I asked ChatGPT 'how certain you are to your answer to my previous question' after I asked it some question, and it replied me the following: 'As an AI developed by OpenAI, I provide information based on a large dataset of text from the internet up until my last training cut-off in September 2021. The definition of a blender as an electrically powered kitchen appliance used for mixing, chopping, or liquefying foods is a standard definition found in many dictionaries and widely accepted. Therefore, I can say with high confidence that this definition is accurate. However, please note that in different contexts, the word "blender" might have different meanings. For example, Blender is also the name of a popular open-source 3D computer graphics software.'

Now, from what I understand, Bayesian people would not accept LLMs have a proper way to quantify uncertainty, since they are just trained to maximize likelihood (i.e., trained to do MLE). According to them a model with true uncertainty quantification should be something like Bayesian neural networks where the distribution of weights is given. But by the same logic, I can argue neither are we humans capable of truly quantifying uncertainty, since we don't have a distribution of how likely a neural is activated in our brain (there is only one instantiation of it).

QUESTION: why do probabilistic/Bayesian machine learning researchers/practiationers claim deep learning models do not have truth uncertainty quantification when even humans do not have the same kind of uncertainty quantification they deem to be genuine, while as humans we clearly have a sense of uncertainty? Is our sense of uncertainty not true then by the same token?

  • $\begingroup$ I find hard to follow and find the question in this. Could you better elaborate what is your question? $\endgroup$
    – Ciodar
    May 24 at 16:49
  • $\begingroup$ I am trying to argue against Bayesian way of thinking uncertainty quantification, because our brain doesn't work that way. $\endgroup$
    – Sam
    May 24 at 19:19
  • $\begingroup$ by 'doesn't work that way' I meant when we try to estimate the uncertainty of certain things, there is only one reality in which the neurons in our brain is activated in a fixed way. This is analogous to there is no distribution of parameters in LLMs like in Bayesian NN $\endgroup$
    – Sam
    May 24 at 19:28
  • $\begingroup$ Edit your post to formulate an actual question. I don't see any question. A question is something like "Why is the sky blue?" $\endgroup$
    – nbro
    May 25 at 22:09
  • $\begingroup$ Who are the "Bayesian people"? $\endgroup$
    – Kostya
    May 27 at 9:30