Machine Learning

Breakthrough in AI: Fine-Tuning LLMs with Less Data and Power Researchers at the University of California San Diego have pioneered a new method for fine-tuning large language models (LLMs) that requires significantly less data and computing power. This innovative technique avoids retraining a model’s entire parameter set, instead selectively updating only the most critical components. This approach dramatically reduces costs and minimizes the risk of overfitting. In tests on protein language models, the method achieved high accuracy with limited training data, proving to be 326 times more parameter-efficient than traditional fine-tuning. This breakthrough is poised to democratize AI, enabling smaller organizations with limited resources to customize powerful AI models for specialized applications. The findings were published in Transactions on Machine Learning Research. ...

Google DeepMind Reveals AI Agents with Human-Like Computer Skills Google’s DeepMind has made a significant advancement in agentic AI, developing agents that can interact with computers in a manner similar to humans. This development is seen as a major step forward in creating more capable and autonomous AI systems. Researchers at the Bay Area Machine Learning Symposium also discussed the future of AI, with a focus on creating models that can learn through interaction and exploration rather than just being trained on massive datasets. The goal is to achieve systematic generalization, allowing AI to combine known elements into novel meanings and learn with significantly less data than current models require. Google DeepMind’s work on models for more intelligent robots, including the recently released Gemini Robotics 1.5 and E.R. 1.5, further illustrates this push towards more interactive and intelligent AI. ...

Today’s Top Tech Stories: Waymo, NASA, Oracle, and Quantum Breakthroughs Here’s a summary of the most significant developments in AI, autonomous driving, cloud infrastructure, and quantum computing for October 15, 2025. Waymo Announces 2026 London Robotaxi Service Launch Waymo, the autonomous driving technology company under Alphabet, has officially announced its plan to launch a driverless robotaxi service in London. The full rollout is scheduled for 2026, making London the first European city to host Waymo’s autonomous ride-hailing service. Initial vehicle testing with safety drivers will commence in the upcoming weeks. The service, accessible via the Waymo app, will directly compete with established ride-hailing platforms. This expansion represents Waymo’s second international market after Tokyo. UK Transport Secretary Heidi Alexander voiced support, emphasizing the initiative’s potential to advance the UK’s autonomous vehicle sector and enhance transportation accessibility. ...

NVIDIA Unveils RLP: A New Era of Reinforcement Learning for AI Pretraining NVIDIA researchers have developed a groundbreaking training objective called Reinforcement Learning Pretraining (RLP), which integrates reinforcement learning directly into the pretraining stage of AI model development. The core concept of RLP treats a short chain-of-thought as an action that is rewarded based on the information gain it provides for the model’s next-token prediction. This innovative method creates a verifier-free and dense reward system that can be applied at scale to standard text streams. Unlike previous methods that depend on sparse signals or external verifiers, RLP’s approach allows for updates at every single token position. Experiments demonstrate that RLP can yield significant improvements in a model’s reasoning capabilities, with one model showing a substantial increase in its average score on scientific reasoning benchmarks. ...

Robotics in Action: Midlakes Schools Automate Field and Pool Maintenance Midlakes Schools are leveraging robotic technology to enhance the efficiency of maintaining their athletic fields and swimming pool. A Tiny Mobile Robot Pro X now paints lines on sports fields using precise GPS-guided templates, while a Wave 200XL robotic vacuum handles the cleaning of the new pool surface. Ryan Tilburg, the Director of Facilities, notes that these robots are generating significant time savings on the athletic fields and conserving water in the pool. This implementation of automation helps the maintenance staff better manage their workload, especially during peak summer months, ensuring maximum availability of the school’s facilities. ...

Infineon Launches Advanced Power Protection for AI Data Centers Infineon Technologies AG has introduced a new line of power path protection solutions engineered for the increasing demands of AI data center infrastructures. The new products include a family of 48-volt smart eFuses and a reference board for hot-swap controllers for 400V and 800V power architectures. These components are designed to improve the reliability, scalability, and monitoring of energy flow in high-performance computing environments. The company stated that these solutions aim to maximize server uptime, reduce operational costs, and extend the lifespan of critical equipment in AI data centers. ...

S&P Global and IBM Launch Agentic AI to Revolutionize Supply Chains S&P Global and IBM have announced a partnership to integrate IBM’s watsonx Orchestrate agentic AI framework into S&P Global’s offerings, starting with supply chain management. This collaboration aims to provide businesses with AI-powered tools to address global supply chain challenges by enhancing visibility and insights into vendor selection. S&P Global will develop new AI agents for IBM’s watsonx Orchestrate Agent Catalog, utilizing its proprietary data and analytics to help accelerate decision-making. The partnership is a key part of S&P Global’s strategy to use trusted AI agents to transform its enterprise services. Beyond the initial supply chain focus, S&P Global intends to integrate IBM’s AI technology into other risk intelligence services. ...

Google Unveils CodeMender AI to Automatically Fix Software Vulnerabilities Google DeepMind has introduced CodeMender, a groundbreaking AI agent designed to automatically detect, patch, and rewrite vulnerable code to thwart security exploits. This powerful cybersecurity tool operates both reactively, addressing new threats as they emerge, and proactively, hardening existing codebases to eradicate entire classes of vulnerabilities. Powered by Google’s advanced Gemini AI models, CodeMender can debug, identify, and resolve security flaws by targeting their root cause and validating patches to prevent regressions. Alongside this launch, Google has initiated an AI Vulnerability Reward Program to incentivize the discovery of AI-related issues in its products, reinforcing its commitment to AI-driven security. ...

MIT Unveils TX-GAIN, The Most Powerful University AI Supercomputer MIT’s Lincoln Laboratory has launched the TX-Generative AI Next (TX-GAIN) computing system, now the most powerful AI supercomputer at any U.S. university. Optimized for generative AI, TX-GAIN is set to drive innovation in critical research areas such as biodefense, materials discovery, and cybersecurity. With a peak performance of two AI exaflops, the supercomputer will support a wide range of research and development projects at Lincoln Laboratory and across the MIT campus. Researchers are already utilizing TX-GAIN to model complex protein interactions for biological defense, demonstrating its significant impact on scientific breakthroughs. The system is housed in an energy-efficient data center in Holyoke, Massachusetts. ...

Cornell’s ME-AI Model Captures Human Intuition for Quantum Materials Discovery Researchers at Cornell University have introduced a novel machine-learning model, Materials Expert-Artificial Intelligence (ME-AI), designed to quantify and encapsulate the intuition of human experts to accelerate the discovery of new quantum materials. The model effectively “bottles” expert knowledge into descriptors that can accurately predict a material’s functional properties. In a proof-of-concept test, ME-AI was trained on expert-curated data to identify a specific desirable characteristic across 879 materials. The model not only reproduced the human expert’s intuition but also expanded upon it, proving its ability to generalize predictions to other compound sets. This research establishes a powerful model for future collaborations between materials scientists and AI experts to push the boundaries of scientific discovery. ...