Hugging Face Blog
AI News
Real headlines from AI-dedicated outlets (TechCrunch AI, The Verge AI, OpenAI News, Hugging Face, and more) plus AI-relevant Hacker News stories, pulled every 6 hours — not another analysis, just the fastest real feed of what's happening.
The Sources/Vertical/Function filters above don't apply here — those filter launches, and a news headline isn't a launch. Use Source below to narrow by news outlet instead.
Headline volume, last 14 days
Showing 9551–9600 of 9720 headlines.
OpenAI News
OpenAI has agreed to license its GPT-3 technology to Microsoft. Microsoft plans to utilize the model to power its own products and services.
Hugging Face Blog
OpenAI News
OpenAI News
OpenAI has applied reinforcement learning from human feedback to train language models. This approach enables the models to perform significantly better at text summarization.
OpenAI News
The third class of OpenAI Scholars presented their final projects during a virtual Demo Day. The presentations showcased research results developed by the participants over the course of five months.
Hugging Face Blog
OpenAI News
OpenAI is co-organizing two NeurIPS 2020 competitions alongside AIcrowd, Carnegie Mellon University, and DeepMind. The challenges will center on the Procgen Benchmark and MineRL.
OpenAI News
OpenAI found that large transformer models trained on pixel sequences can generate coherent image completions and samples just as they do with text. The study also established a link between sample quality and classification accuracy, showing that the generative model learns unsupervised features competitive with top convolutional networks.
OpenAI News
OpenAI has announced the release of an API to provide access to its newly developed AI models. The platform allows users and developers to interface with the organization's technology.
OpenAI News
OpenAI News
An analysis from OpenAI reveals that the compute required to train a neural network to AlexNet-level performance on ImageNet has decreased by a factor of 2 every 16 months since 2012, requiring 44 times less compute overall. This improvement significantly outpaces the 11-fold gain expected from Moore’s Law over the same period. The findings suggest that algorithmic advances have contributed more to efficiency gains in high-investment AI tasks than traditional hardware improvements.
OpenAI News
OpenAI has introduced Jukebox, a neural network capable of generating music and rudimentary singing as raw audio across numerous genres and artist styles. Along with the announcement, OpenAI released the model's weights, code, and an interactive tool to explore the generated audio samples.
OpenAI News
OpenAI contributed to a multi-stakeholder report featuring 58 co-authors across 30 organizations that outlines 10 mechanisms to improve the verifiability of claims about AI systems. The proposed tools allow developers to demonstrate that their models are safe, fair, secure, and privacy-preserving. Additionally, the framework provides methods for civil society, policymakers, and users to assess AI development practices.
OpenAI News
OpenAI Microscope is a new collection of visualizations displaying every significant layer and neuron across eight vision models commonly used in interpretability research. The tool makes it easier for researchers to examine features forming within neural networks. OpenAI aims for the resource to help the wider scientific community better understand complex vision systems.
Hugging Face Blog
Hugging Face Blog
OpenAI News
OpenAI has announced that it is standardizing its deep learning framework on PyTorch. The move consolidates the organization's ongoing research and development onto a single framework.
OpenAI News
OpenAI News
OpenAI News
OpenAI demonstrated that the double descent phenomenon occurs across CNNs, ResNets, and transformers, where performance initially improves, degrades, and then improves again as model size, data size, or training time grows. Although this issue can often be mitigated with careful regularization, the exact mechanisms behind it remain incompletely understood. The researchers highlight further exploration of this widespread behavior as an important area for future study.
OpenAI News
OpenAI has released Procgen Benchmark, a set of 16 procedurally generated environments for reinforcement learning. The suite is designed to provide a direct assessment of how quickly AI agents can learn generalizable skills.
OpenAI News
OpenAI News
OpenAI has released Safety Gym, a suite of environments and tools designed for reinforcement learning research. The platform aims to measure progress toward developing agents that can adhere to safety constraints during training.
OpenAI News
OpenAI has released the largest version of GPT-2, featuring 1.5 billion parameters, alongside code and model weights to assist in detecting model outputs. This release concludes the staged rollout plan intended to provide the AI community with a case study for responsible publication processes.
OpenAI News
OpenAI has trained neural networks entirely in simulation to solve a Rubik's Cube using a human-like robot hand. By utilizing Automatic Domain Randomization alongside existing reinforcement learning code, the system can adapt to unprecedented physical disturbances, demonstrating that reinforcement learning can manage complex real-world dexterity tasks.
OpenAI News
Applications have opened for the third class of OpenAI Scholars for the year 2020.
OpenAI News
OpenAI has fine-tuned the 774-million-parameter GPT-2 language model using human feedback across various tasks, noting that the model learned to mirror the specific preferences of external labelers, such as copying input text for summaries. This research is aimed at advancing AI safety techniques centered on machines communicating effectively with humans to extract human values.
