Parse notes
Meta's new models have been fine-tuned and instruction-tuned on significantly more code. They achieve impressive performance on HumanEval, approaching GPT-4 in some respects. The models are available under the Meta license and on HuggingFace
RWKV is an architecture alternative to the dominant Transformer model. It uses a novel linear attention mechanism that makes it extremely efficient. This new checkpoint and configuration was trained on 1T tokens and surpases many other powerful 7B Transformer based models.
OpenAI is suspected of violating European Union privacy laws. Italy's data protection authority has given it 30 days to respond to allegations following an investigation into ChatGPT.
One key challenge with 3D object generation is the diversity of items that can be created. This work uses a modified architecture to improve sample efficiency and scales up the system to work with a much larger set of objects in each 3D category.
Another potential distillation work. Importantly, this one can work on models as small as Phi-2. This means you can remove 90% of the rows and columns of weight matrices with minimal reduction to quality at almost all scales.
Meta Learning is the process of training systems that learn to learn and can quickly adapt to new tasks. This work from Google improves Meta Learning with synthetic data generated from a Universal Turing Machine and analyzes the results experimentally and theoretically.
Sponsored
Microsoft is marketing its Fabric Lakehouse architecture as an ideal way to manage Power BI semantic models. This tech talk [12] will examine these claims in light of AtScale’s TPC-DS benchmark [12], measuring Power BI performance on Microsoft Fabric on 100GB, 1TB and
CreativeSynth is a method for artistic image editing by seamlessly integrating text and image inputs. Its diffusion model, equipped with customized attention mechanisms, enables precise style and content manipulation while preserving the original artwork's essence.
Inconsistency Masks (IM) are a new method for image segmentation that works even with limited data. This approach, tested on the ISIC 2018 dataset, beats traditional techniques and even outperforms models trained on fully labeled datasets.
A new image distortion technique makes facial images unrecognizable to humans but still identifiable by AI, striking a balance between privacy and accuracy in biometric systems.
Researchers have improved the ability of generative models like GANs and diffusion models to create realistic hand images by adding three extra channels to training images for hand annotations.
Looking for technical sales specialist interested in AI/LLMs with 3+ years of sales experience in cloud computing/infrastructure to help drive adoption of our NVIDIA GPU cloud. Apply here [18].
Shortwave released an AI-powered autocomplete feature that's similar to Github Copilot but designed for writing emails. It learns from sent emails and provides contextual suggestions.
An open source platform for AI-native application development.