Policy Inflection, Evaluation Revolution, and Memory Breakthrough: AI Enters the Era of Governance and Refinement

New York became the first U.S. state to pause new data center approvals, directly targeting the environmental and energy bottlenecks of AI compute expansion. OpenAI proposed a 'reverse federalism' AI governance framework, advocating state-led laws to build a national safety system. Microsoft leveraged AI to fix a record 570 security vulnerabilities, showcasing AI's practical value in security governance, while OpenAI's AI-era scorecard provided a new yardstick for measuring AI return on investment, accelerating the deployment of governance and evaluation tools.

• New York State Halts Approval of All New Data Centers, First State to Do So
• OpenAI Proposes 'Reverse Federalism' for AI Governance: State Laws First, National Framework Next
• Microsoft Patches Record 570 Security Vulnerabilities, Citing AI
• OpenAI Introduces AI Age Scorecard: A New Way to Measure ROI

Human naturalness in speech AI was for the first time quantified by Real World VoiceEQ, scoring 87 points and topping the week's charts, marking a measurable phase in evaluation. Meanwhile, two 79-point studies dissected the internal bias mechanisms of LLMs as judges and unified the evaluation framework for visual tool-calling agents. Issues such as test oracle problems in synthetic corpora and risk audit statements for RL models further pushed evaluation from model performance toward trustworthiness and reliability.

• Real World VoiceEQ: Measuring the Human Quality of Voice AI
• Inside the Unfair Judge: A Mechanistic Interpretability Account of LLM-as-Judge Bias
• MM-ToolSandBox: A Unified Framework for Evaluating Visual Tool-Calling Agents
• The Test Oracle Problem in Synthetic LLM-as-Judge Corpora: Disappearance, Distortion and a Validation Protocol
• Auditing the Risk Claims of Distributional Reinforcement Learning: Majority of Assertions Refuted

Online neural-spatial memory synthesized real-time memory from dynamic perspectives, opening new paths for temporal understanding; JoLT achieved near-lossless KV cache compression via joint Tucker decomposition and JL residual allocation, significantly reducing inference costs. SCOPE-RL optimized reasoning paths before and after success, while FlashDecoder stream-decoded latent codes to pixels in real time—together pointing toward more efficient and dynamic memory and reasoning architectures.

• Online Neural Space Time Memory: Real-Time Memory Mechanism for Dynamic Novel View Synthesis
• JoLT: Near-Lossless KV Cache Compression via Joint Tucker and JL-Residual Allocation
• SCOPE-RL: Optimizing Reasoning Paths Before and After Success
• FlashDecoder: Real-Time Latent-to-Pixel Streaming Video Decoder
• Leveraging Instruction Tuning and Merging for Reasoning Model Adaptation

The Lobsters community successfully migrated to SQLite, validating single-server architecture and challenging the mainstream narrative of large-scale distribution. Tools converting Mermaid to ASCII art and Unicode box drawing improved in-browser chart rendering, and Firefox compiled to WebAssembly made it possible to run a browser inside a browser. Though diverse in focus, these innovations collectively reflect a pursuit of resource efficiency and toolchain simplification.

• Lobsters Community Successfully Migrates to SQLite: A Win for Single-Server Architecture
• Mermaid to ASCII Art: In-Browser Diagram Rendering with Go+WASM
• Mermaid to Unicode Box Art: Introducing grok-mermaid
• Firefox Compiled to WebAssembly, Runs Inside a Browser
• SQLite Query Explainer: An Interactive Tool Running in the Browser