Shuxin Chen

𝐅𝐨𝐫 𝐍𝐨𝐧-𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐞𝐫𝐬, 𝐭𝐡𝐞 𝐭𝐞𝐫𝐦𝐢𝐧𝐚𝐥 𝐢𝐬 𝐭𝐡𝐞 𝐛𝐚𝐫𝐫𝐢𝐞𝐫. 𝐈 𝐫𝐞𝐦𝐨𝐯𝐞𝐝 𝐢𝐭. 𝐯𝟎.𝟏 𝐢𝐬 𝐨𝐩𝐞𝐧 𝐬𝐨𝐮𝐫𝐜𝐞 𝐭𝐨𝐝𝐚𝐲. White paper: https://lnkd.in/gH3u3KPp Website: https://lnkd.in/gtfjxGa2 GitHub: https://lnkd.in/gABj_9eE --- Most local AI tools have the same problem: they're built for engineers. You need a terminal, a shell script, and patience. Everyone else is left out. I wanted to fix that. I spent the last two months building a local AI agent. The result is OxcerAI, a local-first document intelligence tool where the terminal never shows up. The shell exists internally as a tool the agent can call. The user just types. Three architectural decisions I'm proud of: 1. 𝗧𝗵𝗲 𝘁𝗲𝗿𝗺𝗶𝗻𝗮𝗹 𝘀𝘁𝗮𝘆𝘀 𝗵𝗶𝗱𝗱𝗲𝗻. The user never sees a terminal. llama.cpp runs inside the app. The shell exists internally as a tool the agent can call, not a user-facing surface. The user just types. 2. 𝗡𝗼𝘁𝗵𝗶𝗻𝗴 𝗹𝗲𝗮𝘃𝗲𝘀 𝘆𝗼𝘂𝗿 𝗺𝗮𝗰𝗵𝗶𝗻𝗲. No HTTP calls during inference. No external API by default. The model runs locally via llama.cpp + Metal on Apple Silicon. Cloud providers are available as an opt-in, explicitly configured by the user. Off by default. 3. 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 𝗮𝗰𝗰𝘂𝗺𝘂𝗹𝗮𝘁𝗲𝘀 𝗹𝗼𝗰𝗮𝗹𝗹𝘆. Most tools are stateless. Every session starts from zero. I designed OxcerAI so that workflow patterns, document relationships, and command vocabulary accumulate on the user's machine over time. On month six, the cost of switching isn't inconvenience. It's the loss of a context graph that exists nowhere else. The underlying code for all three memory layers, memory.rs, db.rs, agent_session_log.rs, is fully implemented and tested in v0.1. Wiring them to the main agent loop is the primary work of v1.0. Feedback welcome. Looking for collaborators. If you're interested in local agent architecture or want to contribute, check out the GitHub. Built with Meta Llama, Anthropic Claude, and Cursor

4

2 Comments

The 🚀 open-source NLP project: Fine-Tuning MuRIL for Nepali & Code-Mixed Text Classification I built a complete MuRIL (google/muril-base-cased) fine-tuning pipeline using Hugging Face Transformers, optimized for Nepali and Nepali–English code-mixed social media text—a common challenge in low-resource NLP. 🔗 GitHub: 👉 https://lnkd.in/gDRSr8fm Highlights: Sentence-level text classification Clean training & evaluation pipeline Weighted metrics for class imbalance Production-ready inference setup This project is suitable for sentiment analysis, social media mining, tourism reviews, and academic research, and can be extended toward API deployment. Always excited to work on low-resource and multilingual NLP 🚀 Feedback and collaboration are welcome. #NLP #MachineLearning #Transformers #MuRIL #HuggingFace #OpenSource #NepaliNLP #AI #DataScience

6

Based on the provided articles, here are 3 promising AI startups with recent funding news: 1. Flint - Raised $5 million in seed funding (October 14, 2025) - Led by Accel with participation from Sheryl Sandberg's fund and Neo - Building an AI platform that autonomously builds and updates websites - Source: https://lnkd.in/dYMYVyWu 2. Leonardo.ai - Acquired by Canva (announced July 29, 2024) - Had previously raised over $38.8 million from investors including Smash Capital, Blackbird, Side Stage Ventures and others - Specializes in generative AI content creation with over 19 million registered users - Source: https://lnkd.in/d3J5DAMA 3. Coco Robotics - While specific funding isn't mentioned, the company announced launching a new physical AI research lab (October 14, 2025) - Partnering with OpenAI and led by UCLA professor Bolei Zhou - Focuses on autonomous last-mile delivery robots - Source: https://lnkd.in/dUbYZDCf

3

37 years of work. Replaced in 2 weeks. For $20K. A researcher and 16 AI agents just built a C compiler from scratch. 100K lines. Passes 99% of GCC’s torture tests. Compiles FFmpeg, Redis, PostgreSQL, QEMU. Runs Doom. GCC took thousands of engineers and 37 years to get here. They say they (mostly) walked away. That mostly matters. No human wrote a single line. But the researcher kept redesigning tests, fixing CI pipelines when agents broke each other, and unsticking all 16 when they hit the same wall. The human didn’t write code. They engineered the system that let AI write code. That’s not replacement. That’s a role shift. The developers who thrive next won’t be the best coders. They’ll be the best system thinkers.🤞

