sillymath2

From a practical point of view, if the prior probability of the disease is low but the disease is mortal and there is a $0.1 test and $0.2 pill that cure the disease without having negative effects, I think that you should take the test and if positive take the pill. Knowing the posterior probability is not enough to decide what to do, you have to know what are the tools and trade-offs.

Perhaps there is a model of the world that is 90% compatible with the most usual phrases used in the training dataset. That model could be 10% of the time absolutely wrong about the real state of our world. Since the way in which LLM construct a model of the world is not ruled based, it is difficult to design an adversarial attack to detect the pitfalls of the model.

We know that LLMs are trained just to predict the next token and that the model of the world that we assume she has is based on that. This intuitively suggests that such an adversarial attack could be devised and exploited, but at the same time it indicates that such attack is very difficult to design.

Because LLMs model are weak at math problems and at logic reasoning we could look for those fields that reflect the weakness of LLMs, so we should base the attack on concepts that are an expansion of those known weaknesses.

I haven't found any deep thought about probabilistic analysis and hypothesis of how the weakness of LLMs are multiplied when applied to more elaborated concepts, but we all know that the problem of logic is that one weak or false argument break the strength of a chain of logical reasoning, to summarize: we should look for those strands that when loosed break concepts apart.

Bayes' theorem is not enough, I think you have to weight how expensive is the test, what are the consequences of false positives and true positives. For example a low price test able to detect a genetic illness that can be save life to patients with a low costs and great confidence treatment, and almost zero false positive rate seems to me a test that should be applied to anyone at risk.

I am sure that if we are trying to answer posts with title like "How do you approach a problem you are not sure has a solution?" then we have a big problem of run round in circles. The first step to solve a problem is to define well the problem, the title omitted the anxiety/paralysis bit, so I bited this. Anyway, to add something constructive, I recommend "How to Solve It" by Polya. From a reviewer: In lucid and appealing prose, Polya reveals how the mathematical method of demonstrating a proof or finding an unknown can be of help in attacking any problem that can be “reasoned” out (don't know with or without anxiety).

Skimming the paper: “What are rich areas of exploration in the field of NLP that could lead to a PhD thesis and cover a space that is not within the purview of LLMs.”

With all due respect to the authors, I am not interested in skimming the 27 pages. What I am interested at just now is to know whether NLP can be used to reduce LLMs costs?, bonus if you can use NLP to enhance the LLM ability for using tools and/or reducing hallucinations. Life is short and hard. Why am I posting this comment here?, simply because I am expecting someone to provide hints about how to reduce cost of (or way of enhancing) LLMs by using other NLP techniques or tools.

Perhaps fuzzy logic is like naive Bayes: When you don't have enough data to train a DL model or to compute joint probabilities with little error, you have to use simple tools with less power. Fuzzy logic is a system of simple (thumb rules) that in some cases can provide good results.

If, in a concrete case, fuzzy logic can provide you with some insight about a system then it is a useful tool for that case. For example, it is a tool for exploration about new hypothesis in case of having few data about a complex system. But when a lot of data is available then statistical methods perform better.

So I suppose fuzzy logic is a tool to use when having few data about a complex system, nothing more and nothing less.