Second, there are already several multimodal models out there; GPT-4 is still a pure text model. Here’s where open-source innovated first.
LLMS
-
Open-Source Model Finetuning Outperforms Generalist Models
By
–
Disagree. You can finetune open-source models. And if done right, finetuned models outperform generalist models. You can’t finetune GPT-4 w/o access.
-
Multimodal LLMs and reasoning advancement in foundation models
By
–
Good points from @ylecun albeit it may just be a matter of time before we have foundation AI LLM models trained on + taking inputs on a multimodal basis (including video and text). However, logic and reasoning whilst improving with Chain of Thought may still take longer to
-
Zero and Few Shot Recommender Systems with Large Language Models
By
–
Zero and Few Shot Recommender Systems based on Large Language Models https://
bit.ly/42TPcHS #AI #MachineLearning #DeepLearning #LLMs #DataScience -
AI Training Fair Use Argument and Copyright Infringement Liability
By
–
On the U.S. side, the argument is a common one that's been made many times before:
"Training is Fair Use." Reproductions of protected works *are* made during training, so AI companies could be liable for infringement, but the argument goes, that's Fair Use. -
Legal battles over LLaMA weights and AI output ownership rights
By
–
VanL was hired by theshawwn to handle a DMCA claim by Meta about taking down the LLaMa weights, so they're trying to prove that weights are in the public domain. He was also hired by icreatelife to attempt to show that it's possible to own rights on the outputs of AI systems.
-
Stochastic Parrots Paper Accessible via ACM Digital Library
By
–
Thank you, @cher0x801 Also — Stochastic Parrots was published as Open Access. There's absolutely no need to point to random storage locations. It will remain accessible at its point of publication in the ACM Digital Library:
-
Language Modelling Success: Next Token Prediction and Transformer Models
By
–
Language modelling has wildly succeeded due to current objective functions(most notably, next token prediction) and evidently powerful, efficient and scalable models(ie. Transformer) . With next token prediction, you have both input text and labels from same text(input is 1 word
-
AI Community Fragmentation Over LLM Understanding Debate
By
–
"I've never seen the AI community become fragmented….with different camps shouting at each other"
— Eric Topol (@EricTopol) 11 juin 2023
on the risks and whether LLMs "understand" https://t.co/U1XUXmyXUK"I've never seen the AI community become fragmented….with different camps shouting at each other" on the risks and whether LLMs "understand"
-
New Book Machine Learning Q and AI Released Today
By
–
Today is the day! My new book Machine Learning Q and AI is now complete! https://
leanpub.com/machine-learni
ng-q-and-ai/
… Covering – Explanations of multi-GPU training paradigms.
– Using and finetuning transformers.
– Differences between encoder- and decoder-style LLMs.
– And many more! 1/3