Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Qwen has released a 32B parameter reasoning model that performs similarly to OpenAI's o1 series of models. The company observed scaling performance when sampling longer reasoning traces. The model is available to use and excels at mathematics and coding.
OpenAI has filed a trademark application for its latest AI model, o1, as the firm moves to shield its IP.
ElevenLabs has introduced GenFM, a feature in its iOS app that creates AI-generated multispeaker podcasts from uploaded content. Supporting 32 languages, the feature adds natural human elements like 'ums' and pauses to enhance authenticity in the podcast experience. ElevenLabs has planned further customization options and an expansion into new markets, including India and Poland.
In early image synthesis, traversing the latent space was a great way to generate diverse images. Now, the same can be done with video to generate sequential edits to a single image while maintaining semantic consistency.
Models trained for less time on fewer tokens experience less degradation when being quantized after training. This suggests, as has been seen in other work, that models use the extra precision to fully compress more and more information during extended training.
Star Attention offers a block-sparse approach to speed up Transformer-based large language models (LLMs) during long-sequence inference.
SketchAgent leverages an off-the-shelf multimodal LLM to facilitate language-driven, sequential sketch generation through an intuitive sketching language. It can sketch diverse concepts, engage in interactive sketching with humans, and edit content via chat.
A deep-learning-based dense visual SLAM framework that achieves real-time global optimization of poses and 3D reconstruction.
P2DFlow is a protein ensemble generative model with SE(3) flow matching based on ESMFold. The ensembles generated by P2DFlow could aid in understanding protein functions across various scenarios.
Cradle has solved a critical challenge in optimizing protein shapes. It is now expanding its team and efforts to land this technology in the hands of practitioners everywhere.
Rabbit R1 has introduced a teach mode feature which allows users to train its AI to automate tasks on any website. Aiming to improve functionality and productivity, the feature supports complex multi-platform interactions. It potentially offers a better experience than using dedicated apps. Rabbit envisions a marketplace for user-created automations. It is hoping for broader adoption despite potential platform challenges.
Recent experiments suggest that GPT-3.5-turbo-instruct can play chess well compared to other LLMs due to potential differences in base model training and data quality. Various prompting techniques and examples enhanced chess performance of other models like GPT-4o, though not to the same extent. The analysis indicates OpenAI's strategy might involve curating high-quality chess data for better LLM performance.
The team behind ShowUI has introduced a vision-language-action model designed to improve how AI interacts with graphical user interfaces.
MIT researchers have developed an efficient algorithm for enhancing reinforcement learning models that involves strategically selecting training tasks for improved decision-making in variable environments.
Apple's 2024 iPhone App of the Year finalists largely exclude AI-driven apps, favoring traditional applications like Kino and Runna, though some AI apps are recognized in iPad and Mac categories.
Sponsored
As the stock market reaches new highs, the potential for a correction could raise concerns for investors heavily exposed to equities.
Mochi 1, the leading open source video model, has LoRA fine tuning available. It runs on a single GPU. This repository shows a number of uses, including custom effects and character consistency.
Anthropic now supports custom styles, allowing users to tailor the AI's responses to their communication preferences and workflows.
A group leaked access to OpenAI's unreleased video generator Sora as a protest against alleged unfair practices and "art washing". They published a frontend on Hugging Face allowing users to generate videos, which OpenAI reportedly shut down after a few hours. OpenAI maintains that Sora is in research preview.
Nvidia has showcased a new 2B audio model that is far more flexible than previous models. It can mix and match audio, music, and speech in novel and interesting ways.
A continuation of the fully open source training efforts, Allen AI has released a new generation of language models that are transparent from top to bottom. They match or surpass the best available open weight models.
OneDiffusion is a large-scale diffusion model designed for a wide range of tasks, from text-to-image generation and image editing to reverse processes like depth estimation and segmentation.
Sponsored
New flash attention fork that can have ragged Q/V matrix sizes.
MVGenMaster is a multi-view diffusion model designed to improve Novel View Synthesis tasks by integrating 3D priors.
This work benchmarks discrete flow matching methods for generating novel 3D molecular structures, critical for chemical discovery.
This project explores the emerging "LLM-as-a-judge" paradigm where large language models are used for scoring, ranking, and selection tasks across various AI and NLP applications.
