Top story
Meituan's LongCat-2.0 (1.6T MoE, 48B active) trained on tens of thousands of Huawei Ascend chips. a major signal that Chinese AI infrastructure is maturing beyond NVIDIA. The model shipped stealthily on OpenRouter as "owl-alpha" before its formal reveal. LongCat-2.0 · Reddit thread
Research
TUA-Bench (Meta). A new benchmark for general-purpose terminal-use agents. HuggingFace
OSWorld2.0. Extends computer-use agent evaluation to long-horizon, real-world desktop tasks. HuggingFace
SWE-INTERACT. Reimagines SWE benchmarks as user-driven, long-horizon coding sessions. arXiv
Agent Memory-Poisoning Forensics. Introduces trajectory signatures to detect memory-poisoning attacks on LLM agents. arXiv
One-Step Gradient Delay. Empirical/theoretical analysis showing async pipeline-parallel pretraining isn't blocked by a one-step gradient delay. arXiv
Tools
Google's Agentic Peer-Reviewer. Formal paper on an AI peer-reviewer that handled ~10K submissions at ICML/STOC and caught 34% more math errors than zero-shot prompting. Reddit
DeepSeek V4 in llama.cpp. PR merged, enabling local GGUF builds of DeepSeek V4. Reddit
Base44 / Base1. First app-creation platform to launch a proprietary LLM (fine-tuned on tens of millions of Wix user interactions), as coding-AI startups chase defensibility. TechCrunch · @Avishai_ab
Cursor Mobile. A mobile app for monitoring and steering your coding agent on the go. TechCrunch
Ornith-1.0. Open-source, self-improving Qwen fine-tune for agentic coding. GitHub
Industry
LMArena hits $100M. The widely-used AI leaderboard becomes a $100M-scale business. TechCrunch
Cerebras/OpenAI Deal Squeezes Everyone Else. The ~$20B Cerebras-OpenAI arrangement has effectively killed near-term high-throughput inference capacity for other customers. Reddit
Meta Was Secretly Running on Gemini. Report alleges Meta used Google's Gemini internally for CS, ads, and moderation before being cut off for excessive usage. Reddit
NASA Tests Local LLM Inference. NASA is evaluating onboard local LLM inference for future space missions. Reddit
OpenAI Teases Codex Hardware. OpenAI hints at a hardware device tied to its Codex coding product. The Verge
Community
Zvi vs. WSJ on "China matched Anthropic". Zvi Mowshowitz argues the WSJ claim is obvious nonsense and overstated. Substack
Apple Neural Engine Reverse-Engineering. Deep write-up on ANE architecture, ISA, and performance, including a ModernBERT running example. Hacker News
Krea-2-Turbo. New image model with editing capabilities and easy uncensoring. Reddit
Omen AI's Water-Cooled Data Centers. Startup pitches liquid-based cooling to optimize AI data center efficiency. TechCrunch