-
This AI Paper Introduces Lemur and Lemur Chat For Harmonizing Natural Language and Code For Language Agents
The University of Hong Kong, XLang Lab, Salesforce Research, Sea AI Lab, University of Washington, and MIT CSAIL have developed Lemur and Lemur-Chat, two state-of-the-art models for language agents. By combining natural language and coding abilities, Lemur and Lemur-Chat outperform other open-source models in agent benchmarks, bridging the gap between open-source and commercial alternatives. The…
-
Rethinking the Role of PPO in RLHF
Researchers propose Pairwise Proximal Policy Optimization (P3O), a new approach to Reinforcement Learning with Human Feedback (RLHF) that addresses the inconsistency between the reward learning and RL fine-tuning stages. By using a comparative training process, P3O improves alignment with human values and outperforms existing methods in terms of the KL-Reward frontier and GPT-4 win-rate. The…
-
A method to interpret AI might not be so interpretable after all
Formal specifications, which use mathematical formulas to describe AI behavior, are not easily interpretable by humans, according to researchers at MIT Lincoln Laboratory. In an experiment, participants were asked to validate an AI agent’s plan for a virtual game based on formal specifications, and they were correct less than half of the time. The researchers…
-
CMU & Google DeepMind Researchers Introduce AlignProp: A Direct Backpropagation-Based AI Approach to Finetune Text-to-Image Diffusion Models for Desired Reward Function
The paper discusses the emergence of text-to-image diffusion models for image generation. It introduces “AlignProp,” a method to align diffusion models with reward functions through backpropagation during the denoising process. AlignProp outperforms alternative methods in optimizing diffusion models, achieving higher rewards in fewer training steps and improving both sampling efficiency and computational effectiveness. The approach…
-
The US government moves to further restrict tech exports to China
The US government plans to implement additional sanctions to prevent American chipmakers from circumventing export restrictions on AI chips going to China. The upcoming regulations will close loopholes that allowed Chinese companies to obtain specialized AI chips through foreign distributors. The new measures will also prohibit the sale of advanced chipmaking machinery and semiconductors to…
-
Another researcher identifies singed text from the Herculaneum scrolls
Ancient scrolls from Herculaneum, buried for centuries, have started to reveal their secrets. Using AI technology, a computer science student and a data science graduate have made breakthroughs in deciphering the charred papyrus. They have identified the word “porphyras” using different AI techniques. The competition to understand the Herculaneum scrolls is heating up, thanks to…
-
How Veriff decreased deployment time by 80% using Amazon SageMaker multi-model endpoints
Veriff is an identity verification platform partner for organizations in various industries. They use advanced technology, including AI-powered automation and human feedback, to verify user identities. Veriff standardized their model deployment workflow using Amazon SageMaker, reducing costs and development time. They use SageMaker multi-model endpoints and Triton Inference Server to manage and deploy ML models…
-
Carbon Emissions of an ML Engineering Team
This text discusses the significance of the hidden costs of development. It emphasizes the importance of recognizing and considering these costs in order to ensure accurate decision-making and successful project outcomes.
-
Unlocking AI Transparency: How Anthropic’s Feature Grouping Enhances Neural Network Interpretability
Researchers have developed a new framework using sparse autoencoders to make neural network models more understandable. The framework identifies interpretable features within the models, addressing the challenge of interpretability at the individual neuron level. The researchers conducted extensive analyses and experiments to validate the effectiveness of their approach, and they believe it can enhance safety…
-
From 2D to 3D: Enhancing Text-to-3D Generation Consistency with Aligned Geometric Priors
Researchers have developed a method called SweetDreamer to address the issue of geometric inconsistency in converting 2D images to 3D objects for text-to-3D generation. This method aligns 2D geometric priors with well-defined 3D shapes to ensure consistency from all viewpoints. The researchers achieved high success rates compared to other methods and believe their work will…