153

5 Comments

Most existing #lean4 datasets contain only correct proofs. Models learn error correction with RL, that's expensive. With UW Math AI lab we release a dataset of 260k erroneous Lean proofs with - compiler feedback - reasoning trace - corrected proof Improvements in Error Correction: - Goedel 8B: 2x - Kimina 8B: 3x Paper: https://lnkd.in/gaybt4bd

117

2 Comments

In the past, I used pip tools like repo2txt (https://lnkd.in/esqVJ8nU) to prepare repositories for LLM prompting. Now tools like Uithub can save a lot of time as well. However, if you're privacy-conscious and working with private repositories, it's better to avoid third-party processing and instead use an open-source tool locally.

6

Happy to share that our paper "Octax: Accelerated CHIP-8 Arcade Environments for Reinforcement Learning in JAX" has been accepted to ICLR 2026 🇧🇷! Octax is a high-performance JAX-based framework that transforms CHIP-8 arcade games into fully vectorized RL environments. CHIP-8, a cousin to Atari 2600, offers a simpler yet rich set of arcade games that are nice for RL benchmarking. We have implemented a complete CHIP-8 emulator entirely in JAX that can load ROM files directly, creating the first end-to-end GPU-accelerated alternative to traditional CPU-bound arcade environments like Atari. Why it matters: - We achieve a 14x speed improvement over Envpool-accelerated Atari environments (the current standard), with orders-of-magnitude gains over traditional CPU emulators. This democratizes RL research on arcade games: researchers with low-end GPUs or small clusters can now run experiments in reasonable timeframes and conduct statistically rigorous evaluations with many more seeds than the traditional 5-seed standard, improving research reliability. - Unlike existing solutions that remain CPU-bound, Octax runs entirely on GPU, eliminating the computational bottleneck that has limited large-scale RL experimentation with image-based environments. - Extensible and modular design: The framework makes it easy to add new games or modify existing ones, providing flexibility for custom research needs. - As a proof-of-concept, we demonstrate that LLMs can generate entirely new, playable CHIP-8 games directly from prompts. This capability paves the way for exciting research directions in open-ended learning, procedural content generation, and automatic curriculum design. We are currently polishing the camera-ready version and developing comprehensive documentation to make Octax accessible to the broader RL community. Our goal is to provide the JAX ecosystem with a powerful, scalable tool for large-scale RL research. Huge thanks to my brilliant co-authors Thomas Michel and Hector Piteau for making this collaboration such a rewarding experience! 📄 Paper: https://lnkd.in/eYDz8M7Y 💻 GitHub: https://lnkd.in/eXKUpfd8 #Jax #ICLR2026 #OpenSource

149

6 Comments

I kept reusing the same PyTorch training pipeline across many of my projects (for tasks like tracking metrics, managing checkpoints, and handling training loops). I realized this setup could save others a lot of setup time, so I turned it into an open-source package. simpletrainer is a lightweight toolkit for clean training loops. Works for binary, multiclass, and multilabel classification, as well as regression tasks. Install: pip install torch-simpletrainer 🔗 GitHub: https://lnkd.in/dzd9ncfQ If you try it out, I’d be happy to hear what you think. #PyTorch #MachineLearning #OpenSource #DeepLearning

97

1 Comment

#MAVEN : #AI_Powered #Chip_Design: Spec to Silicon by : SIVAKUMAR P R Founder and CEO / Maven Silicon Source : https://lnkd.in/gKKYSvu9 [ 23 Slides : https://lnkd.in/gjJNzs8B) : 2025 #AutoDV - #Click, #Verify and #Relax by : Allen Chen, Eric Cheng / TESDA 

8

Attention-Based Stock Prediction AI for NEPSE🚀 I’ve developed an AI model that predicts NEPSE stock prices for the next 3 days using the past 15 days of data.📉And it doesn’t just rely on historical prices , it also considers technical indicators, calendar patterns, and other market factors that influence price movements. I’ve written a full blog about this on Medium: https://lnkd.in/dAU8TKZP 💥I tested the model for 3 consecutive days (yesterday, today, and tomorrow), and the prediction was off by just -0.99! 🤯I’m excited to see how the cumulative error performs for today and tomorrow. You can try it yourself and see the results firsthand☺️ And I’m open-sourcing the project on GitHub 💻so you can try it, explore the code, and even contribute new features. github link: https://lnkd.in/dXSG432y #NEPSE #StockPrediction #AIModel #DeepLearning #AttentionMechanism #AI #DeepLearning #MachineLearning #HuggingFace #LanguageModel #PersonalProject #GenerativeAI #OpenSource Here’s how my model’s predictions compare with the actual NEPSE prices on the test dataset:

51

6 Comments