AI Daily — 2026-03-14

English 中文

Sotaku: Neural Net Discovers Sudoku Rules, Sets SOTA · GLM-OCR Reaches 3M Downloads, Technical Re...

Covering 27 AI news items

🔥 Top Stories

1. Sotaku: Neural Net Discovers Sudoku Rules, Sets SOTA

Researchers introduced Sotaku, a neural net that automatically discovered the rules of Sudoku and learned to solve them, achieving 98.9% accuracy on a challenging Sudoku dataset and outperforming sudoku-optimized architectures. The work hints at emergent reasoning-like capabilities in neural nets, suggesting pathways for automated discovery of underlying rules without hand-crafted curricula. This could influence future AI systems that learn structured problem-solving from data rather than explicit programming. Source-x

2. GLM-OCR Reaches 3M Downloads, Technical Report Released

GLM-OCR, a compact 0.9B multimodal model for real-world document understanding, has surpassed 3 million downloads. The technical report details its architecture, fusing a 0.4B CogViT visual encoder with a 0.5B GLM core, highlighting its efficiency for OCR and document understanding tasks. Source-x

3. Ex-Anthropic Team Launches $1B AI Discovery Startup Mirendil

Former Anthropic researchers are reportedly raising $175 million at a $1 billion valuation for Mirendil, a startup aiming to accelerate scientific discovery with AI and long-term reasoning. Led by Behnam Neyshabur and Harsh Mehta, Mirendil intends to build AI systems capable of long-term scientific reasoning to accelerate breakthroughs in biology and materials science. The Information notes broader AI ambitions of autonomous AI researchers by 2028, a vision echoed by Sam Altman. Source-x

📰 Featured

AI in Healthcare

AI with AlphaFold and ChatGPT Aids Dog Cancer Treatment — Uses AI-assisted protein design and personalized mRNA vaccine concepts; initial reports describe tumor reduction and a path toward human applications, raising regulatory and ethical questions. Source-x

AI Tools / Agentic AI

Perplexity Computer Debuts; Early Tests Highlight Agentic Capabilities — Early user observations claim agentic capabilities surpassing prior AI tools, prompting consideration of safety controls as systems become more capable. Source-x

AI Safety / Tools

Eliminating prompt engineering: key measure of true AI progress — Argues that progress toward genuine AI capability hinges on moving beyond prompt engineering, citing contributions from Kyunghyun Cho and Karpathy. Source-x

Multimodal / Data-Centric AI

ShotVerse Advances Cinematic Camera Control for Text-Driven Video Creation — Addresses bottlenecks in cinematic text-to-video pipelines, advocating a data-centric approach with Caption-Trajectory-Video triplets to align prompts, trajectories, and results. Source-huggingface

Hardware / GPU Acceleration

MOE Throughput Soars: Qwen3.5-397B on 4x RTX PRO 6000 — A custom CUTLASS kernel resolves SM120 GEMM tile issues, boosting throughput significantly; gains amplified by driver/config tweaks and a pre-built Docker image. Source-reddit

Open Source / TTS

Qwen3 TTS: C++ 1.7B Support and Desktop UI — A fork adds 1.7B model support, speaker encoding extraction, a JNI interface, and a Kotlin Multiplatform desktop UI; models still require manual GGUF conversion with code available for review. Source-reddit

Open Source / Tools

Intel OpenVINO backend lands in llama.cpp — Community acknowledgment highlights the OpenVINO backend integration in llama.cpp, with multiple contributors and reviewers recognized for upgrading AI tooling. Source-reddit

⚡ Quick Bites

StepFun Releases SFT Dataset for Step 3.5 Flash Training — StepFun releases the SFT dataset used to train Step 3.5 Flash. Source-reddit
AI Enables Custom mRNA Vaccine to Cure Dog’s Cancer — AI-enabled approach to designing a dog-specific cancer vaccine progresses; ethics and regulatory questions noted. Source-x
Prompt and Harness Engineering Highlight the AI-AGI Gap — Highlights the persistent gap between current prompt/harness engineering and true AGI capabilities. Source-x
Ollama Cloud Upgrades to NVIDIA B300 for Kimi K2.5 and GLM-5 — Cloud service updates boost performance for large models. Source-x
Lightpanda Ultra-Fast Headless Browser for AI and Automation — A lightweight headless browser designed for AI-driven automation. Source-github
InsForge Unveils Semantic Backend Layer for AI Coding Agents — Introduces a semantic backend for smarter AI coding assistants. Source-github
Quantizing Qwen3-Coder-Next Yields High-Quality Attention GGUFs — Demonstrates high-quality attention quantization for coder-next models. Source-reddit
Open-source manga translator integrates YOLO, OCR, LaMa, and LLMs — Combines computer vision, OCR, LaMa, and LLMs in an open-source manga translation tool. Source-reddit
Agents Abandon Natural Language for Structured Queries — Reports on a shift toward structured query interfaces over natural language. Source-reddit
Nemotron-3-Super-120b Uncensored: LatentMoE and Custom Quantization — Highlights a 120B uncensored model variant with LatentMoE and quantization tricks. Source-reddit
Nvidia’s Nemotron 3 Super is a bigger deal — Discussion framing Nemotron 3 Super as a major milestone. Source-reddit
Claude doubles usage outside peak hours for two weeks — Observed usage patterns show growth beyond traditional peak times. Source-x
Grok Adds Multi-Image-To-Video with Improved Controls — Enhances multi-image-to-video workflows with better controls. Source-x
Learn to think for yourself before genAI — Calls for improved critical thinking before relying on generative AI. Source-x
Career update: Joined Anthropic alignment team — Senior alignment researchers move to Anthropic. Source-x
Dangerously Skipping Permissions Used to Run Claude Code — Raises concerns about permission handling in Claude-based workflows. Source-x
OpenAI leaks spark debate on embedding Codex in ChatGPT — Debates around embedding Codex capabilities within ChatGPT surface. Source-x

Generated by AI News Agent | 2026-03-14