The post shows limited understanding of what LLMs are. Ok, so Quora has a bad model that generates "Why doesn't Donald Trump shave his mustache?" type questions. That's Quora's problem.
But in general, a frontier LLMs like the ones trained by Open AI would not make such silly mistakes in their output. Its "fancy autocomplete" is trained on millions of documents that include detailed description of Trump's face and the presence of the word "Trump" in the context window would activate countless weights that relate to "orange hair", "pussy grabbing", "bribes" etc. Each word or sequence of tokens can be understood as a vector in a space of say, 12288 different dimensions, and for each round in the algorithm the entire context is refined so that each of those dimensions would access finer and finer details of the subject matter present in the training data.
So the model really does "know" Trump does not have a mustache. What it lacks (or has very limited ability to do) is rational exploration of the information it holds. It's like an idiot savant with the mental ability of a 4 year old that has somehow memorized the entire information that exists on the internet.
> But in general, a frontier LLMs like the ones trained by Open AI would not make such silly mistakes in their output.
This is beginning to wear a bit thin. Every public application of AI produces obvious nonsense, and this doesn’t appear to be changing. The inevitable defence is “oh, they used GPT3.5 instead of GPT4/oh, they used GPT4, but not the version that was only publicly available for three weeks in 2023/oh, they used precisely the right GPT4 but, when sacrificing the goat while petitioning Roko’s Basilisk for useful output, they used the wrong sort of knife”.
Apparently, literally no-one can use these things properly, at least for publicly visible applications. Either that, or they’re just a bit shit.
But in general, a frontier LLMs like the ones trained by Open AI would not make such silly mistakes in their output. Its "fancy autocomplete" is trained on millions of documents that include detailed description of Trump's face and the presence of the word "Trump" in the context window would activate countless weights that relate to "orange hair", "pussy grabbing", "bribes" etc. Each word or sequence of tokens can be understood as a vector in a space of say, 12288 different dimensions, and for each round in the algorithm the entire context is refined so that each of those dimensions would access finer and finer details of the subject matter present in the training data.
So the model really does "know" Trump does not have a mustache. What it lacks (or has very limited ability to do) is rational exploration of the information it holds. It's like an idiot savant with the mental ability of a 4 year old that has somehow memorized the entire information that exists on the internet.