Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Getty Images and Nvidia have enhanced their AI collaboration with the launch of Generative AI by iStock, a platform for creating stock photos tailored for small and medium businesses that uses Nvidia's Picasso model and trained on Getty and iStock libraries.
A new Copilot key, the first big change to the Windows PC keyboard layout in nearly three decades, will ship on a variety of new PCs and laptops from Microsoft’s partners, providing quick access to Microsoft’s AI-powered Windows Copilot experience straight from a keyboard button press.
One of the best coding models of last year was the DeepSeek LLM. It got close to GPT-3.5 on many benchmarks (even if it is likely 3x the size). Information about model training, token counts, model architecture, and more has been released in a technical report.
Vision Transformers (ViT) have taken over the vision community. They still sometimes display grid-like artifacts in their embeddings. This has made the community hesitant to adopt them for downstream tasks. This work proposes an update to the positional embeddings that eliminates this issue and gives a 25%+ boost for downstream vision tasks.
Researchers have developed a face video editing architecture that combines GAN-NeRF technology for 3D consistency and a novel stabilizer for smooth temporal coherence. This method excels in editing videos by maintaining consistent views and seamless transitions between frames.
Sponsored
QA Wolf [11] gets web apps to 80% automated test coverage in just 4 months. They build and maintain your automated test suite, plus provide unlimited parallel test runs.
in QA engineering + executes >300 tests in parallel on every PR in minutes.
Portkey AI recently open sourced its gateway, which allows for routing between many different language models.
The ALOHA robotics system took the internet by storm with its incredible cooking and household task demonstration. It uses a number of interesting training techniques (mostly imitation learning). This repository contains some baseline training algorithms for the ALOHA system in simulation.
CrewAI is a platform for orchestrating role-playing, autonomous AI agents aimed at enhancing collaborative intelligence and enabling these agents to work together effectively in complex tasks. It provides a structure for sophisticated multi-agent interactions and is designed for various applications, including smart assistant platforms and automated customer service.
This project introduces the Open-Vocabulary SAM, a framework that combines the strengths of CLIP and SAM models for improved image segmentation and recognition.
The New York Times' lawsuit against OpenAI and Microsoft claims that their AI technologies, including ChatGPT, were trained using millions of the newspaper's copyrighted articles, leading to AI-generated outputs that compete directly with the Times' services. The lawsuit seeks billions in damages and the destruction of AI models and data that used the Times' content, arguing that this constitutes copyright infringement and undermines the newspaper's investment in journalism, raising significant questions about the legalities of AI training practices and the impact of AI on traditional content creators.
How diffusion models advance RL. This repository categorizes their applications and points towards future interdisciplinary research opportunities.
Sponsored
Unstructured news data is now accessible for analytics. Learn how NLP automates the process of sifting through thousands of articles to find ‘diamond in the rough’ insights. Read the article on Dow Jones
CBI Analyst is an always-on AI-driven research analyst.
A very simple implementation of inference of the new Mistral MoE model using the Tinygrad library.
Sharp-NeRF introduces a grid-based approach to Neural Radiance Fields, efficiently transforming blurry images into sharp, detailed views.
Nabla, an AI clinical notes assistant company, raised another round of funding to continue improving its technology and accelerate sales.
Open Interpreter is a great project that emulates OpenAI's interpreter. It now also has an OS mode and a visual mode so it can control your computer with a language model by clicking buttons and viewing the screen.
Apple is set to introduce generative AI-based tools, including an enhanced Siri, at its Worldwide Developers Conference in June as part of iOS 18. These innovations are based on Apple's Ajax large language model and may include features like auto-completion and summarization for core and productivity apps, AI-assisted coding in Xcode, and automated playlist creation in Apple Music
There are lots of ways to align a language model, for example, SFT, LoRa, RLHF, DPO, Prompting, and Threatening. This work suggests using negative embeddings added to the bias terms to push model generations towards a desired outcome.
