it has data not the training data, but it creates meta data from said training data much like a Spark Notes is different from the Novel, yet it contains words and references from the original even if it is greatly reduced / sublimated - ThreadSky

roland00.bsky.social • 7 days ago

it has data

not the training data, but it creates meta data from said training data

much like a Spark Notes is different from the Novel, yet it contains words and references from the original even if it is greatly reduced / sublimated

Comments

allfornaught.bsky.social•7 days ago

You would think that this must be true, but it isn't.

jalf.dk•7 days ago

And yet you can make it regurgitate large reams of its training data with the right prompt. We can quibble about how the training data is stored and encoded (certainly not in a structured and indexed plaintext format, but we cannot pretend that it isn’t there at all.

allfornaught.bsky.social•7 days ago

Perhaps there is text encoded into the static decimal weights somehow, but there is *no logic* to do this or retrieve it. This thing turns your words into numbers (tokens) runs them through a math function that "predicts" output tokens (one at a time, in order!) to create understanding and meaning.

allfornaught.bsky.social•7 days ago

This information is available. Or, you can go ask any AI about this and don't accept the first answer they give because they won't fully explain unless you persist.

allfornaught.bsky.social•7 days ago

allfornaught.bsky.social•7 days ago

These things don't even have any simple parameters you can toggle on their behavior, because the people that made them *don't know how they work*. The best they can do is to completely filter or block a response (obvious), or ask it nicely and repeatedly to behave (flimsy). That isn't a joke.

jalf.dk•7 days ago

Yes, the information is available. I'm a software engineer and CS graduate and have done my share of reading on how LLMs work (whereas you have asked chatgpt?)

I never claimed there was explicit "logic to retrieve it", just that it is encoded in the weights and shows up in generated text

allfornaught.bsky.social•7 days ago

Ok, then you know that the model is performing approximate reconstruction by statistical correlation (which can result in verbatim reconstruction of segments) and that no specific text is encoded anywhere at all. It doesn't really matter how I know this, because it is correct.

Comments

Posting Rules

Reply