* Eli Zaretskii <e...@gnu.org> [2025-03-08 13:09]: > > Date: Sat, 8 Mar 2025 00:51:23 +0300 > > From: Jean Louis <bugs@gnu.support> > > Cc: emacs-tangents@gnu.org > >
> When Large Language Model (LLM) analyzes 130,000 documents it > means it is getting the information, thus uses that information. If > you call it learning or absorbing, whatever, it took the information, > it did not invent new information. > That is inaccurate. Machine learning does produce information not > explicitly included in the learning set. > You mentioned "new original insight in the data" -- that is matter > of viewpoint and consciousness of the observer. > No, it isn't. The model data includes data generated by analyzing the > learning set, and that fits the definition of "new insight". Let us then ask Microsoft Phi-4, fully free "smart" one: https://pageassist.xyz/share/cm80d751k00718c5qil4pwyo0 And look what it says, surprisingly matches my view point: In essence, whether machine learning produces "new information" can depend on how one defines newness and insight. From a strict data-centric view, all information is derived from the input data. However, from a functional or application-oriented perspective, the insights and predictions generated by the model can be seen as new information, as they provide value and understanding beyond the explicit content of the training set. I didn't ask the LLM before I made my own opinion. You said machine learning produce information not included in the learning it. And I asked for example. I can understand people's opinions, but it is not an opinion I am looking for, I am genuinely interested to find such examples, which are beyond personal biased perspective of how it may look like. Deepseek said: Both are right, but they are focusing on different aspects of the process. Louis is emphasizing that LLMs do not create entirely new information, while Eli is highlighting that LLMs can generate new insights by analyzing and synthesizing the data in ways that reveal latent patterns or relationships. The disagreement stems from differing definitions of "new information" and "new insight." But neither Deepseek or Phi-4, I consider authoritative. Providing some rules to the model, like rules related to proteins and asking it to find out the combination beneficial to human is not something I consider "innovation" as the rules have already been in existence. Discovering that there is a fox at the end of the den is not an innovation, there is a rule that says about probability to find it. We have mathematical rules, we do not need to know the result of a complex mathematical operation, but the result cannot be called an innovation. If you have examples, some links, send me. > > The exercises gives me quite a feeling that model "knows" only that > > what it was given to know. > > If this is your conclusion, it is inaccurate. Since the learning > process generates data not explicitly in the learning set, the model > "knows" something beyond what the learning set includes. Especially > if parts of the learning process are supervised (as they many times > are nowadays). That is exactly what I would like it. I just don't see it is so. And when asking LLM, I get different answer. If I try to query search engines for it, I cannot get conclusive answers. If you just say it is wrong, or it is right, without giving me example, I am not getting any new insight. There is no reasoning, and no reference in saying that something is wrong or right. Look at the answer here: https://pageassist.xyz/share/cm80dqliq00728c5qjbrur23x Excerpt: -------- However, it's important to note that the generated data is still constrained by the patterns and knowledge captured from the training data. The model does not "create" information in the sense of forming new concepts that were entirely absent from the data it was trained on. Instead, it recombines and extends existing patterns in innovative ways. In summary, while the specific text generated by an LLM was not explicitly in the training dataset, the ability to generate such text is a result of the model's exposure to and understanding of the data it was trained on. The generated content is a reflection of the patterns and structures learned during the training process. Those responses above are LLM responses, though appearing educative, I do not consider them authoritative. I am really interested to see if LLM could eventually find something "beyond" the knowledge that it combines. > > It is far from any thinking, it is computation based on given goals. By all > > means it is similar to computer program just with the NLP. At least Phi-4 validated my statement above: https://pageassist.xyz/share/cm80dw4fn00738c5q8paavc70 > Given our basic inability to know and understand what exactly happens > during human "thinking", the above assertion is meaningless. I understand you wish to impose your opinion in authoritative way, but that is not how I learn. I need to learn by defining meanings and understanding what occurs in the real life. When one part of equation is unknown but other known, we may have ways to find the result. - We have mathematical understanding of how LLM works; That it is based on mathematical computations, that is strong and established fact. Is it? - We do not have full understanding how human mind works; though we have lot of understanding and we can deal with it. There are many similarities to computing, but what is really known as it cannot be proven is that it doesn't work by computing predictability with vectors like in the context of the LLM. A model doesn't "think". It computes results mathematically. Providing bunch of text of "reasoning" under the badge of "thinking" doesn't make it thinking. It is deceptional. Reasoning process of the LLM is generating text which in turn is making better probabilities for the final answer. It is an algorithm, not thinking. > > The surprise that there is some invention by AI must be based on lack of > > consciousness of the observer on what is actually going on. > > Given our basic inability to understand what constitutes > "consciousness" and how it is translated into the physical and > physiological processes in the human brain, the above assertion is > also meaningless. Let's be practical. A child may look at objects on the table not knowing that there is hot coffee in the cup, and harm itself. We all know what is lack of consciousness. I guess you got well impressed by answers given by the generative text that you now thinkg they are "thinking". Keep using it. I suggest you use it locally on very low end GPU like 4 GB VRAM, beautiful exercise, as that is good way to understand that it doesn't think, doesn't invent anything new, and that it creates an illusion to do so, like a good magician. Such experience you cannot that easy on better GPUs with 12 or 24 GB VRAM, as models are better, and not showing you the reality. Is "Instruction" way of "learning"? I think yes. Instruction: this is how you do it? Duplicating instruction a human could find out why. Maybe not. But telling to human why, would improve his instruction implementation. It seems that human knowledge agrees with the above: https://pageassist.xyz/share/cm80emxgc00748c5q251d4ar4 But there is just no LLM that could follow the instruction and learn out of it. Let's say, there is all the knowledge and functionality, like how to create datasets, how to train the model, how to access all the packages to do so, but the model simply can't train itself and become better. What is point of having all the knowledge and not being able to actually do it? Maybe "thinking" isn't there. > > And I wish I could find somewhere truly inovative, but I don't. > > What is "innovation" for this purpose? Is generating a few paragraphs > of text about some subject "innovation"? No, that is what I sam saying. > Is theorem proof "innovation"? It could be, if there is some new mathematical theorem generated by the LLM model that doesn't exist in our society as such, then I would call that innovation derived from the known rules. I tried finding such theorem, if you have one, let me know. I asked my computer and answer after Internet search is: https://pageassist.xyz/share/cm80h2atk00758c5qbp8904v5 (negative) Then I asked Deepseek: As of now, there is no widely recognized mathematical theorem that has been invented entirely by a Large Language Model (LLM) like GPT-3 or similar AI systems. LLMs are primarily designed to generate human-like text based on patterns and information in the data they were trained on. While they can assist in generating ideas, proofs, or exploring mathematical concepts, they do not possess the capability to independently create new, original, and rigorously validated mathematical theorems. That is why I am really interested to see something new invented beyond calculations of probabilities. > Is a decision which alternative out of several to prefer > "innovation"? Not to me. > All of the above are creative activities that humans are engaged in > quite a lot, where LLM can help to a great extent or even replace > humans. There is nothing that LLM creates. It computes. I understand the excitement as it is like audience sitting and watching magician take rabbit out of the hat. But it is deceptive, rabbit was somewhere there but hidden. :-) JL --- via emacs-tangents mailing list (https://lists.gnu.org/mailman/listinfo/emacs-tangents)
