# Nov 7, 2025 ## Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning model **Kimi K2 Thinking** explores advanced methodologies in artificial intelligence, emphasizing the integration of **cognitive processes** to enhance machine learning capabilities. ## Mathematical exploration and discovery at scale **AlphaEvolve**, developed in collaboration with Google DeepMind, utilizes a **large language model (LLM)** to evolve computer code for solving mathematical problems, enhancing traditional optimization methods by focusing on code structure rather than raw input data. ## Open Source Implementation of Apple's Private Compute Cloud **OpenPCC** is an **open-source framework** for **provably private AI inference**, enabling users to run AI models without compromising data privacy through **encrypted streaming** and **unlinkable requests**. ## LLMs encode how difficult problems are **LLMs encode problem difficulty** in a manner that aligns with human judgment, revealing a strong linear decodability of human-labeled difficulty (AMC: **$\rho \approx 0.88$**) across various model sizes, while LLM-derived difficulty shows poor scaling. ## PyTorch Helion **Helion** is a high-level **Python-embedded DSL** that compiles to optimized Triton code, enabling developers to create performant ML kernels without deep hardware knowledge, thus reducing technical debt and enhancing productivity. ## The Parallel Search API The **Parallel Search API** enables AIs to efficiently navigate the web, enhancing their ability to retrieve and process information in real-time, thus revolutionizing AI interactions with online data. ## We built a cloud GPU notebook that boots in seconds **Modal Notebooks** revolutionizes cloud GPU access by enabling **instant boot times** for Jupyter notebooks, enhancing collaborative workflows with real-time features and a focus on high-performance computing. ## Show HN: TabPFN-2.5 – SOTA foundation model for tabular data **TabPFN-2.5** significantly enhances tabular AI, scaling to **20× data cells** compared to its predecessor, and matches the accuracy of complex models like AutoGluon 1.4 while outperforming tuned tree-based models on industry benchmarks. ## From Memorization to Reasoning in the Spectrum of Loss Curvature This study reveals that **memorization** in transformer models can be effectively disentangled through **loss landscape curvature**, indicating that sharper curvature correlates with memorized data, thus enabling a novel weight editing procedure that outperforms existing unlearning methods like **BalancedSubnet**. [Link to article](https://arxiv.org/abs/2510.24256) ## R N TabPFN-2.5 is now available: Tabular foundation model for datasets up to 50k samples **TabPFN-2.5** is a **pretrained transformer** that significantly enhances tabular data processing, now accommodating **50,000 samples × 2,000 features**, a **5x increase** from its predecessor. ## GT – Experimental multiplexing tensor framework for distributed GPU computing **GT** is an **experimental multiplexing tensor framework** that enhances distributed GPU computing by utilizing dynamic scheduling and asynchronous execution, moving away from traditional lock-step paradigms. ## R WavJEPA: Semantic learning unlocks robust audio foundation models for raw waveforms **WavJEPA** is a novel audio foundation model that operates on **raw waveforms**, achieving superior performance in audio representation tasks with significantly less compute and training data compared to models like Wav2Vec2.0 and HuBERT. ## NVIDIA Nemotron Nano V2 VL **Nemotron Nano V2 VL** significantly enhances real-world document understanding and long video comprehension, outperforming its predecessor, Llama-3.1-Nemotron-Nano-VL-8B, through advanced model architecture and innovative training techniques. ## KernelFalcon: Autonomous GPU Kernel Generation via Deep Agents **KernelFalcon** is a pioneering deep agent architecture that autonomously generates GPU kernels, achieving **100% correctness** across all 250 tasks in the [KernelBench](https://github.com/ScalingIntelligence/KernelBench) suite, utilizing a unique combination of hierarchical task decomposition and execution-based verification.