# Mar 13, 2026

## Can I Run AI locally?

**CanIRun.ai** evaluates whether your machine can effectively run AI models, with a focus on **Meta's Llama 3.1 8B**, which boasts a **great quality/speed ratio** for diverse applications.

## Document poisoning in RAG systems: How attackers corrupt AI's sources

**Document poisoning** in RAG systems can mislead AI outputs by injecting fabricated documents, as demonstrated by a successful attack that altered a company's reported revenue from **$24.7M to $8.3M** using just three documents in a local setup.

## Are LLM merge rates not getting better?

**LLMs have not improved in programming abilities for over a year**, as evidenced by a lack of increase in merge rates since early 2025, contradicting claims of ongoing advancements in AI capabilities.

## Run NanoClaw in Docker Sandboxes

**NanoClaw now integrates seamlessly with Docker Sandboxes**, allowing users to run agents in isolated environments with a single command, enhancing security and ease of use.

## Launch HN: IonRouter (YC W26) – High-throughput, low-cost inference

**IonRouter** delivers **high throughput** and **low-cost inference** through its **IonAttention engine**, achieving a throughput of **7,167 tokens per second** on a single GH200 GPU, significantly outperforming traditional inference providers which average around **3,000 tokens per second**.

## Forcing Flash Attention onto a TPU and Learning the Hard Way

**Porting Flash Attention to TPU** revealed that while the algorithm is sound, the TPU's architecture and XLA's optimization capabilities often outperform manual implementations, especially for single-head attention tasks.

## [R] LEVI: Beating GEPA/OpenEvolve/AlphaEvolve at a fraction of the cost

**LEVI** achieves superior performance in LLM-guided evolutionary optimization by utilizing **stratified model allocation** and **fingerprint-based CVT-MAP-Elites**, enabling it to outperform competitors like GEPA and OpenEvolve at a fraction of the cost.

## Exploring JEPA for real-time speech translation

**JEPA-v0** is a **self-supervised audio encoder** designed to enhance real-time speech translation by preserving **voice, emotion, and timing**, addressing limitations of traditional cascaded translation models that discard paralinguistic features.

## [R] Beyond Prediction - Text Representation for Social Science (arxiv 2603.10130)

**Text representations** in ML/NLP must bridge the **prediction–measurement gap**, as effective tools for prediction may fail as scientific instruments in social science contexts.

## [P] Visual verification as a feedback loop for LLM code generation

The **autonomous pipeline** generates playable Godot games from text prompts, addressing the challenge of LLMs writing correct code in underrepresented languages like GDScript, which lacks sufficient training data for reliable API usage.

## [R] HoloPASWIN: Integrating Physics into Swin Transformers for Holographic Reconstruction (Code/Dataset/Paper)

**HoloPASWIN** employs **Swin Transformers** to effectively address the "twin-image" problem in lensless in-line holography, enhancing the model's ability to capture long-range dependencies in diffraction patterns.

## [P] ColQwen3.5-v2 4.5B is out!

**ColQwen3.5-v2** is a **4.5B parameter** visual document retrieval model that outperforms its predecessor by utilizing a **simplified training recipe** with only **two phases** instead of four, enhancing efficiency and results.

## 🤗How NVIDIA AI-Q Reached #1 on DeepResearch Bench I and II

**NVIDIA AI-Q** achieved **first place** on both [DeepResearch Bench I](https://github.com/Ayanami0730/deep_research_bench) and [DeepResearch Bench II](https://github.com/imlrz/DeepResearch-Bench-II) by leveraging a **multi-agent architecture** that enhances research quality through modular design and fine-tuned models.

## Into the Omniverse: How Industrial AI and Digital Twins Accelerate Design, Engineering and Manufacturing Across Industries

**Industrial AI and digital twins** are revolutionizing design and manufacturing by enabling rapid simulation and optimization, as seen in partnerships like that of NVIDIA and Dassault Systèmes, which leverage **AI physics** and **virtual twin technology** for enhanced product development.

## 🤗Build an Agent That Thinks Like a Data Scientist: How We Hit #1 on DABStep with Reusable Tool Generation

The **NVIDIA KGMON (NeMo Agent Toolkit) Data Explorer** achieved **1st place** on the DABStep benchmark, demonstrating a **30x speedup** over the baseline by employing a multi-phase approach that separates foundational knowledge from rapid inference.