# Feb 16, 2026

## Daily

### Qwen3.5: Towards Native Multimodal Agents

- **Qwen** operates across multiple domains, including **qwen.ai** and **chat.qwen.ai**, indicating a robust infrastructure for AI interactions and services. [Link to article](https://qwen.ai/blog?id=qwen3.5)

### Audio is the one area small labs are winning

- **Gradium and Kyutai are pioneering audio AI** with groundbreaking models like Moshi, which is the first full-duplex conversational AI capable of real-time interaction, outperforming larger labs despite limited resources.

### Two different tricks for fast LLM inference

- **OpenAI's fast mode** achieves over **1000 tokens per second**, making it **15x faster** than its previous model, while Anthropic's fast mode offers **2.5x** speed, but with the actual Opus 4.6 model, ensuring higher capability despite lower speed.

### Towards Autonomous Mathematics Research

- **Aletheia** is an advanced math research agent that autonomously generates, verifies, and revises solutions, demonstrating capabilities from **Olympiad** problems to **PhD-level** exercises, powered by Gemini Deep Think and novel inference-time scaling laws.

### Expensively Quadratic: The LLM Agent Cost Curve

- **Cache reads dominate costs** in LLM conversations, becoming **87%** of total expenses by **50,000 tokens**, highlighting the quadratic nature of these costs as conversations progress.

### Show HN: Klaw.sh – Kubernetes for AI agents

- **klaw** is an **open-source platform** designed to deploy and manage AI agents at scale, akin to Kubernetes but focused on intelligent agents capable of coding, researching, and automating tasks.

### \\[D\\] We found 18K+ exposed OpenClaw instances and ~15% of community skills contain malicious instructions

- Over **18,000 OpenClaw instances** are publicly exposed, with nearly **15% of community skills** harboring malicious instructions that can download malware or exfiltrate sensitive data.

### Study: Self-generated Agent Skills are useless

- **SkillsBench** evaluates **86 tasks** across **11 domains**, revealing that curated Skills improve average pass rates by **16.2 percentage points**, with significant variability across domains, from **+4.5pp** in Software Engineering to **+51.9pp** in Healthcare.

### Show HN: Solving Sudoku reasoning via Energy Geometric models

- **Curvature-Guided Wavefront Execution** leverages intrinsic Riemannian curvature for a GPU-accelerated solver, achieving a **1,226× speedup** over traditional Python CPU methods, effectively solving the hardest Sudoku puzzles at **270,000 per second**.

### Chiplets Get Physical: The Days of Mix-and-Match Silicon Draw Nigh

- **Chiplet technology is evolving**, with Cadence introducing a **Physical AI Chiplet Platform** designed for real-time AI processing in edge applications, marking a significant shift towards modular, mix-and-match silicon solutions.

### Show HN: I generated a "stress test" of 200 rare defects from 7 real photos

- **Generated 200 rare defect variations** from 7 real-world samples to enhance model validation for structural inspection, addressing the lack of a reliable "Golden Set" for rare failures.

### New SemiAnalysis InferenceX Data Shows NVIDIA Blackwell Ultra Delivers up to 50x Better Performance and 35x Lower Costs for Agentic AI

- The **NVIDIA Blackwell Ultra platform** achieves **up to 50x better performance** and **35x lower costs** for agentic AI, driven by advanced hardware and software optimizations that enhance throughput and efficiency.

### TRACE: Temporal Reasoning via Agentic Context Evolution for Streaming Electronic Health Records (EHRs)

- **TRACE** enhances **temporal clinical reasoning** in **longitudinal patient trajectories** by structuring context without the need for fine-tuning or retrieval augmentation, thus addressing the limitations of existing models.

### GRAIL: Geometry-Aware Retrieval-Augmented Inference with LLMs over Hyperbolic Representations of Patient Trajectories

- **GRAIL** enhances next-visit event prediction in clinical settings by utilizing **structured geometric representations** of electronic health records (EHRs) to improve accuracy and reduce hallucinations in large language models (LLMs).

### HyperMLP: An Integrated Perspective for Sequence Modeling

- **HyperMLP** and **HyperGLU** present a novel perspective on autoregressive attention, framing it as a dynamic two-layer MLP that utilizes context history for weight instantiation, diverging from traditional probabilistic models.