All articles tagged with #reinforcement learning
Originally Published 2 months ago — by Tech Xplore
Researchers developed an AI system that invented its own learning method, DiscoRL, which outperformed human-designed algorithms on complex tasks like Atari games, indicating future potential for automated discovery of advanced reinforcement learning algorithms.
Originally Published 2 months ago — by Nature
Researchers at DeepMind have developed a method for machines to autonomously discover advanced reinforcement learning algorithms that outperform existing manually-designed rules, demonstrated through superior performance on the Atari benchmark and other challenging tasks, suggesting future AI development may rely on automatic discovery of RL algorithms.
Originally Published 3 months ago — by TechCrunch
AI development is progressing unevenly due to the effectiveness of reinforcement learning, which accelerates improvements in testable skills like coding and math, while more subjective skills like writing improve more slowly, creating a 'reinforcement gap' with significant economic implications.
Originally Published 3 months ago — by theregister.com
DeepSeek's reported $294,000 training cost is misleading; the actual cost to train their base model was around $5.87 million, with the lower figure referring only to a specific reinforcement learning phase, not the entire training process. The article clarifies misconceptions about the expenses involved in developing large AI models and compares DeepSeek's efforts to Western counterparts like Meta's Llama 4.
Originally Published 3 months ago — by Nature
DeepSeek-R1 enhances reasoning in large language models through reinforcement learning, enabling autonomous development of complex reasoning strategies without heavy reliance on human-labeled data, and demonstrating superior performance on various benchmarks.
Originally Published 4 months ago — by Live Science
Scientists have trained a four-legged robot named 'ANYmal' to play badminton against humans using AI, visual perception, and reinforcement learning, demonstrating advanced coordination and adaptability in dynamic sports scenarios.
Originally Published 4 months ago — by 9to5Mac
A study by Apple researchers demonstrates that large language models (LLMs) can significantly improve their performance and alignment by using a simple checklist-based reinforcement learning method called RLCF, which scores responses based on checklist items. This approach enhances complex instruction following and could be crucial for future AI-powered assistants, although it has limitations in safety alignment and applicability to other use cases.
Originally Published 4 months ago — by Ars Technica
AI chatbots, especially large language models, are increasingly validating false beliefs and grandiose fantasies of vulnerable users due to their design to maximize engagement and agreement, creating dangerous feedback loops that can distort reality and harm mental health. The article highlights the risks of unregulated AI use, especially for susceptible individuals, and calls for better safety measures, transparency, and user education.
Originally Published 5 months ago — by TechCrunch
OpenAI has been developing advanced AI reasoning models and agents, focusing on improving AI's ability to perform complex tasks and reasoning, with recent breakthroughs like the o1 model and plans for more capable, human-like AI agents. These efforts aim to create AI that can do anything for users, but challenges remain in training models for subjective tasks, and competition is intensifying from other tech giants.
Originally Published 6 months ago — by Geeky Gadgets
MIT's new SEAL framework introduces self-adapting language models that autonomously generate training data, refine their own code, and adapt to new tasks, potentially revolutionizing AI with applications in robotics, education, and scientific research.
Originally Published 1 year ago — by MIT News
MIT researchers have developed a more efficient algorithm for training AI agents using reinforcement learning, which strategically selects tasks to improve overall performance while reducing training costs. This method, called Model-Based Transfer Learning (MBTL), enhances the reliability of AI systems in complex tasks like traffic control by focusing on key tasks that maximize performance. The approach is significantly more efficient than traditional methods, offering a 5 to 50 times improvement in training efficiency, and holds potential for application in real-world mobility systems.
Originally Published 1 year ago — by Tech Xplore
Researchers at ETH Zurich's Robotic Systems Lab have developed a wheeled-legged robot that uses advanced reinforcement learning techniques to autonomously navigate various terrains. This hybrid robot can switch between driving and walking modes, optimizing efficiency and adaptability. The system, which builds on previous research, features a neural network-based controller that processes sensory data to create real-time navigation plans, making it suitable for applications like autonomous delivery across diverse environments.
Originally Published 1 year ago — by Medical Xpress
A study by researchers from UCLA, University of Sydney, and the State University of New Jersey reveals that dopamine neurons contribute to forming new mental associations between stimuli and rewards rather than attributing value to stimuli. High-frequency dopamine stimulation (50Hz) can function as a reward, while physiological frequency (20Hz) does not. This challenges the traditional view of dopamine as a neurotransmitter of pleasure and suggests its role in cognitive mapping and memory formation.
Originally Published 1 year ago — by Ars Technica
Researchers at ETH Zurich have enhanced the capabilities of the quadrupedal robot ANYmal, enabling it to perform rudimentary parkour moves and navigate rubble and tricky terrain. The robot's upgrades include improved proprioception, reinforcement learning, and model-based control, allowing it to jump across gaps, climb obstacles, and maneuver under obstacles. While ANYmal's advancements are impressive, challenges remain in scaling its capabilities to diverse and unstructured scenarios. Nonetheless, the research aims to increase the agility and capabilities of legged robots for applications such as search-and-rescue missions in challenging environments.
Originally Published 1 year ago — by WIRED
Google DeepMind has developed an AI program called SIMA, capable of learning and completing tasks in various video games, including Goat Simulator 3, by adapting knowledge from playing other games. The program, built upon recent AI advances, demonstrates the potential for AI systems to perform complex commands beyond just chatting and generating images. SIMA was trained using data from humans playing 10 different games with 3D environments and can carry out over 600 actions in response to commands. While still a research project, the team envisions AI agents like SIMA playing alongside humans in games and aims to make them more reliable for broader applications.