Lex: Can you speak to what Q-Star is? Sam: We’re not ready to talk about that. https://
x.com/0fficialDinesh
/status/1769939750071472165/video/1
…
RESEARCH
-
Lex asks about Q-Star; Sam declines to discuss
By
–
-
Yi Language Model: Full-Stack Development and Key Insights
By
–
Final Thoughts: (5/6) In this report, we discuss the full-stack development of the Yi language model family. Key takeaways highlight the importance of data quantity and quality, along with our unique finetuning procedure outlined in the paper.
-
Dynamic Resource Allocation for Yi Model Upscaling from 6B to 9B
By
–
Depth Upscaling: (4/6) Scaling laws pose significant challenge in resource allocation. Discover our novel approach to dynamically adjust resource allocation, leading to the upscaling of Yi-6B base model to the Yi-9B base model with enhanced training efficiency & performance.
-
Yi Vision Language Models: Architecture and Three-Stage Training Process
By
–
Visional Language: (3/6) Earlier this year, we launched the Yi Vision Language models (Yi-VL-34B, Yi-VL-6B) with bilingual multimodal understanding and generation capabilities. Check out the architecture of the Yi-VL models and our three-stage training process.
-
Yi Model Series Extended to 200K Long-Context Capability
By
–
Long Context: (2/6) The long-context capability offered by the Yi model series (Yi-34B-200K and Yi-6B-200K) has gained immense traction from the community. Learn how we extended the Yi base model to 200K long-context through various methods outlined in the paper.
-
Yi: Open Foundation Models – New Extensions and Capabilities
By
–
“Yi: Open Foundation Models” by http://
01.AI has arrived. Take a look into the specifics with us and discover the groundbreaking extensions of our Yi model series – which include long context, vision language, depth upscaling, and more. (1/6） -
01.AI Infrastructure Team Showcases FP8 Training at GTC2024
By
–
Hello #GTC2024!! Impressive work done by http://
01.AI amazing infra team, presented as a @nvidia official best practice for end-to-end FP8 training and inference. Our larger model is under way -
Wikipedia Boosts Legal Citations Through AI-Written Law Articles
By
–
Wikipedia sets the law. In an experiment in Ireland, a random set of cases were given detailed Wikipedia articles written by law students. The cases that were added became 25% more cited in actual legal opinions by busy judges than they had been before. https://
pubsonline.informs.org/doi/full/10.12
87/isre.2023.0034
… -
ImageNet-12 Dataset Fits in DGX GPU Memory Easily
By
–
yes, imagenet-12 classification set is only 200GB compressed. can literally load it into gpu memory on a dgx box easy.
-
Sam Altman Discusses GPT-5 Capabilities and AGI Vision
By
–
On a Lex Fridman podcast released today, Sam Altman spoke on GPT-5, Sora, his vision for AGI, and more.
— Rowan Cheung (@rowancheung) 19 mars 2024
When asked about GPT-4, Altman said it “sucks“ and that the leap in capabilities for GPT-5 will be similar to GPT-3’s jump to GPT-4. pic.twitter.com/Of1Vh5l0NzOn a Lex Fridman podcast released today, Sam Altman spoke on GPT-5, Sora, his vision for AGI, and more. When asked about GPT-4, Altman said it “sucks“ and that the leap in capabilities for GPT-5 will be similar to GPT-3’s jump to GPT-4.