22
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 23 Dec 2024
22 points (100.0% liked)
TechTakes
2043 readers
77 users here now
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
founded 2 years ago
MODERATORS
so openai is claimed to be doing great on the FrontierMath dataset. I've already seen the usual sort of dipshits using this to pump ai on reddit, and here's a post that went to the frontpage on HN:
https://xenaproject.wordpress.com/2024/12/22/can-ai-do-maths-yet-thoughts-from-a-mathematician/
(tl;dr only a few problems from the dataset are public but if representative the problems are about 25% survivable by an undergrad; coincidentally this is the % openai says their models are completing.)
this post is by kevin buzzard. he has a let's say not easily beloved personality, but I don't think of him as credulous or grifty, and people in his area regard him as an excellent mathematician.
he points out but I think does not focus enough on how discrediting the secretive nature of the dataset is. the fact that you can't make it public is necessary to run such experiments in a scientifically reasonable way, but also makes it totally impossible to run the experiment in a scientifically reasonable way. an experiment which cannot be examined or reproduced is actually the opposite of science. it's pure grift fuel
If you release the test set, all models magically jump to 87.3% accuracy.