OpenAI News
OpenAI researchers observed artificial agents independently developing sophisticated tool use while playing a simulated game of hide-and-seek. Over time, the agents generated six distinct strategies and counterstrategies, demonstrating that multi-agent co-adaptation can produce emergent, complex behaviors.
OpenAI News
OpenAI has introduced a new metric called Unforeseen Attack Robustness (UAR) to evaluate how well neural network classifiers defend against adversarial attacks not encountered during training. The methodology highlights the necessity of assessing model resilience across a broader range of unexpected threats.
OpenAI News
OpenAI has released the 774-million-parameter version of the GPT-2 model following staged rollouts of smaller iterations and subsequent research into potential misuse. Alongside the release, the organization published a technical report on publication norms and introduced an open-source legal template to facilitate model-sharing partnerships between organizations.
OpenAI News
OpenAI holds a weekly Learning Day every Thursday, allowing employees to dedicate time to self-study technical skills outside their daily tasks. The initiative is designed to help staff enhance their professional capabilities.
OpenAI News
Microsoft is investing $1 billion in OpenAI to support the development of artificial general intelligence. As part of the partnership, Microsoft Azure will become OpenAI's exclusive cloud provider, and both companies will collaborate to build advanced AI supercomputing technologies.
OpenAI News
OpenAI has published a policy paper outlining four strategies to encourage long-term industry collaboration on AI safety norms, including technical collaboration and enhanced transparency. The paper warns that competitive market pressures could cause companies to under-invest in safety unless proactive cooperation is established.
OpenAI News
OpenAI hosted its inaugural OpenAI Robotics Symposium on April 27, 2019.
OpenAI News
The second class of the OpenAI Scholars program has concluded, with eight participants completing their final projects. The scholars presented their work during a Scholars Demo Day at OpenAI.
OpenAI News
OpenAI has concluded its second class of OpenAI Fellows, a six-month apprenticeship that transitioned machine learning beginners into core contributors. Applications are currently being reviewed on a rolling basis for the upcoming Summer 2019 fellowship class.
OpenAI News
OpenAI News
OpenAI has introduced MuseNet, a deep neural network capable of generating four-minute musical compositions across ten instruments in various styles ranging from Mozart to the Beatles. The system utilizes the same unsupervised transformer architecture as GPT-2, learning musical harmony and rhythm by predicting subsequent tokens across hundreds of thousands of MIDI files.
OpenAI News
OpenAI has introduced the Sparse Transformer, a deep neural network that sets new records in predicting sequence data across text, images, and audio. It features an algorithmic improvement to the attention mechanism, allowing it to extract patterns from sequences 30 times longer than previously possible.
OpenAI News
OpenAI Five has defeated reigning Dota 2 world champion team OG in two consecutive games, becoming the first AI to beat world champions in an esports title. The victory also marks the first time an AI system has defeated professional esports players during a live broadcast.
OpenAI News
OpenAI announced it will host its final live event for OpenAI Five. The event is scheduled to take place at 11:30 AM PT on April 13.
OpenAI News
Researchers have made progress in the stable and scalable training of energy-based models, yielding improved sample quality and generalization. By expending more compute to iteratively refine outputs, these models achieve sample quality competitive with GANs while retaining the mode coverage guarantees of likelihood-based approaches.
OpenAI News
OpenAI has introduced its 2019 class of OpenAI Scholars, selecting eight participants from a pool of 550 applicants. The cohort features backgrounds spanning literature, philosophy, cell biology, statistics, economics, quantum physics, and business innovation.
OpenAI News
OpenAI has formed OpenAI LP, a new capped-profit entity. The structure is designed to allow rapid scaling of investments in compute and talent while maintaining checks and balances aligned with OpenAI's core mission.
OpenAI News
In collaboration with Google researchers, OpenAI has introduced activation atlases to visualize how interactions between neurons represent information. The technique aims to provide a clearer view into neural networks' internal decision-making, helping developers identify weaknesses and investigate system failures.
OpenAI News
OpenAI has released Neural MMO, an open-ended multiagent reinforcement learning environment that supports a large, variable population of agents. The presence of diverse agents and species encourages improved exploration, niche formation, and overall agent competence.
OpenAI News
OpenAI hosted its inaugural Spinning Up Workshop on February 2. The event was held as part of the organization's broader educational initiatives.
OpenAI News
OpenAI published a paper emphasizing that long-term AI safety research requires social scientists to ensure alignment methods function effectively with human behavior and psychology. The authors seek to foster collaboration between machine learning and social sciences, with plans to hire full-time social scientists at OpenAI.
OpenAI News
OpenAI has developed a large-scale unsupervised language model capable of generating coherent text and setting new state-of-the-art benchmarks. Without task-specific training, the model can also handle basic reading comprehension, translation, question answering, and summarization.