1. Three Labs, One Bet: The AI Industry Rejects Its Scaling Ceiling Anthropic signed a multi-gigawatt compute deal. Within the same week, OpenAI declared a new phase for enterprise AI. Microsoft's head of AI published an essay arguing the technology won't plateau.

2. Developers Question Who Should Guard the Software Supply Chain The argument that consumed Hacker News wasn't about whether Claude Mythos is dangerous. It was about who gets to decide.

3. Researchers Say AI Makes Everyone Think Alike. Developers Say That's Their Edge Individuals using AI chatbots generate more ideas on their own. Groups using those same tools produce fewer and less creative ideas collectively.

In Brief

• OpenAI Publishes Child Safety Blueprint for AI Developers OpenAI released a framework for building age-appropriate safeguards into AI products, covering content filtering, age verification, and reporting mechanisms. The blueprint is aimed at third-party developers building on its APIs. OpenAI Blog

• Claw-Eval Targets Blind Spots in AI Agent Benchmarks A new evaluation suite exposes three gaps in how autonomous agents are tested: opaque grading that ignores intermediate steps, weak safety checks, and narrow interaction coverage. Claw-Eval includes 300 human-verified tasks across real software environments. Hugging Face Papers

• ThinkTwice Trains LLMs to Solve and Self-Correct in Two Phases Researchers propose a training framework that alternates between solving reasoning problems and refining answers, using the same binary reward signal for both phases. The method improved accuracy across five math benchmarks without requiring external critique or correctness annotations. Hugging Face Papers

• Retrieval Systems Trained on Agent Behavior Outperform Human-Click Models A study retrains information retrieval models using LLM agent interaction logs instead of human click data. Agent-optimized retrieval performed better when embedded in multi-turn reasoning loops, where traditional ranking signals break down. Hugging Face Papers

• ACES Proposes Ranking-Based Selection for LLM-Generated Code and Tests When both candidate code and candidate tests may be wrong, counting test passes is unreliable. ACES uses a leave-one-out AUC method to rank code by test agreement patterns, sidestepping the circular dependency between code correctness and test correctness. Hugging Face Papers

• Video-MME-v2 Benchmark Exposes Gap Between Leaderboard Scores and Real Video Understanding Existing video understanding benchmarks are saturating, inflating model scores beyond actual capability. Video-MME-v2 introduces a three-tier evaluation hierarchy designed to test robustness and faithfulness rather than pattern matching. Hugging Face Papers

• Game Bug Benchmark Tests Whether LLMs Can Do QA Work GBQA ships 30 playable games containing 124 human-verified bugs at three difficulty levels. Current LLMs detected bugs in controlled settings but struggled with complex runtime interactions that require extended gameplay. Hugging Face Papers

• Memory Intelligence Agent Cuts Storage Costs for Long-Running Research Agents Deep research agents that store full reasoning trajectories face ballooning retrieval and storage overhead. MIA compresses past experiences into reusable patterns instead of raw trajectories, reducing memory costs while preserving reasoning quality. Hugging Face Papers

• LIBERO-Para Reveals VLA Robots Break When Instructions Are Rephrased Vision-language-action models fine-tuned on limited robotic data overfit to specific instruction wording. LIBERO-Para systematically varies action expressions and object references to measure brittleness, finding significant performance drops from simple paraphrases. Hugging Face Papers

Read the full edition →