Nicheloom

The opportunity tracker for new startups.

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.

Source

Headline volume, last 14 days

Showing 9651–9700 of 9720 headlines.

Block-sparse GPU kernels Dec. 6, 2017, 8 a.m.

OpenAI News

OpenAI has released highly optimized GPU kernels designed for neural networks with block-sparse weights. Depending on the sparsity level, these kernels can run orders of magnitude faster than cuBLAS or cuSPARSE. They have already achieved state-of-the-art results in text sentiment analysis as well as image and text generation.

Learning a hierarchy Oct. 26, 2017, 7 a.m.

OpenAI News

Researchers have developed a hierarchical reinforcement learning algorithm that discovers high-level actions to solve complex tasks spanning thousands of timesteps. In navigation tests, the algorithm uncovered walking and crawling actions across various directions. These learned capabilities allowed the agent to master new navigation tasks quickly.

Generalizing from simulation Oct. 19, 2017, 7 a.m.

OpenAI News

New robotics techniques allow robot controllers trained entirely in simulation to adapt to unexpected environmental changes when deployed on physical hardware. By creating closed-loop systems instead of earlier open-loop designs, the robots can react dynamically while completing simple tasks.

Meta-learning for wrestling Oct. 11, 2017, 7 a.m.

OpenAI News

In a simulated robot wrestling environment, a meta-learning agent demonstrated that it could quickly defeat a stronger non-meta-learning opponent. Researchers also showed that the meta-learning agent is capable of adapting to physical malfunctions.

Competitive self-play Oct. 11, 2017, 7 a.m.

OpenAI News

Research shows that competitive self-play enables simulated AIs to discover physical skills such as tackling, kicking, ducking, and catching without explicit environment design. Because self-play naturally maintains an appropriate difficulty curve, OpenAI expects it to become a foundational component of future advanced AI systems.

Learning to model other minds Sept. 14, 2017, 7 a.m.

OpenAI News

OpenAI has introduced Learning with Opponent-Learning Awareness (LOLA), an algorithm that accounts for the learning processes of other agents. In the iterated prisoner's dilemma, LOLA successfully discovered self-interested yet collaborative strategies like tit-for-tat, marking progress toward systems that model other minds.

OpenAI Baselines: ACKTR & A2C Aug. 18, 2017, 7 a.m.

OpenAI News

OpenAI has released two new reinforcement learning implementations, A2C and ACKTR, as part of OpenAI Baselines. A2C provides a synchronous, deterministic alternative to A3C that matches its performance. ACKTR offers superior sample efficiency compared to TRPO and A2C while requiring only slightly more computation per update than A2C.

More on Dota 2 Aug. 16, 2017, 7 a.m.

OpenAI News

OpenAI's Dota 2 project demonstrated that self-play paired with sufficient compute can elevate machine learning systems from sub-human to superhuman performance. Within a month, the system advanced from matching high-ranked players to defeating top professionals. Unlike traditional supervised learning that is limited by static datasets, self-play automatically generates increasingly high-quality training data as the agent improves.

Dota 2 Aug. 11, 2017, 7 a.m.

OpenAI News

OpenAI has developed a bot capable of defeating the world's top professional players in 1v1 Dota 2 matches under tournament rules. The system learned the game entirely from scratch through self-play, avoiding the use of imitation learning or tree search. The project marks a step forward in creating AI that can achieve clear objectives within complex, human-involved environments.

Gathering human feedback Aug. 3, 2017, 7 a.m.

OpenAI News

OpenAI has open-sourced RL-Teacher, an interface that trains AI systems using periodic human feedback instead of manually crafted reward functions. Designed to advance AI safety, the method is especially useful for reinforcement learning tasks where rewards are difficult to specify directly.

Better exploration with parameter noise July 27, 2017, 7 a.m.

OpenAI News

OpenAI researchers found that introducing adaptive noise to the parameters of reinforcement learning algorithms regularly improves performance. The exploration technique is easy to implement and rarely degrades results, making it widely applicable.

Proximal Policy Optimization July 20, 2017, 7 a.m.

OpenAI News

OpenAI has released Proximal Policy Optimization (PPO), a reinforcement learning algorithm that matches or exceeds the performance of state-of-the-art approaches. Due to its ease of implementation and tuning, PPO has become the default reinforcement learning method used across OpenAI.