Great research that shows that adding new blocks to a language model and then tuning it on an additional task prevents losing knowledge. This is demonstrated in both code and language tasks.
A framework called GUESS has been introduced for creating human motion from text. It simplifies complex human poses into more abstract forms at various levels, making the synthesis of motion from text more concise and stable.
DiffusionEdge is a novel edge detection model that outperforms existing methods. By integrating a diffusion probabilistic model, DiffusionEdge delivers more accurate and crisp edge maps while being resource-efficient.
This project introduces a method to adapt CLIP-like vision-language models for various tasks while retaining its generalization ability. The approach learns prompts from LLM data, avoiding the need for labeled images.
LLaVA-Phi is an innovative multi-modal assistant that uses the Phi-2 language model. Despite its small size of 2.7B parameters, LLaVA-Phi excels in multi-modal dialogues, combining text and visuals.
A really great introduction to Transformers, accessible with basic math and Python skills.
The "2023 Expert Survey on Progress in AI" reveals a scientific community divided on the pace and impact of AI development. Predictions of significant milestones have been reached earlier than expected. Over half of the respondents are optimistic about AI's impact on humanity, but concerns remain about its potential misuse, highlighting the need for research on risk mitigation.
An extremely in-depth look into how one might build an AI with human-like thought processes.
Sponsored
Discover the capabilities of Hugging Face models for question answering, Q&A architectures, and hands-on coding tips. DATE & TIME: Jan 10 at 6:30 PM PST.
V* is a new grounded visual search algorithm that boosts GPT-V (and other VLM) captcha-solving performance significantly.
Microsoft VP Dee Templeton has joined the OpenAI board as a nonvoting observer.
Two nonfiction authors filed a class action lawsuit against Microsoft and OpenAI, alleging copyright infringement in the creation of AI systems like ChatGPT. The lawsuit follows a similar case by The New York Times and includes claims that the defendants used copyrighted works without permission. The authors are seeking to represent a broad class of impacted writers and damages of up to $150,000 per infringed work.
The generative search company, which has seen massive adoption in tech circles, has raised another round of funding from top VCs.
OpenAI plans to launch its GPT Store sometime in the coming week. To be eligible for listing, developers have to ensure their GPTs are compliant, verify their user profile, and make their creations “public”.
Google is reportedly developing "Bard Advanced," an enhanced version of its Bard AI tool that will be accessible through a paid Google One subscription. This upgraded version, powered by Google's top-tier large language model Gemini Ultra, offers advanced math and reasoning capabilities compared to the current Bard running on Gemini Pro.
MUSE is a masked image generation model that uses Transformers. HuggingFace and Stability AI have worked to reproduce it in the open and have released training code, model checkpoints, and a paper. The models were not trained for as long as some of the best image diffusion models, but they show promise!
Researchers have developed CAINet, a novel approach for improving RGB-T semantic segmentation, crucial for autonomous driving. This system uniquely combines different types of data, focusing on their complementary aspects and global context.
While Visual question answering (VQA) has seen advancements, most models only tackle 2D reasoning, neglecting the complexity of 3D visual scenes. This research presents 3D-aware VQA.
Retrieval Augmented Generation (RAG) is a way to include external knowledge into a language model’s generation. This library allows for training and researching state-of-the-art RAG systems.
ODTrack, a new video tracking system, improves the way we track objects in videos. Unlike older systems that struggled with connecting past and future frames, ODTrack uses a method called 'online token propagation' to better link consecutive video frames.
This survey provides an in-depth look at the vulnerability of recommender systems to 'poisoning attacks', where harmful data is intentionally added to manipulate results.
The new chip from Intel, available on a number of platforms and on-premise deployment, has excellent training throughput, flops, and decoding bandwidth.
Dropbox's new AI features have raised privacy concerns among users who fear their data might be used to train OpenAI's models, despite Dropbox's denial and a policy that requires customer consent for such use. This situation highlights a broader crisis of trust in AI and data privacy, emphasizing the need for clear communication and transparency from companies about their data usage practices.