# Mar 28, 2025

- **Google's A.I. Gemini source code was partially leaked** during a bug bounty event, revealing internal structures and proprietary code, including sensitive **Google3** directories and **internal proto files** that should not have been exposed.

- **Innovative training methods** enable the effective scaling of a **300B Mixture-of-Experts LLM** on lower-performance hardware, achieving results comparable to high-end systems while significantly reducing costs.

- **Claude 3.5 Haiku employs complex internal mechanisms** for tasks like multi-step reasoning and planning, revealing sophisticated strategies that include forward and backward planning, as well as the use of abstract features across various contexts.

- **Recent multi-modal models** like Gemini 2.5 and GPT-4o excel in **native image generation** by integrating advanced image token encoders/decoders with LLM backbones, enhancing their ability to adhere to prompts during both generation and editing tasks.

- The optimized **FP32 matrix multiplication** on AMD RDNA3 GPU achieves a **60% performance increase** over rocBLAS, demonstrating significant improvements through iterative kernel enhancements and architectural insights.

- This work introduces a **novel framework** that utilizes **motion blur** as a valuable cue for **motion estimation**, enabling the prediction of a dense motion flow field and monocular depth map from a single motion-blurred image.

- **Plasmonic modulators** enable the transfer of signal information from electrical to optical waves at unprecedented speeds, potentially revolutionizing **6G networks** and **AI data centers**.

- **Researchers have identified that phasons, low-temperature quasiparticles, facilitate the movement of interlayer excitons in stacked transition metal dichalcogenides (TMDs)** even at temperatures near absolute zero, challenging previous assumptions about exciton behavior.

- **ReaRAG** enhances the **factual accuracy** of Large Reasoning Models (LRMs) by integrating a **novel data construction framework** that limits reasoning chain length and improves decision-making in question answering tasks.

- The **LEGO-Puzzles benchmark** evaluates multimodal language models (MLLMs) on **spatial reasoning tasks**, revealing a significant performance gap between human (85.8%) and AI (59.8%) capabilities, particularly in complex scenarios.

- **AI is transforming 2D engineering drawings into 3D parametric models** through methods like **Text-to-CAD** and **Machine Learning Pipelines**, with companies such as zoo.dev and AdamCad leading the charge in this innovation.

- **Intel Gaudi hardware** is now natively integrated into Text Generation Inference (TGI), enhancing deployment options for Large Language Models (LLMs) and eliminating the need for a separate fork.

- To **train a model** that enhances video quality, focus on techniques that **remove compression artifacts** and generate finer details, leveraging a robust dataset of thousands of videos for effective learning.

- **Collab** introduces a **mixture of agent-based decoding** strategies for aligning Large Language Models (LLMs) at inference time, enhancing adaptability to diverse tasks without the need for retraining.

- **LLM-based generative retrieval (GR)** can produce irrelevant documents, leading to **hallucination** issues that undermine its practical application credibility; this study introduces an optimized GR framework that integrates **knowledge distillation** and a **decision agent** to enhance retrieval accuracy.