4 ms·
I mean to be flippant so can `wc -l` but seriously it also matters how capable is the LLM
by ftxbro 3y ago
I mean to be flippant so can `wc -l` but seriously it also matters how capable is the LLM
- bestcoder69 3y agoSupposedly this is between gpt3.5 and 4, so pretty good.
- inhumantsar 3y agoAnthropic has hobbled it by going hard on the "harmless" aspect of alignment. We've been trialing it at work and found that it's easy to accidentally bully it into absolute uselessness with a couple wrong words in a follow-up prompt. It's good, don't get me wrong, but as an example: We brought the slack bot in. One guy said to it "explain DevOps to me like I'm five". The answer was decent but the language would have been beyond a 5 year old. So another person followed up with "that doesn't seem appropriate for a five year old". Claude agreed and promised to never try to explain advanced topics to children. Now we can't get ELI5s. Obviously that's a silly example but I'm not sure I'd trust its alignment with a live system the way I can with GPT3.5/4.
- version_five 3y agoThe headline is odd both because "ingesting" a few 100kb of text is a meaningless qualification as you say, and the great Gatsby is a particularly short book. Did they pick it on purpose because it's somehow near some length limit? Edit: that's exactly what theh did, gg has 72k tokens according to the article (wc says the project gutenberg text has 51257 words) and the new model can do up to 75k (actually 100k tokens see below).
- famouswaffles 3y agoThe model(s) can do up to 100k tokens not 75k. But yes they picked a book that would fit.
- version_five 3y agoAh, I mixed up tokens and words (100k tokens, 75k words), the ratio of which also seems to square with the difference between the quoted number of tokens in gg and the number of words I counted.