AI boosts material scientists' efficiency, leading to a 44% increase in material discoveries but a 44% drop in work satisfaction due to reduced idea generation tasks. GitHub Copilot enhances productivity for less skilled developers, shifting focus from project management to coding. AI narrows skill gaps but risks alienation by automating creative tasks, echoing past automation impacts in various industries.
Hybrid approaches that combine classical search methods and machine learning are showing promise in overcoming the complexities of chip design, particularly in floorplanning tasks. While AI alone struggles with multi-constraint scenarios, integrating AI for guiding search-based algorithms like simulated annealing enhances efficiency and effectiveness. This combination can accelerate design processes and enable more complex chip solutions.
An easy way to work with a variety of API based models in a single packaged environment.
Amazon announced a $4 billion investment in Anthropic, bringing its total investment to $8 billion while establishing AWS as Anthropic's primary cloud and training partner.
OpenAI is funding Duke University to research AI algorithms that predict human moral judgments.
Sponsored
Secure your AI Pipeline with Piiano's Data Privacy Vault [5]:
The Model Context Protocol (MCP) is a new open standard designed to connect AI systems directly to data sources like business tools and content repositories. It simplifies the process of giving AI access to the data it needs by replacing fragmented, custom integrations with a universal protocol, making systems more scalable and effective.
OpenAI has expanded its red teaming efforts by releasing two papers: one detailing how external experts are engaged in red teaming and another that introduces a new method for automated testing.
"Test-time scaling" is becoming more important as AI models evolve. Nvidia is ready for this shift. This method, which enhances AI inference by increasing compute power, poses competitive challenges as startups develop fast AI inference chips. Despite concerns of diminishing returns, Nvidia remains focused on leveraging its dominant platform advantage for pretraining but anticipates growth in AI inference.
"Coalescence" is a framework that makes LLM inference up to 5x faster when generating structured output like JSON. The technique works by converting structured formats into finite state machines and identifying redundant paths that lead to the same output, allowing them to skip unnecessary LLM calls. While it significantly improves speed, it is important to maintain output quality by ensuring the optimization doesn't prevent more probable sequences from being generated.
WildLMa is a framework that enables quadruped robots to perform complex manipulation tasks in real-world environments. The system combines three key components: a whole-body controller for VR-enabled teleoperation, a library of generalizable skills learned through imitation learning (WildLMa-Skill), and a language model-based planner (WildLMa-Planner) that coordinates these skills for long-term tasks. In this paper, researchers demonstrate how it can be used for tasks like cleaning up trash in hallways and rearranging items on bookshelves. The framework maintains effectiveness across diverse environments and object configurations.
MMGenBench is a new evaluation pipeline for large multimodal models that focuses on their ability to generate and interpret images. Using this method, models generate descriptions from input images, which are then used to create new images for comparison.
Sponsored
1️⃣ Join Bria's Visual AI Coding Sessions @ AWS RE:INVENT and
2️⃣ Leverage Azure "Model as a Service" Credits" : Bria's fully licensed Text-to-Image model is now in the AZURE AI FOUNDRY! 👉
Marco-o1 is an LLM that aims to handle both standard problems (like math and coding) and open-ended tasks without clear correct answers. The model combines Chain-of-Thought fine-tuning, Monte Carlo Tree Search, reflection mechanisms, and novel reasoning strategies to tackle complex real-world problems where traditional reinforcement learning approaches might struggle due to unclear reward signals and success metrics.
Moondream's Python client library contains tools for image analysis and querying. It offers CPU-optimized inference but isn't recommended for GPU or Mac M1/M2/M3 users yet. Users can install it via pip. Model weights can be downloaded in different formats (int8, fp16, or int4).
Sana is an efficient image generation model that can create high-quality 1024x1024 images in under a second on a laptop GPU. Its key innovations include a 32x image compression autoencoder (DC-AE), linear attention replacing vanilla attention in DiT, a decoder-only LLM for text encoding, and optimized training and sampling methods. The 0.6B parameter model matches or outperforms much larger models like Flux-12B while being 20x smaller and 100x faster. Sana-0.6B requires only 9GB VRAM for inference, making it accessible for consumer hardware. The repository includes code for training, inference, and evaluation, with both 0.6B and 1.6B model variants available.