I wonder if that is the origin of the worst insult - like fighting words in South Africa - of the Afrikaans phrase with the same meaning, “jou ma se poes”, effectively “your mother’s vagina”. We have a huge mix of languages and cultures that all contributed to South Africa, including Arabic.
Obviously being an extreme contrarian, I have issued this insult, on occasion: “jou ouma se vis poes” in response, which is reaching nuclear levels of insult in this country. Literally “your grandmother’s fishy smelling vagina.”
It’s possible, although it does seem like the sort of phrase that’s obviously insulting enough for multiple cultures to independently come up with. Hebrew speakers actually say the phrase in Arabic (although the relevant Arab world for vagina is a loan words on its own, and the Arabic word for “your mom” is almost identical to the Hebrew word - but when using the exclamation, you use the Arab pronunciation).
I appreciate the comprehensive and erudite analysis of the issues as you see them. I agree with much of it, but I have to take issue with the quoted part and some of the conclusions that follow.
I agree that absorbing the structure and logic of natural language is an important contributor to an LLM’s capabilities, but it surely cannot be the whole story. For example, as @Babale said, it cannot explain multi-modal LLMs based on essentially the same technology, like those capable of powerful image processing. Even with more conventional text-based LLMs, it’s hard to see how language competency translates into mathematical competency, especially at higher levels. The two skills are so different that many humans, including quite intelligent ones, might possess strong language competency but be very bad at math, or the other way around.
The article below is part of a series about LLMs and is a fair and well-balanced treatment of the subject of emergent properties. It cites a famous paper on the subject from 2022, then cites an attempted refutation. The refutation is, at best, controversial, and centers on the supposition that when emergent properties appear suddently and unexpectedly at particular levels of scale, that this is merely an artifact of the evaluation methodologies. The authors claim that the artifact, or “mirage”, occurs because “almost-right” LLM responses are dismissed as wrong and therefore the existence of the latent capabilities isn’t observed or acknowledged until larger ANN scales and parameter numbers are achieved. I think in most cases this is simply not true (the capabilities just are not there in any meaningful way) but there it is – a competing hypothesis for your interest.
In short, whether true emergence exists in LLMs is the subject of debate, but I’m firmly in the “yes” camp. As the article says, quoting UC Berkeley professor Jacob Steinhardt, “emergence” can be defined as “qualitative changes that arise from quantitative increases in scale.”
I’d certainly like it to be that way. But, likewise, I can see the risk for a “yes man” machine to help compulsive people to amplify and extremify their own psychoses. I can envision the ability for the owner of the AI to push users towards and away from particular things through subtle influence campaigns.
I was reading a paper a bit ago by a researcher who had determined that there was no way (in his testing) to adjust things like a “Facebook algorithm” that didn’t push users towards one form of extremism - even if it was just to focus on a particular celebrity, excessively. That’s likely due to the ease of building social connections and the strength of some individuals’ ability to make use of those connections. It’s sort of (in my mind) a gravitational outcome of giving people access to the entire world of social opportunity.
It’s entirely plausible that there may be similar things here.
Or, inversely, that something like a personal, rational and dispassionate companion is just what’s needed when you’ve also got the social pressures imposed by being connected to the whole world…
I’ve not been particularly sensitive to the risk of “algorithms” because I avoid social media like Facebook and Twitter/X like the plague. And although “algorithm” is not a technically appropriate word to apply to something like ChatGPT, I get your point that the same effect might apply.
Thomas Friedman recently made the point, in the context of American divisiveness, that if you as a liberal and your neighbour as a conservative compared your phones (or laptops, or whatever you use to get your fix of social media) the pervasiveness of algorithm-driven selective information will result in your devices displaying completely different information. In effect, you two live in completely different worlds, informed by completely different truths.
But I think this is getting somewhat off topic. Maybe I’m being naive – and I’m sure that someone will rush in to confirm it – but I don’t believe that AI developers like OpenAI have the same maliciously self-serving intent as, say, Zuckerberg and Facebook. To at least some significant extent, OpenAI (and others doing similar work) are working on valuable technology development; Zuckerberg is just working on getting even richer than he already is. Musk I think is just working on restoring the glories of the Third Reich.
I’m not sure why the handling of new data types changes anything, other than the ability to handle new data types. It’s not like they are any closer to actual reasoning. LLMs have gotten better at simulating intelligence over the years, but they haven’t gotten any closer to actual intelligence.
Of course there are many humans who will pretend to know anything and everything. But by and large, people will freely tell you “I’m not sure” or “that’s not my area” or “talk to Dr. XYZ”, or “I don’t have time for this”.
A human’s lack of an answer can be informative in itself. If 10 different people tell you that this question makes no sense and you’re barking up the wrong tree, that can tell you something! But you can’t get uncertainty from an LLM. They pretend to know everything, all the time, and are not at all embarrassed to feign expertise in a topic that it screwed up horribly just 30 seconds ago.
The real issue here isn’t ChatGPT breaking up marriages, it’s people turning to an unknown third party to win a marital argument, although no doubt the faux expertise of the LLM is enticing more people to engage in that sort of behavior.
Although, ChatGPT probably aggravates this not only by its air of certainty, but by letting its user dump long bullet points of newly constructed grievances with possibly new terminology. In my workplace, people do this even with the purest and most helpful intentions, hitting me with a 3-page wall of AI slop that they haven’t even read themselves. It’s infuriating. I cannot imagine someone brewing up an academic-quality bullet list of 20 things that are wrong with me, throwing it in my face, and declaring victory while I’m not even halfway done reading it.
