LLMs that use Chain of Thought sequences have been demonstrated to misrepresent their own reasoning [1]. The CoT sequence is another dimension for hallucination.
So, I would say that an LLM capable of explaining its reasoning doesn't guarantee that the reasoning is grounded in logic or some absolute ground truth.
I do think it's interesting that LLMs demonstrate the same fallibility of low quality human experts (i.e. confident bullshitting), which is the whole point of the OP course.
I love the goal of the course: get the audience thinking more critically, both about the output of LLMs and the content of the course. It's a humanities course, not a technical one.
(Good) Humanities courses invite the students to question/argue the value and validity of course content itself. The point isn't to impart some absolute truth on the student - it's to set the student up to practice defining truth and communicating/arguing their definition to other people.
First, thank you for the link about CoT misrepresentation. I've written a fair bit about this on Bluesky etc but I don't think much if any of that made it into the course yet. We should add this to lesson 6, "They're Not Doing That!"
Your point about humanities courses is just right and encapsulates what we are trying to do. If someone takes the course and engages in the dialectical process and decides we are much too skeptical, great! If they decide we aren't skeptical enough, also great. As we say in the instructor guide:
"We view this as a course in the humanities, because it is a course about what it means to be human in a world where LLMs are becoming ubiquitous, and it is a course about how to live and thrive in such a world. This is not a how-to course for using generative AI. It's a when-to course, and perhaps more importantly a why-not-to course.
"We think that the way to teach these lessons is through a dialectical approach.
"Students have a first-hand appreciation for the power of AI chatbots; they use them daily.
"Students also carry a lot of anxiety. Many students feel conflicted about using AI in their schoolwork. Their teachers have probably scolded them about doing so, or prohibited it entirely. Some students have an intuition that these machines don't have the integrity of human writers.
"Our aim is to provide a framework in which students can explore the benefits and the harms of ChatGPT and other LLM assistants. We want to help them grapple with the contradictions inherent in this new technology, and allow them to forge their own understanding of what it means to be a student, a thinker, and a scholar in a generative AI world."
I'll give it a read. I must admit, the more I learn about the inner workings of LLM's the more I see them as simply the sum of their parts and nothing more. The rest is just anthropomorphism and marketing.
Whenever I see someone confidently making a comparison between LLMs and people, I assume they are unserious individuals more interested in maintaining hype around technology than they are in actually discussing what it does.
Someone saying "they feel" something is not a confident remark.
Also, there's plenty of neuroscience that is produced by very serious researchers that have no problems making comparisons between human brain function and statistical models.
Current LLMs are not the end-all of LLMs, and chain of thought frontier models are not the end-all of AI.
I’d be wary of confidently claiming what AI can and can’t do, at the risk of looking foolish in a decade, or a year, or at the pace things are moving, even a month.
That's entirely true. We've tried hard to stick with general principles that we don't think will readily be overturned. But doubtless we've been too assertive for some people's taste and doubtless we'll be wrong in places. Hence the choice to develop not a static book but rather living document that will evolve with time. The field is developing too fast for anything else.
I think that’s entirely the problem. You’re making linear predictions of the capabilities of non-linear processes. Eventually the predictions and the reality will diverge.
Every time someone claimed “emerging” behavior in LLMs it was exactly that. I can probably count more than 100 of these cases, many unpublished, but surely it is easy to find evidence by now.
Not quite, but it was the closest pithy quote I could think of to convey the point that things can be false for a long time before they are suddenly true without warning.
The post seems to be talking about the current capabilities of large language models. We can certainly talk about what they can or cannot do as of today, as that is pretty much evidence based.
The ground truth is chopped off into tokens and statistically evaluated. It is of course just a soup of ground truth that can freely be used in more or less twisted ways that have nothing to do or are tangent to the ground truth. While I enjoy playing with LLMs I don't believe they have any intrinsic intelligence to them and they're quite far from being intelligent in the same sense that autonomous agents such as us humans are.
Any all of the tricks getting tacked on are overfitting to the test sets. It's all the tactics we have right now and they do provide assistance in a wide variety of economically valuable tasks with the only signs of stopping or slowing down is data curation efforts
I've read that paper. The strong claim, confidently made in the OP is (verbatim) "they don’t engage in logical reasoning.".
Does this paper show that LLMs "don't engage in logical reasoning"?
To me the paper seems to mostly show that LLMs with CoT prompts (multiple generations out of date) are vulnerable to sycophancy and suggestion -- if you tell the LLM "I think the answer is X" it will try too hard to rationalize for X even if X is false -- but that's a much weaker claim than "they don't engage in logical reasoning". Humans (sycophants) do that sort of thing also, it doesn't mean they "don't engage in logical reasoning".
Try running some of the examples from the paper on a more up-to-date model (e.g. o1 with reasoning turned on) it will happily overcome the biasing features.
I think you'll find that humans have also demonstrated that they will misrepresent their own reasoning.
That does not mean that they cannot reason.
In fact, to come up with a reasonable explanation of behaviour, accurate or not, requires reasoning as I understand it to be. LLMs seem to be quite good at rationalising which is essentially a logic puzzle trying to manufacture the missing piece between facts that have been established and the conclusion that they want.
So, I would say that an LLM capable of explaining its reasoning doesn't guarantee that the reasoning is grounded in logic or some absolute ground truth.
I do think it's interesting that LLMs demonstrate the same fallibility of low quality human experts (i.e. confident bullshitting), which is the whole point of the OP course.
I love the goal of the course: get the audience thinking more critically, both about the output of LLMs and the content of the course. It's a humanities course, not a technical one.
(Good) Humanities courses invite the students to question/argue the value and validity of course content itself. The point isn't to impart some absolute truth on the student - it's to set the student up to practice defining truth and communicating/arguing their definition to other people.
[1] https://arxiv.org/abs/2305.04388