On August 27, we returned to Seoul to host Research@ Korea. Researchers spoke about topics including NotebookLM, Med-Gemini, the UX of AI agents, and Connectomics. We also heard from the Korean startup and academic communities. Special thanks to our guests, visiting speakers,
MACHINE LEARNING
-
ULMFiT technique adoption opportunity for major AI labs
By
–
We’ve been teaching this since 2017 and used it in ULMFiT, but still seems no big lab has picked up on it. Big opportunity for someone here…
-
LLMOps Summit Talks Now Available in Video Playlist
By
–
Missed our #LLMOps Summit, but want to catch all the goodness? Now you can! Here's a playlist with all of the talks: https://
pbase.ai/47lSneN Last week 150 #AI leaders and ML engineers joined us in Downtown San Francisco to discuss the future of #GenAI with a focus on how to make -
Transfusion: Multi-Modal Model for Token Prediction and Image Diffusion
By
–
New research paper from Meta FAIR – Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal Model. @violet_zct
, @liliyu_lili and team introduce this recipe for training a multi-modal model over discrete and continuous data. Transfusion combines next token -
COAR Method Improves Attribution in Vision Language Models
By
–
Experiments on large-scale vision & language models show COAR yields accurate component attributions, outperforming prior methods. This makes it a valuable tool for understanding & debugging complex ML systems.
-
COAR: Scalable Predictive Component Attribution Method
By
–
In summary, COAR is a scalable method for estimating predictive component attributions that outperform prior approaches across models & tasks. COAR's attributions act as a counterfactual estimator, helping w/targeted model edits — from fixing errors to boosting subpopulation
-
COAR: Surgical Edits to Fix ML Model Errors
By
–
The approach, called COAR, allows us to make surgical "edits" to ML models. By intervening on the right components, we can fix errors, "forget" harmful labels, or boost performance on subpopulations that are underrepresented in data.
-
COAR: Component Attribution for Targeted Model Edits
By
–
COAR works by estimating the contribution of each component to the model’s final output. These component attributions act as a counterfactual estimator, enabling targeted model edits w/o additional training.
-
MIT Research Reveals How Neural Networks Transform Data Into Predictions
By
–
How do black-box neural networks transform raw data into predictions? Inside these models are thousands of simple "components" working together. New MIT CSAIL research (
https://
bit.ly/473lcfE) introduces a method that helps us understand how these components compose to affect -
NVIDIA Blackwell Achieves 4X Performance on Llama 2 70B
By
–
In the latest round of #MLPerf industry benchmarks, the first submission of NVIDIA Blackwell delivered up to 4X more performance on #Llama 2 70B, while the NVIDIA H200 Tensor Core GPU continued to deliver outstanding results across every data center test.