Parse notes
AI startup Reka has announced a 21B parameter model that was trained completely from scratch. It is natively multimodal and surpasses Llama 70B and Mixtral on a number of key benchmarks. Importantly, given where the founders of the company used to work, it also matches Gemini Pro. The model is available for use in Reka's playground.
Apple has launched MGIE, an open-source AI model for image editing that uses multimodal large language models to interpret text instructions and execute pixel-level edits.
DeepMind and USC researchers have developed SELF-DISCOVER, a new framework that significantly improves LLM reasoning capabilities, outperforming traditional methods by up to 32% on tasks with models like GPT-4. The two-stage framework allows LLMs to form and utilize reasoning structures autonomously. This advancement represents a major step toward more sophisticated AI problem-solving abilities.
"Box o Flows" introduces a novel experimental system for testing reinforcement learning (RL) algorithms in dynamic real-world settings, addressing the challenge of simulating complex fluid dynamics for RL applications. It showcases the ability of model-free RL algorithms to generate complex behaviors from simple rewards and enhances data efficiency via offline RL and paves the way for broader RL applicability in complex systems.
Weblinx is a dataset of 100k web-based interactions in a conversational format. It was released to improve the study of language model guided web-based navigation.
This paper introduces ImplicitDeepfake1, a new approach that combines deepfake technology with Neural Radiance Fields (NeRFs) and Gaussian Splatting (GS) to create highly realistic 3D avatars.
Sponsored
The AI in Production [11] summit is happening this Thursday and next [11]! Learn from 70+ speakers [11] from the companies that are leading the way in AI (Hugging Face, Microsoft, Netflix, Uber, Perplexity…), get practical tips in hands-on workshops, and enjoy fun surprises
Apple released code that uses multimodal language models to improve human-provided natural language edits to images.
Lag-Llama introduces a breakthrough in time series forecasting with a transformer-based model that excels in zero-shot and few-shot generalization, pre-trained on a broad spectrum of time series data. It surpasses existing deep learning methods in state-of-the-art performance through fine-tuning on minimal data, marking a significant advancement in foundation models for time series analysis.
A repository for training end-to-end control policies using deep reinforcement learning. The training is done in simulation and is finished within seconds on a consumer-grade laptop. The trained policies generalize and can be deployed on real quadrotors.
The concept of Massed Muddler Intelligence (MMI) represents a shift from traditional monolithic AI scaling towards a model based on distributed, agent-based systems that learn and adapt in real time. Grounded in principles of embodiment, boundary intelligence, temporality, and personhood, MMI advocates for AI development that emphasizes scalable, interactive agents with a degree of autonomy and mutual governance, moving away from the current focus on accumulating larger datasets and computational resources.
Packing in training models is a way to improve training efficiently by concatenating examples. If done incorrectly, contamination can occur between examples because the attention doesn’t know where to stop. The community has found that EOS is often sufficient, but there still can be problems. This repository provides an implementation in Hugging Face for common models to properly pack input data.
Free AI tool to audit and optimize online forms.
Meta plans to combat the spread of fake images on its platforms by detecting and labeling AI-generated content, even if produced by rival services.
Axolotl is a tool used by many groups to fine-tune language models. This PR adds support for Mac M devices via Pytorch and MPS.