Parse notes
Rabbit created a large action model (LAM) that goes with its R1 pocket companion computer. The company's R1 device is a reimagination of the computer and smartphone, powered almost entirely by its LAM.
OpenAI is offering news publishers between $1M and $5M a year to use their content to train its AI models.
Intel has announced an AI-focused chip for the automotive sector. Described as "the first-generation AI-enhanced software-defined vehicle system-on-chip," the announcement marks a strategic shift towards AI PCs and automotive AI applications.
Google researchers have introduced Self-Play Preference Optimization (SPO), a simpler form of alignment than traditional RLHF. The researchers turned to game theory and figured out single-player self play dynamics that are robust to noisy preferences and give strong performance.
The Mixtral paper has been released. There is no discussion of pretraining datasets. Most of the things disclosed are well known, but there is an interesting new insight on the discussion of routing evaluation between experts.
Researchers have developed a new evaluation metric for text-to-3D generative models, addressing the limitations of current single-criterion metrics. This advanced method uses GPT-4V to create prompts and compare 3D assets. It aligns closely with human preferences and offers versatility by adapting to various user-defined criteria.
AI Toolkit is a header-only C++ library that provides tools for building the brain of your game's NPCs.
SpeechAgents is a multi-modal AI system capable of simulating human communication with remarkable authenticity. This system, powered by a multi-modal LLM, can handle up to 25 agents. It shows potential in creating dramas and audio novels through its ability to mimic human dialogues, complete with consistent content, realistic rhythms, and expressive emotions.
In a significant advancement for Vision Transformers (ViTs), researchers have improved the robustness of the Fully Attentional Network (FAN) models using a new self-emerging token labeling (STL) framework. This approach involves training a FAN token labeler to create meaningful patch token labels, followed by training a FAN student model.
This essay explores Robin Hanson's critique of AI risk discourse, focusing on the concept of "othering" AIs and the ethical implications of controlling or altering future AI and human values. Hanson argues that fearing AI as an "other" is a form of bias. Hanson's views may not sufficiently consider the risks of uncontrolled AI development and the challenges in aligning future AI values with human ethics.
This article introduces the "Anatomy Framework", a tool for assessing the disruptive potential of any innovation, including AI. It explores innovation from five layers: underlying infrastructure, core platform, applications, tools, and ecosystem enablers. It also discusses the role of startup and incumbent innovators and the medium of innovation (hardware vs. software).
Researchers have developed a new Multi-disciplinary Collaboration framework to improve Large Language Models' understanding and reasoning in the medical field. This approach involves AI agents role-playing working together to analyze, discuss, and reach consensus on medical topics.
This project introduces a one-shot method for editing human images that can make significant changes in pose and body shape while preserving the person's identity.
A personalized AI sales coach for each and every cold call.
The switch transformer from Google was one of the first successful Mixture-of-Experts models. It now is available with code on the HuggingFace platform.