I am thrilled to be among faculty members at Georgia Tech who made the CIOS 2023 Fall Honor Roll.
AI
-

MM1: Methods and Insights from Multimodal LLM Pre-training
By
–
MM1: Methods, Analysis & Insights from Multimodal LLM Pre-training McKinzie et al.: https://
arxiv.org/abs/2403.09611 #ArtificialIntelligence #DeepLearning #MachineLearning -
Learning Goals Enhance Team Performance and Knowledge Transfer
By
–
To make teams better, "set learning-based goals for knowledge development, knowledge re-use (knowledge that could be used for future projects), and knowledge transfer (knowledge that could be useful to other teams)." as well as performance goals. Paper: http://
eprints.lse.ac.uk/101239/1/Getti
ng_to_know_you._Motivating_cross_understanding_for_team_and_individual_performance.pdf
… -

Teams Learn Together Perform Better Than Task-Focused Groups
By
–
Teams need to focus on learning, not just doing. When teams focus only on getting a job done, they perform worse than teams that also spend time learning from each other. Groups should be encouraged to think about learning for the future, not just on what they are doing now.
-
Moore’s Law for Language Models: Performance and Speed Trade-offs
By
–
Trying to think about what a Moore's Law for Language Models would look like. the problem i'm running into is that there are two separate axes to improve upon: model ability for a given model size (perplexity) and inference speed (tokens per second) Anyway, we can all agree
-

Cohere Engineering Manager Speaks on Model Fine-tuning and Embeddings
By
–
Hemant Jain, Engineering Manager in Finetuning at Cohere, will speak at the @awscloud Booth during @nvidiagtc on March 21 at 11:45 am PDT & 12:20 pm PDT. Come by Booth #708 to ‘Learn the art of model fine-tuning’ and ‘Why Embeddings are the best kept secret in GenAI’.
-
Why do SD3 and SDXL add excessive moles to generated faces?
By
–
Any reason why, like SDXL, SD3 tends to add so many moles to the face? It seems like a pretty common behavior, and I am not sure where that comes from.
-
Hand Motion Capture Advances Robotic Learning Data Collection
By
–
Data is a key limiting factor for robotic learning. I’m very excited by our work in using a portable hand motion capture system to obtain high fidelity manipulation data. This work is led by @chenwang_j , co-advised by Karen Liu. 😍🦾 https://t.co/qtLZlJFNR6
— Fei-Fei Li (@drfeifei) 14 mars 2024Data is a key limiting factor for robotic learning. I’m very excited by our work in using a portable hand motion capture system to obtain high fidelity manipulation data. This work is led by @chenwang_j , co-advised by Karen Liu.
-
Claude 3 Haiku Now Supports 200k Token Context Window
By
–
We have longer context windows for all Claude 3 models, including the just-released Haiku: https://
poe.com/Claude-3-Haiku
-200k
… -

Upcoming customization options for ChatGPT capabilities in generic conversations
By
–

WIP: Customize ChatGPT section may also get an option to control which capabilities should be available in generic GPT-4 conversations.
