# Oct 16, 2025

## ML Times    Oct 16, 2025

### Highlights:

- **Apple M5 chip**  
  **M5** achieves **over 4x peak GPU compute performance** for AI compared to M4, featuring a **10-core GPU** with a **Neural Accelerator** in each core, enhancing both AI and graphics capabilities significantly.

- **Claude Haiku 4.5**  
  **Claude Haiku 4.5** offers **near-frontier coding performance** at **one-third the cost** and **twice the speed** of its predecessor, Claude Sonnet 4, making it a game-changer for real-time AI applications.

- **A Gemma model helped discover a new potential cancer therapy pathway**  
  **Google's C2S-Scale 27B model, with 27 billion parameters, has successfully identified a novel cancer therapy pathway by predicting the effects of silmitasertib in enhancing antigen presentation in immune-context-positive environments.**

- **New Coding Models and Integrations**  
  **Ollama introduces new coding models**: The **GLM-4.6** and **Qwen3-Coder-480B** are now available on Ollama’s cloud service, featuring seamless integration with familiar tools and enhanced performance for tool calling with **Qwen3-Coder-30B**.

- **A kernel stack use-after-free: Exploiting Nvidia's GPU Linux drivers**  
  **Two critical vulnerabilities** in NVIDIA's Linux GPU drivers, CVE-2025-23280 and CVE-2025-23300, allow local unprivileged processes to exploit kernel memory management, confirmed through a proof of concept that achieves kernel read and write primitives.

- **State of AI Report 2025**  
  The **State of AI Report 2025** reveals that **OpenAI** maintains a slight edge in AI development, while **China's DeepSeek** and others are rapidly closing the gap in reasoning and coding tasks, marking a significant shift in global AI leadership.

- **Recursive Language Models (RLMs)**  
  **Recursive Language Models (RLMs)** enable language models to **decompose and recursively interact** with input contexts of unbounded length, significantly improving performance on long-context tasks while mitigating "context rot."

- **SWE-Grep and SWE-Grep-Mini: RL for Fast Multi-Turn Context Retrieval**  
  **SWE-grep and SWE-grep-mini** are newly trained models that achieve **fast context retrieval** in coding tasks, outperforming traditional models by an order of magnitude in speed while maintaining high accuracy.

- **TaxCalcBench: Evaluating Frontier Models on the Tax Calculation Task**  
  **TaxCalcBench** reveals that **state-of-the-art models** can accurately calculate less than **one-third** of federal income tax returns, highlighting significant limitations in current AI capabilities for tax filing.

- **[P] Nanonets-OCR2: An Open-Source Image-to-Markdown Model with LaTeX, Tables, flowcharts, handwritten docs, checkboxes & More**  
  **Nanonets-OCR2** is a cutting-edge model suite that excels in converting images to markdown, featuring capabilities like **LaTeX recognition**, **signature isolation**, and **multilingual support** for diverse document types.

- **[R]: Create a family of pre-trained LLMs of intermediate sizes from a single student-teacher pair**  
  **Boomerang distillation** allows for the creation of a family of pre-trained LLMs of varying sizes by distilling a large teacher model into a smaller student and then reintegrating teacher layers, optimizing both performance and resource efficiency.

- **Closer to production quality Python notebooks with `marimo check`**  
  **marimo check** is a linter designed to enhance the quality of notebooks, pipelines, and apps by providing actionable feedback for both humans and AI agents, ensuring adherence to best coding practices.

- **Generalized Orders of Magnitude**  
  **Generalized Orders of Magnitude (GOOMs)** extend traditional numerical methods, enabling stable computation over **larger dynamic ranges** than conventional floating-point approaches, crucial for fields like deep learning and finance.

- **[R] Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity**  
  **Verbalized Sampling** mitigates **mode collapse** in LLMs by prompting for probability distributions rather than single outputs, enhancing creative task diversity by **2.1x** without sacrificing quality.

- **PyTorch 2.9 Release Blog**  
  **PyTorch 2.9** introduces significant enhancements, including **symmetric memory** for multi-GPU programming and expanded support for **AMD ROCm** and **Intel XPU**, improving performance across diverse hardware platforms.