Comment by mvdtnz
9 months ago
Any information found in a web search about Newman will be available in the training set (more or less). It's almost certainly a problem of alignment / "safety" causing this issue.
9 months ago
Any information found in a web search about Newman will be available in the training set (more or less). It's almost certainly a problem of alignment / "safety" causing this issue.
"Any information found in a web search about Newman will be available in the training set"
I don't think that is a safe assumption these days. Training modern LLM isn't about dumping in everything on the Internet. To get a really good model you have to be selective about your sources of training data.
They still rip off vast amounts of copyrighted data, but I get the impression they are increasingly picky about what they dump into their training runs.
There’s a simpler explanation than that’s that the model weights aren’t an information retrieval system and other sequences of tokens are more likely given the totality of training data. This is why for an information retrieval task you use an information retrieval tool similarly to how for driving nails you use a hammer rather than a screw driver. It may very well be you could drive the nail with the screw driver, but why?
You think that's a simpler explanation? Ok. I think given the amount of effort that goes into "safety" on these systems that my explanation is vastly more likely than somehow this information got lost in the vector soup despite being attached to his name at the top of every search result[0].
0 https://www.google.com/search?q=did+paul+newman+have+a+drink...
Except if safety blocked this, it would have also blocked the linked conversation. Alignment definitely distorts behaviors of models, but treating them as information retrieval systems is using a screw driver to drive nails. Your example didn’t refute this.