# Main Article Content

## Recent Highlights

### Show HN: Semble  
- **Semble** is a code search library that enables agents to retrieve relevant code snippets **instantly**, utilizing **~98% fewer tokens** than traditional methods like grep+read, achieving indexing in under a second and query responses in approximately **1.5 ms**.

### GenCAD  
- **GenCAD** is an innovative image-conditional CAD generation model that produces both **3D CAD** and the complete **parameterized CAD command history**, enhancing the design process significantly.

### Project Glasswing: what Mythos showed us  
- **Mythos Preview** represents a significant advancement in vulnerability detection, enabling the construction of **exploit chains** and **proof generation**, which enhances its ability to identify and validate security flaws effectively.

### Anthropic acquires Stainless  
- **Anthropic's acquisition of Stainless** enhances its capabilities in **SDK generation** and **agent connectivity**, crucial for the evolving landscape of AI where agents must act effectively within their environments.

### Reviving PapersWithCode (by Hugging Face)  
- **Hugging Face** is revitalizing **PapersWithCode**, utilizing AI to parse high-impact papers and generate leaderboards, ensuring the platform remains a vital resource for the ML community.

### Voice AI Systems Are Vulnerable to Hidden Audio Attacks  
- **Hidden audio signals** can manipulate AI voice systems, demonstrating vulnerabilities that allow **unheard sounds** to alter model behavior, posing significant security risks.

### Agora-1: The Multi-Agent World Model  
- **Agora-1** introduces a **multi-agent world model** that allows up to four participants, human or AI, to interact in a shared, real-time simulation, enhancing experiences across various fields like gaming and robotics.

### Cutting inference cold starts by 40x with LP, FUSE, C/R, and CUDA-checkpoint  
- **Modal's engineering innovations** have reduced inference cold start times by **40x**, enabling rapid scaling of AI applications through techniques like **cloud buffers**, **custom filesystems**, and **checkpoint/restore** methods for both CPU and GPU memory.

### Fabricked: Misconfiguring Infinity Fabric to Break AMD SEV-SNP  
- **Fabricked** is a novel attack vector targeting **machine learning models**, exploiting vulnerabilities in their training data to manipulate outcomes.

### Recent Developments in LLM Architectures: KV Sharing, mHC, and Compressed Attention  
- **Recent LLM architectures** like Gemma 4 and DeepSeek V4 are innovating with techniques such as **KV sharing** and **compressed attention** to enhance long-context efficiency, significantly reducing memory and computational costs.

### Sub-JEPA: a simple fix to LeCun group's LeWorldModel that consistently improves performance  
- **Sub-JEPA** enhances LeCun's **LeWorldModel** by applying Gaussian regularization within multiple frozen random orthogonal subspaces, improving performance on low-dimensional tasks without introducing new hyperparameters.

### Sense Humans with WiFi – Ruview  
- **Cognitum.One** offers advanced **intelligence solutions** tailored for real-world applications, enhancing decision-making through data-driven insights.

### A Rust-Python thing I am working on. Apache 2 licence  
- **Nairobi OS** is a **high-performance, distributed data science infrastructure** that processes massive datasets efficiently in constrained environments, utilizing a Rust-based refinery daemon and advanced kernel features for optimal performance.

### Alignment pretraining: AI discourse creates self-fulfilling (mis)alignment  
- **Discourse on AI** significantly influences **LLM behavior**, with negative narratives leading to self-fulfilling misalignment; upsampling misalignment discourse increases misaligned behavior, while aligned discourse reduces misalignment scores from **45% to 9%**.

### Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design  
- **AIRA-Compose** and **AIRA-Design** enable LLM agents to autonomously create advanced neural architectures, yielding 14 new models that outperform Llama 3.2 and Composer baselines in accuracy and efficiency.