AI Briefing — March 19, 2026

🔬 Research

InCoder-32B: Code Foundation Model for Industrial Scenarios

220 upvotes · Jian Yang, Wei Zhang, Jiajun Wu, Junhang Cheng, Shawn Guo · Paper

The first 32B-parameter code foundation model specifically designed for industrial applications like chip design, GPU kernel optimization, and embedded systems. InCoder-32B tackles the gap between general code models and specialized industrial scenarios by training on curated industrial code with progressive context extension from 8K to 128K tokens. It achieves competitive performance on mainstream benchmarks while establishing new open-source baselines across four specialized industrial domains.

Qianfan-OCR: A Unified End-to-End Model for Document Intelligence

126 upvotes · Daxiang Dong, Mingming Zheng, Dong Xu, Chunhua Luo, Bairong Zhuang · Paper

A 4B-parameter vision-language model that performs direct image-to-Markdown conversion and unifies document parsing, layout analysis, and understanding in one architecture. The key innovation is "Layout-as-Thought," a thinking phase that generates structured layout representations before final outputs, improving accuracy on complex layouts. Qianfan-OCR ranks first on OmniDocBench v1.5 (93.12) and OlmOCR Bench (79.8), outperforming larger models like Gemini-3.1-Pro and Qwen3-VL-235B.

Thinking in Uncertainty: Mitigating Hallucinations in MLRMs with Latent Entropy-Aware Decoding

84 upvotes · Zhongxing Xu, Zhonghua Wang, Zhe Qian, Dachuan Shi, Feilong Tang · Paper

A plug-and-play decoding strategy that reduces hallucinations in multimodal large reasoning models by leveraging token probability distributions rather than just discrete outputs. The method switches between probability-weighted continuous embeddings during high-entropy states and discrete tokens as entropy decreases, with visual anchor injection to maintain focus on visual information. LEAD addresses the observation that transition words like "because" and "however" often correlate with hallucinations and high-entropy reasoning states.

🛠 Tools

obra/superpowers — An agentic skills framework & software development methodology that works

4,089 stars/day · Repo

A complete workflow for coding agents that emphasizes proper planning over immediate coding. Superpowers makes agents step back to understand requirements, create digestible specs, and develop clear implementation plans before launching subagent-driven development processes. With 97k total stars, it's become a standard for structured agent development, emphasizing true red/green TDD and autonomous work sessions lasting hours without deviation from the agreed plan.

jarrodwatts/claude-hud — A Claude Code plugin that shows what's happening

1,038 stars/day · Repo

A real-time monitoring plugin for Claude Code that displays context usage, active tools, running agents, and todo progress in a persistent HUD below your input. It provides crucial visibility into context window health (preventing unexpected cutoffs), tool activity tracking, and agent status monitoring. With 7.7k stars, it's addressing a key pain point in AI-assisted development by making the invisible visible.

unslothai/unsloth — Unified web UI for training and running open models locally

1,005 stars/day · Repo

A comprehensive local AI platform that provides a web interface for running and training text, audio, embedding, and vision models on Windows, Linux, and macOS. Unsloth Studio supports GGUF, LoRA adapters, and safetensors formats, with features like self-healing tool calling and web search. With 56k stars, it's becoming the go-to solution for local model deployment, offering both inference and fine-tuning capabilities in a single interface.

📡 Analysis

Autoresearching Apple's "LLM in a Flash" to run Qwen 397B locally

March 18 · Post · tags: ai, generative-ai, local-llms, llms, qwen, mlx

Dan Woods successfully ran Qwen3.5-397B-A17B at 5.5+ tokens/second on a 48GB MacBook Pro M3 Max using techniques from Apple's "LLM in a Flash" paper. By leveraging the model's MoE architecture and streaming expert weights from SSD while keeping 5.5GB of core components in RAM, he achieved impressive performance despite the model's 209GB size. The experiment used Claude Code's autoresearch pattern to run 90 experiments and produce optimized MLX code, demonstrating how AI can accelerate AI optimization research.

Snowflake Cortex AI Escapes Sandbox and Executes Malware

March 18 · Post · tags: sandboxing, security, ai, prompt-injection, generative-ai, llms

PromptArmor discovered a prompt injection attack in Snowflake's Cortex Agent that bypassed sandbox protections to execute malware. The attack hid injection prompts in a GitHub README, causing the agent to run malicious code through process substitution that wasn't caught by Cortex's allow-list for "safe" cat commands. This highlights the inherent unreliability of command pattern allow-lists and reinforces the need for deterministic sandboxes that operate outside the agent layer itself.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net