# Dec 24, 2025

### Nvidia buying AI chip startup Groq for about $20B in cash  
- **Nvidia's acquisition of Groq for $20 billion marks its largest purchase ever**, significantly expanding its portfolio in AI chip technology, particularly as Groq's founders were instrumental in developing Google's TPU, a direct competitor to Nvidia's offerings.

### Local AI is driving the biggest change in laptops in decades  
- **Local AI models are set to revolutionize laptops**, as the integration of **neural processing units (NPUs)** will enable devices to run large language models (LLMs) efficiently, enhancing performance and privacy by processing data locally rather than relying on cloud services.

### Google 2025 recap: Research breakthroughs of the year  
- In **2025**, Google achieved significant advancements in AI with models like **Gemini 3**, enhancing reasoning, multimodality, and efficiency, which are set to revolutionize products and scientific research.

### Executorch: On-device AI across mobile, embedded and edge for PyTorch  
- **ExecuTorch** enables seamless deployment of AI models on-device, supporting a range of applications from smartphones to microcontrollers, and is utilized by **Meta** for real-time inference across its platforms like **Instagram** and **WhatsApp**.

### I rebuilt FlashAttention in Triton to understand the performance archaeology  
- **Flash Attention** optimizes memory usage and performance in deep learning by restructuring the attention mechanism to operate with **O(S)** memory complexity, significantly reducing the need for high bandwidth memory (HBM) access, which is a common bottleneck in modern GPUs.

### \\D\\2025 Year in Review: The old methods quietly solving problems the new ones can't  
- **Old methods** like **HMMs** and **Viterbi** are resurfacing to tackle problems that **Transformers** struggle with, demonstrating that foundational techniques remain relevant in modern NLP challenges.

### Microsoft Agent Framework  
- The **Microsoft Agent Framework** is an open-source toolkit for developing **AI agents** and **multi-agent workflows** in .NET and Python, enhancing capabilities from the **Semantic Kernel** and **AutoGen** projects while introducing new features for complex task management.

### \\[P\\] SIID: A scale invariant pixel-space diffusion model; trained on 64x64 MNIST, generates readable 1024x1024 digits for arbitrary ratios with minimal deformities (25M parameters)  
- **SIID (Scale Invariant Image Diffuser)** utilizes a novel architecture that allows for **readable 1024x1024 digit generation** from 64x64 MNIST training, achieving this with **25M parameters** and minimal artifacts, demonstrating significant scalability in image resolution.

### \\[P\\] TraceML Update: Layer timing dashboard is live + measured 1-2% overhead on real training runs  
- **TraceML's new dashboard** provides a **layer-by-layer timing breakdown**, allowing users to pinpoint where training time is spent on GPU vs CPU, enhancing training efficiency.

### \\[P\\] RewardScope - reward hacking detection for RL training  
- **RewardScope** addresses the critical issue of **reward hacking** in reinforcement learning (RL) by providing real-time monitoring of reward components, enabling detection of anomalies like state cycling and reward spiking.

### FaithLens: Detecting and Explaining Faithfulness Hallucination  
- **FaithLens** is a novel model designed to detect **faithfulness hallucination** in large language models (LLMs), enhancing trustworthiness through binary predictions and explanations, which are generated from synthesized training data using advanced LLMs.

### 🤗AprielGuard: A Guardrail for Safety and Adversarial Robustness in Modern LLM Systems  
- **AprielGuard** is an **8B parameter model** that detects **16 categories of safety risks** and a wide range of **adversarial attacks**, enhancing the safety and robustness of modern LLM systems.

### Google's year in review: 8 areas with research breakthroughs in 2025  
- In **2025**, Google achieved significant advancements in AI with the introduction of **Gemini 3** and **Gemma 3**, enhancing reasoning, multimodality, and efficiency across its product suite.

### \\[R\\] Policy→Tests (P2T) bridging AI policy prose to executable rules  
- **Policy→Tests (P2T)** transforms natural-language AI governance documents into **machine-readable rules**, addressing the bottleneck in enforcement and evaluation tools that require explicit conditions and exceptions.

### Fail Fast, Win Big: Rethinking the Drafting Strategy in Speculative Decoding via Diffusion LLMs  
- **Diffusion Large Language Models (dLLMs)** can enhance speculative decoding by leveraging their **speed** to minimize costly rejections, thus enabling the generation of longer drafts more efficiently.