Chinese AI start-up Moonshot AI launched a new reasoning AI model, k0-math, which outperformed OpenAI's models in Chinese mathematical examinations and competition-level problem sets. The model uses enhanced learning and chain-of-thought reasoning but has limitations with geometry problems and tends to overthink simple maths problems.
MIT researchers developed 'LucidSim', a platform combining simulation software and AI to train robot dogs. Dogs trained this way successfully performed tasks like chasing balls and avoiding obstacles. The method, which used ChatGPT to generate training descriptions, may offer faster and more accurate robot training.
AI and machine learning are revolutionising scientific prediction across fields like college admissions, elections, and drug discovery. Large datasets and complex 'black box' models are being used, with statisticians developing techniques to quantify uncertainty without understanding the models' inner workings. Concerns about reproducibility exist, and statisticians are creating safeguards for reliable findings. The emerging field of data science incorporates traditional statistics with new techniques like large-scale population tracking.
Researchers from top universities questioned whether large language models (LLMs) truly understand the world. Using new metrics, they tested LLMs on city navigation and a board game. Despite high accuracy in normal conditions, LLMs struggled with unexpected changes and created impossible scenarios. This suggests LLMs may perform well without grasping underlying rules, raising concerns for real-world AI deployment. The study will be presented at an upcoming conference.
An AI Image Fabrication detector has been developed to identify AI-generated images in scientific publications. The tool claims a 98% success rate with a 0.02% false-positive rate and can detect AI-generated microscopy images. Science journal is using the tool, while Springer Nature is developing its own detection tools. Regular updates are planned to recognise images from new AI models.
China aims to become a global AI leader, with goals to achieve major breakthroughs by 2025 and become the world's primary AI innovation centre by 2030. The strategy involves investing in AI education, talent development, and supporting Chinese AI enterprises. China leads in AI publications and patent filings, and plans to leverage international initiatives to expand its AI influence globally. The country also aims to actively participate in and influence global AI standards and governance.
Researchers used GPT-4 to teach low-cost robot arms to clean up spills. The experiment used LangChain to translate the AI's input into robot movements. This showcases AI language models' potential for physical tasks and demonstrates how open-source technology can make robotics more accessible.
Meta FAIR and university researchers developed Thought Preference Optimization (TPO), a method to improve large language model responses. TPO enhances the model's internal thought process before answering, leading to more accurate and coherent outputs. It outperforms baseline models on benchmarks and shows potential for various applications, including creative tasks.
Quantum Machines and Nvidia have made progress in quantum computing control by using reinforcement learning to improve qubit calibration. This approach, demonstrated on a Rigetti quantum chip, enables continuous calibration and could lead to significant gains in quantum error correction. The successful collaboration shows potential for applying this method to more complex quantum systems.
OpenAI won't release GPT-5 this year, focusing instead on GPT-o1 for specialised research. Future ChatGPT advancements may include task-performing 'agents' and a search interface. The company is prioritising urgent developments due to compute limitations.
Google has launched Illuminate, an experimental AI tool that converts academic research papers into AI-generated audio conversations. Currently optimised for computer science papers from arxiv.org, it aims to make complex research more accessible. Users with Google accounts can generate up to five audio discussions per day, which are automatically deleted after 30 days unless saved.
