Post

Exactly this. And in fact, many models *have* been created this way, and decades of prior machine learning systems *were* created this way, consensually. These things are choices! People are STILL working this way! It’s not just that a better world is possible — it already EXISTS. We can nurture it.

my understanding is that they performed very poorly without being able to make sense of our modern vernacular and why they "had" to crawl all of reddit to figure out how modern humans talk (????? reddit? really?)

News flash: Using LLMs to store “knowledge” will never work, and those systems will always hallucinate. The sales fraud and stolen content is a policy choice based on ego and arrogance.

why are you telling someone who works in machine learning that machine learning models cannot perfectly capture the sum total of human knowledge and work with 100% accuracy and 100% recall en.wikipedia.org/wiki/All_mod...

of course models will tell you incorrect things. that's their inherent nature. data scientists are aiming for `most probably true`, above some threshold, answered near instantly, not `definitively true`, which is intractably hard right now - and maybe forever.

meteorological models are a great example of talking about inaccurate forecasts. everyone has a visceral experience with 'it's supposed to be nice for <outdoor event>' and it ends up being not nice. we still don't expect the news to drop the weather report. we just know that it isn't omniscient.

separating ML and LLMs is like separating squares and rectangles. you stated your problem was with hallucinations - false positives - which is true for all ML. projecting into the squares/rectangles analogy, you're expressing displeasure with right angles.

I don't know that I'm being very clear and as such I beg your pardon. I do not take any issue with your ultimate hatred and frustration with LLMs - there's a lot to hate about them right now.

I just think the "it hallucinates" argument is tired; false positives are problems, we try to minimize them, but they *will* exist in a statistical model. (unless your statistical model is ultimately 'return no confidence in existence of anything', I suppose