# Oct 14, 2025

Daily

## First device based on 'optical thermodynamics' can route light without switches
- **Breakthrough in photonics**: Researchers at USC have developed the first optical device utilizing **optical thermodynamics**, allowing light to route itself without switches or external control, guided by thermodynamic principles.

## LLMs are getting better at character-level text manipulation
- **LLMs like GPT-5 and Claude 4.5 show significant improvement in character-level text manipulation**, successfully handling tasks such as character replacement and counting, which were challenging for earlier models.

## NVIDIA DGX Spark In-Depth Review: A New Standard for Local AI Inference
- The **NVIDIA DGX Spark** introduces a compact workstation that delivers **supercomputing-class performance** with a unique **unified memory architecture**, enabling seamless access for both CPU and GPU, which is particularly beneficial for large model inference.

## StreamingVLM: Real-Time Understanding for Infinite Video Streams
- **StreamingVLM** introduces a novel approach for real-time understanding of **infinite video streams**, effectively addressing the challenges of **latency** and **memory usage** through a unified framework that aligns training with streaming inference.

## Prefix sum: 20 GB/s (2.6x baseline)
- **Delta coding** enhances data compression by storing differences between successive values, achieving a remarkable **19.8 GB/s** throughput— **1.8x faster** than naive implementations and **2.6x faster** than FastPFoR, thanks to optimized compute restructuring.

## How AI hears accents: An audible visualization of accent clusters
- **AI models, like BoldVoice's finetuned HuBERT, effectively cluster over 200 accents using a vast dataset of 30 million recordings, revealing unexpected relationships between accents based on geographic and social factors rather than linguistic taxonomy.**

## Beyond the SQLite Single-Writer Limitation with Concurrent Writes
- **Turso's concurrent writes** enhance SQLite's capabilities, achieving up to **4x the write throughput** while eliminating the common `SQLITE_BUSY` error, thus enabling more efficient database operations in modern applications.

## [D] TEE GPU inference overhead way lower than expected - production numbers
- **TEE-secured inference** shows only **7% overhead** compared to standard deployment, a significant improvement for processing sensitive financial documents with LLMs, especially when using Intel TDX on newer Xeon chips.

## Elon Musk Gets Just-Launched NVIDIA DGX Spark: Petaflop AI Supercomputer Lands at SpaceX
- **NVIDIA DGX Spark**, the world's smallest AI supercomputer, delivers **petaflop performance** in a compact design, marking a significant leap in AI capabilities for developers and researchers.

## NVIDIA and Oracle to Accelerate Enterprise AI and Data Processing
- **NVIDIA and Oracle's collaboration introduces the OCI Zettascale10 computing cluster**, leveraging NVIDIA GPUs to achieve **up to 16 zettaflops** of peak AI compute performance, enhancing enterprise AI applications across various sectors.

## All in on MatMul? Don’t Put All Your Tensors in One Basket!
- **The hardware lottery** influences AI research by favoring algorithms that align with dominant hardware, particularly those reliant on **matrix multiplication (MatMul)**, which can stifle innovation in alternative approaches.

## OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators

## NVIDIA, Partners Drive Next-Gen Efficient Gigawatt AI Factories in Buildup for Vera Rubin
- **NVIDIA's Vera Rubin NVL144** is set to revolutionize AI factories with **800 VDC data centers**, supported by over 50 partners, enhancing scalability and energy efficiency for future demands in AI processing.

## Boundary-Guided Policy Optimization for Memory-efficient RL of Diffusion Large Language Models
- **Boundary-Guided Policy Optimization (BGPO)** introduces a **memory-efficient** approach to reinforcement learning (RL) for diffusion large language models (dLLMs) by maximizing a specially constructed lower bound of the ELBO-based objective, addressing the challenge of intractable likelihood functions.

## Iterative Amortized Inference: Unifying In-Context Learning and Learned Optimizers
- **Iterative Amortized Inference** unifies various learning approaches by categorizing them into **parametric, implicit, and explicit regimes**, focusing on how they utilize task-specific information during inference.