# Jan 31, 2026

## Daily

### Show HN: I trained a 9M speech model to fix my Mandarin tones
- **A 9M-parameter deep learning model** for Mandarin pronunciation training utilizes **CTC loss** to provide precise feedback on tone and pronunciation, outperforming traditional methods that rely on hand-tuning.

### Kimi K2.5 Technical Report 
- **Kimi-K2.5** is a significant project by MoonshotAI, focusing on advanced machine learning techniques, as detailed in their comprehensive [tech report](https://github.com/MoonshotAI/Kimi-K2.5/blob/master/tech_report.pdf).

### Code is cheap. Show me the talk
- **Software development has fundamentally changed** with the advent of LLM coding tools, making traditional methods obsolete and allowing for rapid prototyping and high-quality code generation in mere seconds.

### Show HN: Amla Sandbox – WASM bash shell sandbox for AI agents
- **amla-sandbox** offers a **WASM-based** solution for executing LLM-generated code with **capability enforcement**, eliminating risks associated with arbitrary code execution found in traditional frameworks like LangChain and AutoGen.

### Starlink updates privacy policy to allow consumer data to train
- **Starlink's updated privacy policy** now permits the use of customer data for **AI training**, potentially enhancing Musk's AI initiatives and aligning with a planned IPO that could value SpaceX over **$1 trillion**.

### [P] I solved BipedalWalker-v3 (~310 score) with eigenvalues. The entire policy fits in this post.
- The author achieved a **~310 score** on the **BipedalWalker-v3** environment by utilizing a novel approach with **diagonal weight matrices** and **eigenvalues**, simplifying the model to just **69 lines of Python code**.

### Quack-Cluster: A Serverless Distributed SQL Query Engine with DuckDB and Ray
- **Quack-Cluster** is a **serverless distributed SQL query engine** that utilizes **DuckDB** and **Ray** to execute complex SQL queries directly on data stored in object storage, making it a lightweight alternative to traditional big data systems.

### 175K+ publicly-exposed Ollama AI instances discovered
- **Over 175,000 Ollama AI servers are misconfigured, exposing them to LLMjacking attacks, which exploit these instances for malicious activities like generating spam and malware.**

### Claude Code is your customer
- **Claude Code** signifies a paradigm shift in **SAAS** development, emphasizing the necessity for **API-first** products that cater to AI agents, which have become the primary users of software services.

### [P] Open-Sourcing the Largest CAPTCHA Behavioral Dataset
- **Open-sourced dataset** features **30,000 verified human sessions**, breaking records for scale and providing high-fidelity telemetry crucial for modeling human trajectories in CAPTCHA systems.

### Demystifying ARM SME to Optimize General Matrix Multiplications
- **MpGEMM** is an open-source library that optimizes **General Matrix Multiplication (GEMM)** by leveraging ARM's **Scalable Matrix Extension (SME)**, achieving a notable average speedup of **1.23x** over the Apple Accelerate library.

### Show HN: Pinecone Explorer – Desktop GUI for the Pinecone vector database
- **Pinecone Explorer** is a **native macOS application** designed for managing and exploring the Pinecone vector database, featuring advanced capabilities like **dense, sparse, and hybrid search** along with built-in reranking tools for optimizing query results.

### Autonomous cars, drones cheerfully obey prompt injection by road sign
- **Self-driving cars and drones can be hijacked through custom road signs**, as researchers demonstrated that AI systems interpret these signs as commands, leading to dangerous outcomes like ignoring pedestrians or misdirecting drones.

### Show HN: Foundry – Turns your repeated workflows into one-click commands
- **Foundry** is a self-writing meta-extension for **OpenClaw** that autonomously learns from user workflows, researches documentation, and generates new capabilities, effectively becoming an "agent that builds agents."

### [D] Training Image Generation Models with RL
- **Emerging techniques** like DDPO and DiffusionNFT are advancing **reinforcement learning (RL)** fine-tuning for image generation models, suggesting a shift towards training from scratch using only reward signals.