the LLMs’ ability to learn from context and revise its own answers.
IMO citation problem is not testing capability to understand. It’s testing knowledge, memorization, and ability to rate its own confidence. Keep in mind that ChatGPT and most other LLMs will tell you when they perform web searches - if they don’t then they’re likely working off context alone. Enabling web search would greatly increase the accuracy of LLM’s answers.
Unlike LLMs we have somewhat robust ability to rate how confident we are about our recollections, but even in humans memory can be unreliable and fail silently. I’ve had plenty of conversations where I argue with someone about something that one of us remembers happening and the other one is certain didn’t happen - or happened differently. Without lies or misunderstandings, two people who had at some point memorized the same thing can later on confidently disagree on the details. Human brains are not databases and they will occasionally mangle memories or invent concepts that don’t exist.
And even that is completely skipping over people with mental disorders that affect their thinking patterns. Is someone with psychosis incapable of understanding anything because they hold firm beliefs on things that cannot be traced to any source? Are people with frontal lobe damage who develop intense confabulations incapable of understanding? How about compulsive liars? Are you willing to label a person or an entire demographic as incapable of understanding if they fail your citation test?
There are techniques like Chain of Thought that make LLMs think before generating response. Those systems will be able to tell you how they arrived at the conclusion.
But humans are also fairly prone to rationalization after the fact. There was a famous experiment on people who had to have functional hemispherectomy for medical reasons, where the left hemisphere makes up an explanation for right hemisphere’s choices despite not knowing the true reason:
“Each hemisphere was presented a picture that related to one of four pictures placed in front of the split-brain subject. The left and the right hemispheres easily picked the correct card. The left hand pointed to the right hemisphere’s choice and the right hand to the left hemisphere’s choice. We then asked the left hemisphere, the only one that can talk, why the left hand was pointing to the object. It did not know, because the decision to point was made in the right hemisphere. Yet it quickly made up an explanation. We dubbed this creative, narrative talent the interpreter mechanism.”