Technology · Monthly · 星际流动

The through line

In July 2026, the AI industry shifted from pure model capability competition toward deeper battles in engineering, evaluation, and governance. New York State's moratorium on data center approvals marked a symbolic tightening at the policy level, while security incidents at Hugging Face and Modal exposed infrastructure fragility. Meanwhile, AI coding agents moved to the core of software development, and multimodal and reasoning research advanced rapidly. The month's signals suggest a search for a new balance between scaling and control.

Expansion and Constraint of Infrastructure

This month's infrastructure discourse was marked by a tension between expansion and risk. New York State's moratorium on data center approvals applied the first policy-level brake to AI compute expansion, while the idle-GPU problem, likened to 'grounded airplanes', highlighted the urgency of utilization efficiency. Supply-chain security also tightened, with PyPI imposing a 14-day upload deadline and the Hugging Face agent-intrusion disclosure raising security concerns. In response, the industry advanced zero-egress cloud storage and privacy-preserving robustness validation, seeking a balance between efficiency and safety.

AI Coding Agents Reshaping Software Development

The Anthropic Claude Code team's fireside chat was one of the strongest signals this month: AI coding agents are evolving from auxiliary tools into core transformers of software development. Multiple rapid releases of Claude Code, along with GitHub's Agentic Workflows and Copilot Canvases, pushed AI from code completion to cross-repository documentation generation and interactive workspaces. NTT DATA's use of Codex to cut incident analysis to 30 minutes demonstrated practical value in operations.

Evaluation and Security Audit: From Benchmarks to Social Simulators

Evaluation research surged this month, reflecting a systemic rethinking of model capability and behavioral trustworthiness. Two settings on ARC-AGI-3 tripled benchmark scores, hinting at the benchmark's own fragility; a mechanistic interpretability analysis of LLM judge bias exposed the evaluator's blind spots. An audit method for social simulators and a multimodal tool-calling evaluation framework extended the evaluation target from single models to complex interactive systems. The AI-era scorecard also sought to provide a new quantitative yardstick for investment returns.

Reasoning and Architecture Innovation: Pushing Efficiency and Interpretability

At the level of model architecture and reasoning, multiple studies this month pointed to dual advances in efficiency and interpretability. CARE innovated mixture-of-experts LoRA via confidence-aware routing, while κ-LoRA used condition numbers to identify which LoRA matrices deserve updates, both pointing to finer-grained parameter utilization. JoLT achieved near-lossless KV cache compression, and NIFA proposed a nonlinear IMC-enhanced FPGA architecture, offering new hardware-software co-design possibilities for inference deployment. OlmoEarth extended geospatial reasoning to planetary scale, demonstrating domain-specific capability expansion of large foundation models.

Multimodal Perception and Generation: From Voice Naturalness to Surgical Robots

The multimodal field advanced across the full spectrum from perceptual quality to embodied applications. Real World VoiceEQ put quantification of human naturalness in speech AI on the agenda, and visual prompt engineering improved video models with a simple method. NVIDIA Cosmos-H-Dreams brought real-time generative simulation to surgical robots, grounding multimodal generation in medical scenarios. A cognitive structured multimodal agent and an EHR foundation model pushed new paradigms in long-horizon understanding and clinical data modeling.

352 items · 189 selected · 9 sources

Evaluation and Governance: Expansion and Calibration of AI Infrastructure