4 ms·
If LLMs work the way I understand them to, they are trained on large corpora which contain a statistical preponderance of statements which are consistent with t
by dekhn 2y ago
If LLMs work the way I understand them to, they are trained on large corpora which contain a statistical preponderance of statements which are consistent with the current scientific mainstream, so it may not be completely crazy to try doing this; you'd get the scientific mainstream explanation, and possibly a few alternative hypotheses if they had enough literature support.
- opello 2y ago> which are consistent with the current scientific mainstream This seems to require a high amount of curation of training inputs, but I haven't done real digging into it, just going off the more casual "all of stackoverflow" or "all of reddit" type comments frequently thrown around. But if there is such a curation I'd agree, I just don't think there is that curation.
- govg 2y agoWhy would it be consistent with the scientific mainstream? Unless there's evidence that scientific reports and material are specifically up-weighted during training and prioritised somehow, whatever an LLM says will be only consistent with its training material, which could have any proportion of fake articles, Reddit posts, Quora responses, encyclopedia pages and joke blogs
- satvikpendem 2y agoYou don't know what the training data is and what each category's weights are, so how can you assume that?
- dekhn 2y agoMost use common crawl, and most seem to upweight the scientific publication part of it (from what I gather speaking to folks who train LLM models).
- Jerrrry 2y agoYou are literally saying this novel thing is wrong because this thing that can't say novel things says the thing that's not novel is not novel.