OpenAI News
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 9601–9650 of 9720 headlines.
OpenAI News
The inaugural Summer 2018 cohort of the OpenAI Fellows program has concluded. Over the six-month apprenticeship, participants advanced from machine learning beginners to core contributors at OpenAI.
OpenAI News
OpenAI researchers found that the gradient noise scale metric accurately predicts how well neural network training can be parallelized across different tasks. Because complex tasks exhibit noisier gradients, using increasingly large batch sizes can unlock greater scale and help systematize the training process.
OpenAI News
OpenAI has launched CoinRun, a new benchmark environment designed to evaluate how effectively reinforcement learning agents transfer knowledge to unseen situations. The platform is simpler than classic games like Sonic the Hedgehog while still offering a substantial challenge for testing generalization.
OpenAI News
OpenAI has launched Spinning Up in Deep RL, an educational resource aimed at teaching practitioners deep reinforcement learning. The package includes example code, exercises, documentation, and tutorials to guide learners through the field.
OpenAI News
OpenAI has developed an energy-based model capable of identifying and generating spatial concepts like near, above, and between after just five demonstrations with sets of 2D points. The model also demonstrates cross-domain transfer by applying concepts learned in a 2D particle environment to tasks on a 3D physics-based robot.
OpenAI News
OpenAI News
OpenAI has introduced Random Network Distillation (RND), a prediction-based curiosity method designed to encourage reinforcement learning agents to explore environments. Using this technique, an agent has surpassed average human performance on Montezuma's Revenge for the first time.
OpenAI News
OpenAI has proposed iterated amplification, an AI safety technique that specifies complex, beyond-human-scale goals by decomposing tasks into simpler sub-tasks instead of relying on labeled data or reward functions. While currently in early stages with experiments limited to toy algorithmic domains, it is presented as a potentially scalable approach to AI safety.
OpenAI News
OpenAI is accepting applications for the second cohort of its OpenAI Scholars program. The initiative provides stipends and mentorship to 6 to 10 individuals from underrepresented groups to study deep learning full-time for three months and produce an open-source project.
OpenAI News
Applications are now open for the 2019 OpenAI Fellows and Interns programs. Opportunities include the Winter 2019 fellowship and Summer 2019 internships.
OpenAI News
OpenAI News
The first cohort of OpenAI Scholars has finished their program. The participants have now completed their final projects.
OpenAI News
At The International in Vancouver, OpenAI Five was defeated in two games by top human Dota 2 players. Despite the loss, the system maintained a strong chance of winning for the first 20 to 35 minutes of both matches.
OpenAI News
OpenAI News
OpenAI Five defeated a team of 99.95th percentile Dota players, including four former professionals, in a best-of-three benchmark match. The event took place in front of a live audience alongside 100,000 concurrent livestream viewers.
OpenAI News
OpenAI has trained a human-like robot hand capable of manipulating physical objects with unprecedented dexterity.
OpenAI News
OpenAI News
The inaugural class of the OpenAI Scholars program is currently underway. The initiative follows a group of experienced software developers as they transition into becoming machine learning practitioners.
OpenAI News
The OpenAI Five Benchmark match has officially concluded.
OpenAI News
OpenAI has introduced Glow, a reversible generative model that simplifies existing architectures using invertible 1x1 convolutions. It can generate realistic high-resolution images, support efficient sampling, and manipulate data attributes. Code for the model and an online visualization tool have been released publicly.
OpenAI News
OpenAI trained an agent to reach a record score of 74,500 on Montezuma's Revenge using only a single human demonstration. The agent plays sequences starting from selected states in the demonstration and optimizes its score using Proximal Policy Optimization (PPO).
OpenAI News
OpenAI Five, a team consisting of five neural networks, has begun defeating amateur human teams in the game Dota 2.
OpenAI News
The initial run of OpenAI's Retro Contest has concluded. The competition investigated how to build algorithms capable of generalizing from previous experience.
OpenAI News
OpenAI News
OpenAI has achieved state-of-the-art performance across diverse language tasks by combining transformers with unsupervised pre-training. The organization is releasing this scalable, task-agnostic system to demonstrate the effectiveness of pairing supervised methods with unsupervised pre-training and encourage further research on larger datasets.
OpenAI News
OpenAI News
OpenAI is accepting applications for the Fall 2018 cohort of OpenAI Fellows. The program provides a paid, six-month apprenticeship in artificial intelligence research at OpenAI.
OpenAI News
OpenAI has released the full version of Gym Retro, expanding its reinforcement learning platform to over 1,000 games across multiple emulators. The release also includes the internal tool used to integrate new games into the platform.
OpenAI News
A new analysis shows that compute used in the largest AI training runs has grown exponentially since 2012, doubling roughly every 3.4 months for an overall increase exceeding 300,000x. The authors highlight compute improvements as a primary driver of AI progress and urge preparation for systems with capabilities far beyond present-day models.
OpenAI News
OpenAI is proposing a new AI safety technique focused on debate. Under this approach, agents are trained to debate topics against each other, with a human serving as the judge to determine the winner.
OpenAI News
OpenAI has introduced Evolved Policy Gradients (EPG), an experimental metalearning method that evolves the loss function of learning agents to speed up training on new tasks. Agents trained with EPG can successfully execute basic tasks that fall outside their initial training regime, such as finding objects in novel locations.
OpenAI News
OpenAI News
OpenAI has launched the Retro Contest, a transfer learning competition. The contest is designed to evaluate how effectively reinforcement learning algorithms can generalize from past experience to novel environments.
OpenAI News
OpenAI News
OpenAI News
OpenAI hosted its inaugural hackathon on March 3rd. The event brought together 100 members from the artificial intelligence community.
OpenAI News
OpenAI News
OpenAI has developed Reptile, a simple meta-learning algorithm that iteratively samples tasks, applies stochastic gradient descent, and updates initial parameters toward the learned solutions. The algorithm applies the Shortest Descent method to meta-learning and achieves similar computational efficiency and performance to first-order MAML while requiring only black-box optimizer access.
OpenAI News
The OpenAI Scholars program is offering 6 to 10 stipends and mentorship to individuals from underrepresented groups. Participants will study deep learning full-time for three months and release an open-source project.
OpenAI News
OpenAI News
OpenAI News
OpenAI is releasing eight simulated robotics environments alongside a Baselines implementation of Hindsight Experience Replay, which were used to train models on physical robots. In addition, the team is publishing a set of requests for research to guide further work in robotics.
OpenAI News
OpenAI is hosting a hackathon featuring talks at its office in San Francisco's Mission District. The event is scheduled to take place on Saturday, March 3rd.
OpenAI News
OpenAI has co-authored a report analyzing how malicious actors could misuse artificial intelligence and outlining strategies to prevent and mitigate those risks. The publication is the culmination of a year-long collaboration with organizations including the Future of Humanity Institute, the Centre for the Study of Existential Risk, and the Electronic Frontier Foundation.
OpenAI News
OpenAI has announced that it is welcoming a new group of donors. The update highlights continued external support for the organization's work.
OpenAI News
OpenAI has introduced a method enabling AI models to teach one another using examples that are interpretable to humans. The system automatically selects the most informative examples for a given concept, proving effective in experimental evaluations for AI instruction.
OpenAI News
OpenAI has developed a system for entity disambiguation that automatically determines which object a word refers to. The approach uses a neural network to evaluate whether the word belongs to each of roughly 100 automatically discovered non-exclusive categories.
OpenAI News
OpenAI has published Requests for Research 2.0, presenting a new batch of seven unsolved problems. These challenges emerged directly from the organization's ongoing research efforts.
OpenAI News