Machine Learning

A former OpenAI researcher launches a numerical AI model that sidesteps hallucination and cuts costs by 100x, but can it scale?

PrismML's Bonsai 2 brings reasoning AI to phones with 98% performance in a 5.9GB package.

A leaner AI model reads evolution's clues to find disease-causing DNA variants without the massive compute bill.

A bold claim from OpenAI has put a century-old math problem back in the spotlight, but the real test is just beginning.

Google DeepMind has built India-first AI models that map every farm and track crops in near real time, aiming to close the agricultural data gap.

Caltech scientist Anima Anandkumar joins TIME's 100 most influential people in AI for her work bridging AI and physical science.

Harvard researchers found that AI assistance can subtly erode the judgment of experienced evaluators, raising urgent questions for leaders who rely on LLMs for strategic decisions.

MIT CSAIL study shows large AI image generators often produce outputs with no traceable link to any individual training example, challenging copyright assumptions.

AI firms are buying and destructively scanning second-hand books, erasing the human traces that make them irreplaceable.

Secondhand booksellers in the UK and Ireland are seeing a wave of bulk orders for random books, with many suspecting AI companies are buying them for training data.

During Typhoon Dolphin, China operationally integrated AI weather models from Huawei and Alibaba with traditional systems, achieving faster forecast updates and competitive accuracy, while experts caution about AI's limits in extreme events.

During Typhoon Dolphin, Chinese AI models demonstrated landfall predictions accurate to 30 minutes and 30 km, highlighting the country's lead in AI-driven meteorology.

MIT CSAIL's GeoPT uses synthetic dynamics to give AI a feel for physics, slashing data requirements and speeding up industrial simulations.

59 volunteers built India's first open-access ecoacoustic dataset to fix global AI models that fail on tropical wildlife sounds.

ByteDance founder Zhang Yiming has directed the Seed AI team to stop using AI distillation, betting that foundational research will outpace shortcut methods despite short-term competitive risks.

Anthropic's Claude Fable 5 AI model has helped mathematician Levent Alpöge find a counterexample to the Jacobian conjecture, a problem open since 1939.

Argonne National Laboratory repurposed ChatGPT's transformer architecture to simulate fluid dynamics in nuclear reactors, achieving high accuracy at unprecedented speeds and paving the way for real-time digital twins.

Moonshot AI has released the complete technical architecture for Kimi K3, a 2.8 trillion-parameter multimodal model with a one-million-token context window, but training recipes remain closed.

A cascade architecture for RAG generation starts with cheap local models and escalates to hosted flagships only when validation fails, slashing costs while preserving correctness.

Caltech spinout Oratomic raised $300M to build a quantum computer with just 20,000 qubits by 2030, using AI-driven design and mobile atoms for efficient error correction. The approach could break RSA encryption far sooner than expected.

A new study measures the actual GPU electricity cost of running local LLMs on an RTX 3090, finding that some models are cheaper than cloud APIs while others are surprisingly expensive. The key factor is effective throughput, not model size.

As AI matures, CPUs are emerging as a cost-effective, energy-efficient alternative for inference workloads. This shift could democratize AI deployment across industries.

A Peking University team has created an all-optical link for standard chips, achieving a 100x speedup in distributed AI inference with drastically lower resource use. The breakthrough could reshape data center design amid soaring AI demand.

Despite rapid progress, frontier AI models still hallucinate with alarming confidence. This article dissects recent embarrassing failures, explains the underlying mechanisms, and offers practical mitigation strategies.

Ant Group's AI unit, Robbyant, has developed LingBot-Vision, a model that detects object edges with sub-pixel precision. It outperforms larger models on depth benchmarks while using far less data, potentially reducing costly glass crashes in robotics.

Anthropic researchers have discovered a hidden internal workspace in Claude AI, called J-space, where the model silently processes concepts. The J-lens technique reads these thoughts, revealing safety insights and flexible reasoning.

A CUHK Business School study of 3,137 U.S. firms finds that AI significantly boosts resilience during natural disasters, but its protective effect vanishes in human-caused crises. The research offers a strategic roadmap for deploying AI where it matters most.

Chinese researchers have built a memory chip that models brain structures in real time, achieving speeds up to 478 times faster than Nvidia's A100 GPU. The breakthrough could enable real-time surgical guidance and personalized brain disease treatments.

Suraj Biswas, a former Zomato delivery partner from West Bengal, now leads two AI startups focused on individualized AI. His journey from earning Rs 1,500 a day to challenging AI's one-size-fits-all paradigm is a testament to resilience and vision.

Anthropic is exploring a custom AI chip with Samsung to gain control over computing costs. The early-stage talks could reshape the AI hardware landscape and challenge NVIDIA's dominance.