Reports indicate that Anthropic, a prominent AI development company, has acquired biotech startup Coefficient Bio in a deal valued at approximately $400 million. This acquisition signifies a growing trend of AI technology expanding its applications into the healthcare and life sciences sectors.

By integrating Coefficient Bio's expertise, Anthropic is likely to accelerate the application of AI in areas such as drug discovery, protein design, or other biological research. This move underscores the increasing influence of AI companies beyond traditional software development into broader industrial domains.

👉 Read the full article on TechCrunch

Major AI companies, including Microsoft, Meta, and Google, are reportedly constructing large natural gas power plants to fuel their energy-intensive data centers. This development raises significant concerns about the environmental impact of the rapid advancement of AI technology.

The immense power demands for training and operating AI models are leading to a reliance on natural gas, potentially contradicting corporate sustainability goals amidst calls for a transition to renewable energy. The article delves into the potential negative environmental consequences of the AI boom and the issues arising from a 'FOMO' (Fear Of Missing Out) driven energy strategy.

👉 Read the full article on TechCrunch

This paper affirmatively answers whether a Large Language Model (LLM) can improve its code generation capabilities using only its own raw outputs, without the need for a verifier, a teacher model, or reinforcement learning. The proposed method, termed Simple Self-Distillation (SSD), involves sampling solutions from the model with specific temperature and truncation configurations, followed by standard supervised fine-tuning on these generated samples.

SSD significantly improved the Qwen3-30B-Instruct's pass@1 score on LiveCodeBench v6 from 42.4% to 55.3%, with gains concentrated on more challenging problems. The method demonstrates generalization across Qwen and Llama models at 4B, 8B, and 30B scales, including both instruct and thinking variants. The effectiveness of this simple approach is attributed to its ability to resolve a precision-exploration conflict in LLM decoding. SSD reshapes token distributions in a context-dependent manner, suppressing distractor tails where precision is crucial while preserving useful diversity where exploration is beneficial. Consequently, SSD offers a complementary post-training direction for enhancing LLM code generation.

👉 Read the full article on arXiv