sayzardfly51fly (@fly51fly)
Sakana AI와 NVIDIA 연구진이 더 작고 빠르며 가벼운 트랜스포머 언어모델을 제안하는 논문을 공개했다. 대형 언어모델의 효율성을 높이기 위한 구조 개선 연구로, 경량화와 추론 속도 향상 측면에서 AI 개발자들에게 중요한 내용이다.
Ramesh #NotGoingBack⬆️ >> #AI got the blame for #Iran school bombing…
Excellent example of how a #languageModel is NOT the same as #worldModel or #realTime #realWorld data.
The #Maven system that #Palantir embedded into the #US military infrastructure relies on BOTH #LLM and #realTime #realWorld data, but it cannot prevent catastrophes when there is a failure in either or both of them.
In this case, it was #staleData at the very least— possibly a faulty/imprecise language model as well.
Leon FurzeTeaching AI Ethics
Update: since I wrote this original post covering the nine areas, I've expanded each one into a complete article. Have a read through this post, and then when you're ready to dive deeper into AI ethics, check out the full series here. If you linked to this post as part of a course or university resource, I suggest updating your links with the complete series. https://leonfurze.com/ai-ethics/ As we head into the start of Term 1 it's already looking like Artificial Intelligence is going to be […]
Using ChatGPT for Conferencing and Feedback
I've used conferencing for years as my main form of feedback and assessment. I stopped collecting piles of books, stopped writing margin notes that no-one ever read, and stopped correcting work like a human spell-checker. Aside from the hours of time saved by not "correcting" work, I also built stronger relationships with students as a result of regularly sitting with them 1:1 to go through their work. At the moment, ChatGPT has been banned by the Department of Education in most states […]https://leonfurze.com/2023/02/08/using-chatgpt-for-conferencing-and-feedback/
Artificial Intelligence and Film Studies: Midjourney versus CAMELS
CAMELS? As an English teacher, I’ve done a lot of analytical film studies over the years. The biggest problem is always students’ grasp of technical language: film analysis tends to drift off into character and plot description, even more so than for written texts. So, when I teach film, I lean more into my Media Studies background. In the UK I completed my undergrad in English and American Literatures. When I trained to teach, it was in English, Media, and Drama. Then, when I first […]https://leonfurze.com/2023/03/29/artificial-intelligence-and-film-studies-midjourney-versus-camels/
Kilo (@kilocode)
Kilo Coders 커뮤니티(혹은 그룹)에서 최신 모델 GLM-5-Turbo에 큰 호응을 보이고 있다는 소식입니다. 출시 직후부터 KiloClaw 내에서 선두주자로 자리잡으며 주목받고 있다는 내용으로, GLM-5-Turbo의 초기 채택과 트렌드 형성을 알리는 발표성 트윗입니다.
Chetaslua (@chetaslua)
MiniMax M2.5가 가격의 1/10 수준에서 모든 모델보다 우수하다는 강한 주장. MiniMax(약 10B 활성 파라미터)의 가성비와 성능 우수성을 강조하며, Opus 4.5와 4.6 비교를 언급하는 내용이다.
N-gated Hacker News🤔 Oh, look—a language model that can supposedly explain itself! Because clearly, what we needed was a robot that can eloquently justify its own nonsensical ramblings. 🚀 Trained on a modest 1.35 trillion tokens, because who needs a life when you can count to a trillion! 😂
https://www.guidelabs.ai/post/steerling-8b-base-model-release/ #languageModel #AItechnology #selfexplanation #humor #technews #HackerNews #ngated
https://www.guidelabs.ai/post/steerling-8b-base-model-release/ #languageModel #AItechnology #selfexplanation #humor #technews #HackerNews #ngated
Hacker NewsSteerling-8B, a language model that can explain any token it generates
https://www.guidelabs.ai/post/steerling-8b-base-model-release/
#HackerNews #Steerling8B #LanguageModel #AI #Explainability #TokenGeneration
Awni Hannun (@awnihannun)
'Recursive LM' 논문 관련 추가 논의에서 핵심은 프롬프트를 세분화해 하위 LLM들이 각각 부분 작업을 수행하고 결과를 통합하는 재귀적 구조라는 점이 강조됩니다. 또한 LLM에 REPL을 제공하여 코드 실행과 같은 상호작용을 가능하게 하는 부분이 주요한 혁신 포인트로 언급됩니다.
Awni Hannun (@awnihannun) on X
Some replies along the lines of the key idea is breaking down the prompt and recursively running sub LLMs on it and stitching them back together. I understand that’s the central premise of the paper. But the lasting nugget is giving the LLM a REPL (which may not even be novel in