I would guess a problem with AI therapists is that they would be too supportive. They are probably great at being a supportive listener and telling you what you want to hear. That’s what lots of people want. Someone to listen to their problems and make them feel good about themselves. An AI therapist is probably going be like, “Hmm… Wow.. It does sound like you are the best person ever and everyone around you is a huge jerk. Since they are the problem, not you, you need to dump them and just hang around friends who like you, like the all new ChattyBesties from ChatGPT for the low price of $10/month. Sign up now and get a free trial week to hang with your new best friends!”
In the real world, problems often come about because of problematic behaviors and aspects of the person themselves. A real counselor will be trying to hone in on those things to help the person realize them and give them ways to address them. It doesn’t seem like an AI therapist will really challenge the person in a meaningful way to get them to change. Also, a real therapist will be encouraging a person with marriage issues to go to joint marriage counseling. Chatting with AI is probably good for blowing off some steam about work drama, but not very good about making real improvements to your self or relationships.
You’re not? “Data types” are conventionally regarded as things like integers, strings, or floating point numbers. They’ve always been handled within the same basic computational protocols. But the difference between responding to a text query versus asking “here is a picture of a new electronic gadget – tell me what it is and tell me how to use it” – and people have indeed done this – has to be regarded as a qualitative new difference in the level of pretty broad LLM understanding, not just some consequence of the ability to process natural language.
So a more granular level of object recognition, then using the text label of that object identification as information in the LLM, is some sort of miracle or something? You’re losing me. If you wish to grant understanding to an LLM, can you at least define “understanding” first. I feel like I’m looking at a sophisticated (and impressive) computer model and you’re seeing WALL-E.
First of all, I’m using the word “understanding” in the context of results, not impuging what the inner processes necessarily look like.
Secondly, while you might technically argue that you’re not entirely wrong about “different data types”, it profoundly trivializes what’s really happening here. Which is that words and digitized image elements are both being tokenized and processed by the same transformer architecture. Easy-peasy, right? No. It speaks to the incredible generality and power of predictive modeling to operate on the same kind of token embeddings independently of what media they represent. This is not only fundamentally different from what are normally regarded as computational “data types”, but it’s a profoundly important advance in AI as demonstrated not only by the success of traditional LLMs, but by their extension to multi-media applications.
I think it’s a mistake to compare ChatGPT to a fake psychic because LLMs don’t intentionally deceive. As I understand it, they don’t intentionally do anything. They just provide a reasonable sounding statistically generated result based on the data they were trained on. They aren’t reasoning or fact checking the responses.
Which is another way of saying that they can handle additional data types.
Look, there is no one in this or any other thread on this board who doesn’t find LLMs impressive. Hell, I find pretty much the entire field of data science (AI, if you will) fascinating, which is why the past couple of decades of my career were focused in that area. It still doesn’t mean that it has any sort of reasoning ability, actual understanding as I understand (I know, right?) it to be defined, or anything remotely approaching intelligence. It simulates those things, sometimes better or worse than other times, but simulated nonetheless. I was going to compare an LLM interaction as something like a child talking to their favorite doll, a schizophrenic talking to the voices in their head, or someone having a conversation in their dreams. The problem with that is the conversation partners of all of the above actually do reason and understand, as they are simply extensions of the human on the other side. So an LLM is a lower level of intelligence than little Susie’s imaginary friend inside of her Cabbage Patch doll.
The other problem is when we pretend that LLMs are smart because they have access to a giant corpus of data, which some might call knowledge (not me, but some). Note to Babale and anyone else that thinks that we’ve suddenly crossed some imaginary bridge, I don’t care how large that corpus is 3 years from now, nor how many data types it contains, it doesn’t change my arguments.
Knowledge - facts, information, and skills acquired by a person through experience or education; the theoretical or practical understanding of a subject.
I’m perfectly fine with removing the word “person” from that definition, but it still doesn’t change the fact that LLMs don’t have knowledge.
I look at "2 + 2 = " and I know the response is “4”. An LLM predicts that “4” is the most likely response. I know, it predicts. You could try and throw a curveball and say what about "22 + 22 = " and when I respond 44, you tell me that you mean in base 8. My response is that I know the answer is still 44, just in base 8 (18 + 18 = 36 in base 10, which is 44 in base 8). I have gained knowledge of numeric bases and I have knowledge of simple arithmetic. My answers are all factual, not predictions or probabilities. If you say you meant the 22s as string literals, my answer is 2222, because I also happen to have knowledge of concatenation. Again, facts not predictions. So, while you can keep reframing the question, I’ll continue to use my knowledge to answer them concisely and accurately, assuming that I have knowledge in the appropriate area. I remember back when the "r"s in strawberry was throwing off most LLMs. While that’s patchable, that’s what was required to fix it. With actual knowledge, a human could also get the answer wrong. The difference is that if you simply said “you’re wrong”, just about any human would look at the word more carefully, a letter at a time, and then give the correct answer.
Let me ask you one question on LLMs, if you ask your LLM of the day to give you the answer to “2 + 2 =” do you think that it knows the answer factually?
And one question on other deep learning models, if you upload a picture of a cat, and it responds by identifying it as a cat, do you think the model knows what a cat is factually?
I think it’s safe to say that if you can boil an entire branch of philosophy (Epistemology, in this case) to a 20 word sentence, you’re likely oversimplifying.
