Okay I must come back to this. * Rudolf Adamkovič <rud...@adamkovic.org> [2025-02-28 16:00]: > Jean Louis <bugs@gnu.support> writes: > > > [...] that answer is not innovative. > > Thank you for the explanation. I think I now see what you mean. > > In that case, we are simply talking about two different topics. > > I stated that large language models can solve never-seen instances of a > problem. That is, the model has seen the problem in the training set > but not the particular instance in the test set.
It's actually always so; your test lacks facts. If you input "I am an Emacs enthusiast" as a system message into the model, it can infer numerous related concepts without needing explicit question-answer pairs in its dataset for probabilistic resolution of correct answers. This is fundamentally how Large Language Models (LLMs) function. You can't take it as a test that if question wasn't in the dataset that Large Language Model (LLM) invented something new... come on. Imagine, if we would have all the questions people are asking inside of the dataset? That would not make any sense. Let me ask DeepSeek about it: LLMs and Novelty: LLMs like GPT-4 do not "invent" in the human sense. They generate responses based on patterns and information in their training data. If a question or concept wasn't explicitly in the dataset, the model can still produce a novel response by combining and rephrasing learned information in new ways. This might appear inventive, but it’s fundamentally an interpolation or extrapolation of existing knowledge, not true creation. So the Large Language Model (LLM) said basically same thing as me, same thought with different words. I am speaking about what I learned about vectors and tensors, and I am beginner. > I use the terms (1) problem, (2) problem instance, (3) training set, > (4) test set, per their standard, precise, mathematical definitions. > I said nothing more and nothing less, and with the standard > terminology, my statement holds true, as can be trivially > demonstrated, and as I explained. Hmm. That paragraph I am not getting fully. Training set is just that, tuning the vectors in probabilistic manners. If you bring probability accuracy very high by training it, model starts becoming apparently "smarter" while in reality it just matches probability of what could it be better than other models less "trained". Another statement by DeepSeek: Limitations of LLMs: LLMs don’t have true understanding, intent, or creativity. They rely on statistical correlations in their training data. While they can generate outputs that seem novel or insightful, these are still rooted in the data they were trained on, not in independent thought or invention. > Now to your views. As for "innovation" and "true invention", I have > nothing to say to that end. In fact, I have no idea what what those > terms mean, if we were to measure. I have the same issue with the > linked GNU article on "artificial intelligence". That a computer can be programmed to find by probability something that human can't find that fast, it is great tool, not necessarily innovation. What is innovation? * Overview of noun innovation The noun innovation has 3 senses (first 2 from tagged texts) 1. (3) invention, innovation -- (a creation (a new device or process) resulting from study and experimentation) 2. (1) invention, innovation, excogitation, conception, design -- (the creation of something in the mind) 3. initiation, founding, foundation, institution, origination, creation, innovation, introduction, instauration -- (the act of starting something for the first time; introducing something new; "she looked forward to her initiation as an adult"; "the foundation of a new scientific society") (note that single word has multiple definitions depending of the context) Model doesn't study anything. It doesn't see. It doesn't evaluate. We are human who anthropomorphize the Large Language Model (LLM) because of our fascination. The Hells Angels often refer to their motorcycles as 'babies', which can be seen as a form of personification or anthropomorphism, attributing human-like qualities to inanimate objects such as bikes. We have got new words in the area of artificial intelligence. So we have got a new context. to train a child: discipline, train, check, condition -- (develop (children's) behavior by instruction and practice; especially to teach self-control; but to train a model is in different context, it is not related to human, so dictionaries must get updated: - to train the LLM means: To train an LLM (Large Language Model) means to expose it to large amounts of data so it can learn patterns, relationships, and structures in the text to generate coherent and contextually relevant responses. - to learn in the LLM context means: To learn, in the context of an LLM, means to adjust its internal parameters (weights) based on patterns and relationships in the training data, enabling it to predict and generate text that aligns with the input it receives. "Adjusting internal parameters"" does not equate to "learning" as in normal context; rather, it's merely a modification of numerical values influencing the probability of outcomes. > My *narrow* statement was true, and you are taking a *wider* view. > > Thank you again for taking the time to explain your view to me! But all that wasn't important to me. I wish to find out anything truly innovative that was invented by some "AI" Large Language Model (LLM). I can't find it on Internet. -- Jean Louis --- via emacs-tangents mailing list (https://lists.gnu.org/mailman/listinfo/emacs-tangents)
