Month: September 2023

(In case anyone was wondering, Midjourney still has trouble counting fingers.) As noted in the last post, a paper has come out suggesting that Large Language Models have difficulty with reverse identity relationships. That is, if trained on A = B, they don’t necessarily exhibit knowledge of the reverse fact, B = A. Some folk…

A paper has come out suggesting that LLMs trained on A=B cannot infer that B=A. For anyone who has spent time with GPT4, this seems to be obviously false – this is well within the domain of what GPT4 understands. (The claim being made is actually slightly different, as will be discussed in Part 2.)…

Previously, I’ve posted several examples of where GPT4 has done better than some might expect in discussing various concepts. Here’s a conversation that exposes some of its cognitive limitations. I have seen several people on Reddit claiming that GPT4 can pass a Turing test. This is clearly not the case, if the Turing test is…