# Sep 19, 2025

## Discovering new solutions to century-old problems in fluid dynamics

- **New AI methods** have enabled the discovery of unstable singularities in fluid dynamics equations, potentially transforming our understanding of complex physical phenomena.

## Nvidia buys $5B in Intel

- **Nvidia and Intel** have announced a partnership to develop **Intel x86 RTX SOCs**, integrating Intel CPUs with Nvidia RTX graphics, aimed at enhancing the gaming PC market and custom data center CPUs for AI applications.

## Learn Your Way: Reimagining Textbooks with Generative AI

- **Generative AI (GenAI) enhances textbooks** by creating personalized, multimodal learning experiences, leading to a **9% improvement** in immediate assessments and an **11% increase** in retention compared to traditional methods.

## [R] Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens

- **Chain-of-Thought (CoT) reasoning in LLMs may be a superficial phenomenon**, as it often relies on learned patterns from training data rather than genuine inferential processes, suggesting a need for critical evaluation of its effectiveness.

## Launch HN: Cactus (YC S25) – AI inference on smartphones

- **Cactus** is a **cross-platform framework** enabling local deployment of LLM, VLM, and TTS models in **Flutter** and **React-Native** applications, supporting a wide range of models from **Huggingface**.

## Llama-Factory: Unified, Efficient Fine-Tuning for 100 Open LLMs

- **LLaMA Factory** enables **zero-code fine-tuning** of over **100 large language models** through a user-friendly CLI and web UI, streamlining the process for developers and researchers alike.

## [P] Open dataset: 40M GitHub repositories (2015 → mid-2025) — rich metadata for ML

- A **new open dataset** of **40M GitHub repositories** offers extensive metadata, surpassing existing public snapshots like BigQuery’s **~3M** repositories, and includes a **1M-repo sample** for rapid experimentation.

## [P] We built mmore: an open-source multi-GPU/multi-node library for large-scale document parsing

- **mmore** is an open-source library designed for **multi-GPU/multi-node** document parsing, achieving significant **speed and accuracy** improvements over existing tools like Docling, particularly in handling diverse formats such as PDFs, DOCX, and multimedia files.

## Overcoming accuracy limitations of Analog In-Memory Computing hardware

- **Analog in-memory computing (AIMC)** enhances neural network inference speed and power efficiency but faces challenges like _noisy computations_ and _input/output quantization constraints_, limiting conventional LLM performance on AIMC hardware.

## NVIDIA Blackwell: Born for Extreme-Scale AI Inference

- **NVIDIA's Blackwell architecture** is engineered specifically for **extreme-scale AI inference**, promising enhanced performance and efficiency in processing large datasets.

## [R] Uni-CoT: A Unified CoT Framework that Integrates Text+Image reasoning!

## Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning

- **Convolutional decoding (Conv)** and **Rejecting Rule-based Fine-Tuning (R2FT)** enhance diffusion language models by addressing the **long decoding-window problem**, allowing for faster and more fluent text generation without sacrificing bidirectionality.

## Internalizing Self-Consistency in Language Models: Multi-Agent Consensus Alignment

- **Multi-Agent Consensus Alignment (MACA)** enhances **self-consistency** in language models by utilizing a **reinforcement learning framework** that promotes reasoning aligned with internal consensus through multi-agent debates, leading to improved decision-making.

## Enhancing Retrieval Augmentation via Adversarial Collaboration

- The **Adversarial Collaboration RAG (AC-RAG)** framework enhances **Retrieval-augmented Generation (RAG)** by utilizing a **generalist Detector** and a **domain-specialized Resolver** to combat **Retrieval Hallucinations** effectively.

## RationAnomaly: Log Anomaly Detection with Rationality via Chain-of-Thought and Reinforcement Learning

- **RationAnomaly** introduces a **novel framework** for log anomaly detection that combines **Chain-of-Thought (CoT)** fine-tuning with **reinforcement learning**, enhancing both interpretability and accuracy in software systems.