They are! But both are mitigated by RL-based tuning and have similar costs in diversity and creativity. To tune for correctness is to optimize for tokens that, in practice, lead to true output, regardless of how well those choices represent real text. Consider whether an LLM
RESEARCH
-
Segment Anything 8x Faster with PyTorch Advanced Features
By
–
make "Segment Anything" 8x faster by just using PyTorch's more advanced features. No C++, no custom extensions. great post (and work) by @cpuhrsch and the team!
-
Self-Attention Mechanism in Large Language Models Explained
By
–
Understanding and Coding the Self-Attention Mechanism of Large Language Models From Scratch https://
bit.ly/3YL821l #AI #DeepLearning #MachineLearning #DataScience -
Google DeepMind Unveils Lyria GenAI Model for Music Creation
By
–
Google DeepMind unveils Lyria, a powerful #GenAI model for music creation https://
venturebeat.com/ai/google-deep
mind-unveils-lyria-a-powerful-genai-model-for-music-creation/
… via @VentureBeat -
Yann LeCun defends Galactica: Hallucination isn’t a crime, valuable models need less.
By
–
Yann LeCun is right: Galactica was wronged. Hallucination is not a crime. Models designed to hallucinate less (none eliminate it) hold great value — in dollars, perhaps most. Alignment instills reason and diligence as much as any other of our mores. In many ways, models are more
-
Google BRAIDS Team Advances Responsible AI Practices
By
–
The Building Responsible AI Data & Solutions (BRAIDS) team within #GoogleResearch aims to simplify adoption of RAI practices through the utilization of scalable tools, high-quality data, streamlined processes and novel research. Learn about their work ↓
-
ABC Science Series Explains Modern AI Implications Clearly
By
–
If you want to understand the implications of what's happening in AI today, or have non-techie friends & family that you want to explain it to, this whole series from @ABCscience is absolutely top notch:
-
ABC Australia Series on AI and ULMFiT Creation Now Available
By
–
ABC Australia does some of the best science content in the world, and now they're creating a pop-sci series on AI which is the best I've heard. An episode in which I describe the creation of ULMFiT, & my dreams and worries about AI, is now available:
-
RAG Training Flaw: Models Fail to Ignore Irrelevant Context
By
–
I think they made to common mistake during RAG-training of pretty much *always* having relevant context available, so the model never learns to ignore it.
-
ML Community’s Growing Interest in Reinforcement Learning and RLHF
By
–
And ML people are very interested in RL nowadays, thanks to RLHF's success!