Post B6POFPe3F4LuwBZCKW by ariadne@social.treehouse.systems
(DIR) More posts by ariadne@social.treehouse.systems
(DIR) Post #B6PBcXUuugcEY3zXKS by ariadne@social.treehouse.systems
0 likes, 0 repeats
I dunno but chatgpt 5.5 instant sure is a piece of shit
(DIR) Post #B6PJcKAtFOAirS25Fw by ariadne@social.treehouse.systems
0 likes, 0 repeats
trillions of parameters and yet...
(DIR) Post #B6PJnLtrAmy3OhK8fY by ariadne@social.treehouse.systems
0 likes, 0 repeats
I don't know what they are doing over at OpenAI but it wasn't this bad last time I evaluated their product, seriously
(DIR) Post #B6PL0dTLFbOzcEosQC by ariadne@social.treehouse.systems
0 likes, 0 repeats
"bikanel desert the literal tidus because final fantasy x-2"
(DIR) Post #B6PL8mQdpoIvi6RLpA by ariadne@social.treehouse.systems
0 likes, 0 repeats
It's like I predicted that adding more and more and more parameters would make things worse not fix anythingoh wait I did š
(DIR) Post #B6PLRMpdEAnBpCzZiK by ariadne@social.treehouse.systems
0 likes, 0 repeats
glue on pizza isn't gross -- it's topping adhesion consistency
(DIR) Post #B6PLY53rhDMEjoqoee by xan@xantronix.social
0 likes, 0 repeats
@ariadne been too fucking addicted to LinkedIn (please euthanise me) so i keep trying to reach for the "Insightful" react here
(DIR) Post #B6PLuPdrrGVMwGL6v2 by ariadne@social.treehouse.systems
0 likes, 0 repeats
gpt 5.5 "thinking" (pretraining) does not have this weird behavior
(DIR) Post #B6PM4bPWxOYmKoUaJc by ariadne@social.treehouse.systems
0 likes, 0 repeats
this makes sense because the pretraining before generating a response should adjust enough weights for the final output passbut you're trading spatial dimension for compute time
(DIR) Post #B6PMGl3cut3JjE2jr6 by ariadne@social.treehouse.systems
0 likes, 0 repeats
cuil made the same mistake: pursuing more and more parametersand cuil's usefulness as a search engine got worse and worse
(DIR) Post #B6PMLzBzn4GJ10Rzsm by cr1901@mastodon.social
0 likes, 0 repeats
@ariadne ChatGPT is infected with Sin Toxin...
(DIR) Post #B6PMWKIu9ZexO3c492 by jannem@fosstodon.org
0 likes, 0 repeats
@ariadne I was misreading it as "curl" and the post still made a kind of horrified sense...
(DIR) Post #B6PMbIvsgBqSa8DXeq by ariadne@social.treehouse.systems
0 likes, 0 repeats
I will keep saying this for as long as I can: language models regardless of size are only good for turning unstructured text into structured data and vice versano amount of parameters will ever change this!
(DIR) Post #B6PNagOU46ab0hrc0W by ariadne@social.treehouse.systems
0 likes, 0 repeats
so, if adding more parameters increases the probability of errorand pretraining smooths the probability of errorthen to avoid errors during generation you must either pretrain at each turn (so-called "thinking" mode does this)or you must pursue more and more computationally intensive training runsbut really you need both to reduce error to something reasonable
(DIR) Post #B6PNdKWdpOYo3UwRUW by lanodan@queer.hacktivis.me
0 likes, 0 repeats
@ariadne As in horribly inefficient NLP?
(DIR) Post #B6PNtEIuj7RZOS6OzA by ariadne@social.treehouse.systems
0 likes, 0 repeats
at the same time there is a minimum amount of parameters you need to avoid high probabilities of error toobut that is much lower than what these commercial AI companies are pursuing
(DIR) Post #B6POFPe3F4LuwBZCKW by ariadne@social.treehouse.systems
0 likes, 0 repeats
(this is trivially demonstrated by the fact that many models exist which can generate plausible text given structured data as an input)
(DIR) Post #B6PPSHz98FQCtr8agK by dvshkn@social.treehouse.systems
0 likes, 0 repeats
@ariadne if I'm understanding your argument correctly, it sounds a lot like yann lecun's argument about being damned trying to steer the model through the answer space
(DIR) Post #B6PSRWandX8HmD4oka by ariadne@social.treehouse.systems
0 likes, 0 repeats
@lanodan it depends. if you want to handle random user input then the inefficiency might be worth it
(DIR) Post #B6PUubDZiawQQcCmjQ by ariadne@social.treehouse.systems
0 likes, 0 repeats
@evana I dunno, for whatever reason it is final fantasy x-2 content specifically that makes it freak out
(DIR) Post #B6Px87pIac5o2i5WwS by wronglang@bayes.club
0 likes, 0 repeats
@ariadne have they rediscovered overfitting?
(DIR) Post #B6Q6ISYJHiiCo4915E by be_far@social.treehouse.systems
0 likes, 0 repeats
@ariadne that makes sense. Iām convinced that innovations much earlier up the langchain are far more useful to humans