3 ms·
I think I see what you mean and I should have been more precise. I'll rephrase: most LLMs based on transformer architectures do not rely on the simplistic state
by semking 2y ago
I think I see what you mean and I should have been more precise. I'll rephrase: most LLMs based on transformer architectures do not rely on the simplistic state transition logic of Markov chains... BUT, while they can use alternative mechanisms such as relative weight (attention), embeddings... they STILL require the pre-existence of exact words, synonyms, or similar/related words in the datasets, correct?
- david-gpu 2y agoThey do not require instances of exact words in the training set, either. They can cope with e.g. typos or novel portmanteaus of your own making. Can you think of an example of the limitations you are thinking of?
- semking 2y agoSorry, I just saw your question. I'm still confused by the comment system here! To answer your question, I've noticed LLMs overfit the patterns in their training data. The performance was poor when asked to "reason" in ways that diverge from what they "learned". Things less frequently mentioned in the datasets were clearly weighted as less important. And the opposite is also true. I've seen bias being amplified. For some reason, I've seen words such as "elevate" or "landscape" being overused. And obviously, I've seen so many factual errors. Search for niche topics/individuals that were not covered in the training data and you'll see incredible hallucinations.