BLUE
Llunamoth.bsky.social

- 애플 연구원들은 LLM에서 형식적 추론의 증거를 발견하지 못했으며, 정교한 패턴 매칭으로 설명 - 더 큰 문제나 무관한 변화에 따라 성능이 저하 됨 LLMs don’t do formal reasoning - and that is a HUGE problem garymarcus.substack.com/p/llms-dont-...

LLMs don’t do formal reasoning - and that is a HUGE problem
LLMs don’t do formal reasoning - and that is a HUGE problem

Important new study from Apple

0

A new study on LLMs by Apple: "We found no evidence of formal reasoning in language models," The behavior of LLMS "is better explained by sophisticated pattern matching" which the study found to be "so fragile, in fact, that [simply] changing names can alter results." arxiv.org/pdf/2410.05229

1
Ffmanu.bsky.social

Totalmente. Una cosa es la lógica formal y otra las piruetas que hacemos para justificar nuestras tesis. Lo que pasa es que nos gustaría que esas justificaciones fueran tan sólidas como la lógica formal y decimos que son "lógicas" por si cuela

0
FPfpaulus.bsky.social

"LLMs don't do formal reasoning" — adding an irrelevant piece of information worded in a way that seems relevant leads LLMs to break down and provide wrong answers. https://garymarcus.substack.com/p/llms-dont-do-formal-reasoning-and

0
Ccornelius75.bsky.social

This was not a good deal obviously. Jimmy Carter said it was only deal and people had to make it better. And little to nothing has been done by anyone to change it. You need 2 houses to brake a formal treaty. The president can’t just pull us out of NATO ether.

0
JKmthrjo.bsky.social

My standard is “Bless, o Lord, this food to our use and us in your service, and keep us ever mindful of the needs of others.” I sometimes break out something longer and more formal if the occasion seems to want it.

1
TKtetonkali.bsky.social

My beloved opticianry had no formal trades system but had developed a tiered mastery system just in time for the destruction of the profession. A few states still require licsensure (sp?) but for the most part opticianry is now unskilled sales only. 1/?

1
HGmaxsloan.blog

In other news (more exciting), this is almost done

0
BJbootjack.bsky.social

you'll never guess what this actually turned out to be and i would like to extend a formal apology to my cats for falsely putting them on blast for being stinky on the internet

1