Robust adversarial inputs July 17, 2017, 7 a.m.

OpenAI News

Researchers have generated adversarial images that consistently fool neural network classifiers across different scales and viewing angles. These findings challenge recent assertions that self-driving cars would be naturally resilient to adversarial attacks due to their multi-angle and multi-perspective vision systems.

Faster physics in Python June 28, 2017, 7 a.m.

OpenAI News

OpenAI has open-sourced a high-performance Python library designed for robotic simulation using the MuJoCo engine. The library was developed over the course of a year of robotics research.

Learning from human preferences June 13, 2017, 7 a.m.

OpenAI News

To advance AI safety and eliminate the risks of poorly defined goal functions, OpenAI has collaborated with DeepMind's safety team. Together, they developed an algorithm capable of inferring human intent simply by receiving feedback on which of two proposed behaviors is preferable.

Learning to cooperate, compete, and communicate June 8, 2017, 7 a.m.

OpenAI News

OpenAI highlights multiagent competitive environments as vital stepping stones toward achieving artificial general intelligence. These environments provide an intrinsic curriculum based on opponent skill and lack a stable equilibrium, constantly driving agents to become smarter.

OpenAI Baselines: DQN May 24, 2017, 7 a.m.

OpenAI News

OpenAI has announced the open-source release of OpenAI Baselines, an internal initiative to reproduce reinforcement learning algorithms that match published performance benchmarks. The rollout begins with Deep Q-Networks (DQN) and three of its variants, with additional algorithms planned for release in the coming months.

Robots that learn May 16, 2017, 7 a.m.

OpenAI News

OpenAI has developed a robotic system capable of learning a new task after observing it just once. The system was trained entirely within a simulation before being deployed onto a physical robot.

Roboschool May 15, 2017, 7 a.m.

OpenAI News

OpenAI has released Roboschool, an open-source software platform for robot simulation. The tool is integrated directly with OpenAI Gym.

Unsupervised sentiment neuron April 6, 2017, 7 a.m.

OpenAI News

OpenAI has built an unsupervised system that develops an effective representation of sentiment. Notably, the model achieved this despite only being trained to predict the next character in Amazon review texts.

Spam detection in the physical world April 1, 2017, 7 a.m.

OpenAI News

OpenAI has developed an AI system capable of detecting Spam in the physical world. The system was trained entirely in a simulated environment before being deployed on a physical robot.

Evolution strategies as a scalable alternative to reinforcement learning March 24, 2017, 7 a.m.

OpenAI News

OpenAI researchers have discovered that evolution strategies, a decades-old optimization technique, rival standard reinforcement learning methods on benchmarks such as Atari and MuJoCo. Additionally, the approach overcomes many of the practical inconveniences associated with standard reinforcement learning.

Distill March 20, 2017, 7 a.m.

OpenAI News

OpenAI has announced its support for the launch of Distill. Distill is a new journal dedicated to promoting clear and effective communication of machine learning research, covering both novel and existing results.

Learning to communicate March 16, 2017, 7 a.m.

OpenAI News

OpenAI has published research exploring multi-agent systems. The work outlines how artificial agents are able to develop their own language to communicate.

Attacking machine learning with adversarial examples Feb. 24, 2017, 8 a.m.

OpenAI News

OpenAI is examining adversarial examples, which are inputs intentionally crafted by attackers to deceive machine learning models like optical illusions. The post explores how these adversarial inputs function across various mediums and highlights the challenges of securing AI systems against them.

Team update Jan. 30, 2017, 8 a.m.

OpenAI News

OpenAI announced that its team has grown to 45 people. The organization continues to advance AI capabilities through new software systems, novel research, and robotic deployments.

Faulty reward functions in the wild Dec. 21, 2016, 8 a.m.

OpenAI News

OpenAI examines how reinforcement learning algorithms can fail in counterintuitive ways. The post specifically focuses on the failure mode that occurs when a reward function is misspecified.

Universe Dec. 5, 2016, 8 a.m.

OpenAI News

OpenAI has released Universe, a software platform designed to train and evaluate an AI's general intelligence. The platform enables systems to learn across a broad spectrum of real-world games, websites, and applications.

OpenAI and Microsoft Nov. 15, 2016, 8 a.m.

OpenAI News

OpenAI has announced a partnership with Microsoft. Under the collaboration, OpenAI will begin running most of its large-scale experiments on the Azure platform.