Generic AI alone will not create durable advantage; companies must control the models, agent systems, context, and memory that determine how intelligence behaves, along with its cost, quality, and risk. The author argues firms should buy commodity infrastructure but own the agent layer and feedback loop where intelligence compounds with use.
Feedback Loop
反馈闭环
Iterative mechanism that lets AI systems improve with use, framed as the source of durable advantage.
Curated Intelligence
The article introduces 'Harness Engineering,' advocating for optimizing system environments—such as documentation, context loading, and validation loops—to enhance the reliability of coding agents like Codex and Claude Code, rather than relying solely on prompt engineering. Based on the HyperCode project and internal testing, the authors outline ten actionable patterns aimed at reducing manual intervention and achieving more stable automated delivery.
As model capabilities commoditize, the competitive moat for Agents shifts to engineering systems. A QCon discussion highlights the paradigm shift from static execution (Harness) to dynamic feedback loops (Loop), emphasizing that humans must retain decision-making authority while noting that true self-evolution remains distant due to context gaps.
QuEra leverages AI to automate laser frequency locking and calibration for its quantum computers, reducing reliance on human experts for severe lock failures. The system replaces manual on-site recovery with automated sequences to improve uptime and operational efficiency. While the auto-relock controller runs as a standalone production script, the full AI-in-the-loop tuning workflow is being packaged into an independent tool.
Alibaba's Qwen team released Qwen 3.8-Max, the strongest model in its series with 2.4 trillion total parameters (95B active). It delivers significant upgrades in coding, office tasks, and long-horizon complex workflows, marking the first time a Max-tier model will have open weights next week. Benchmarks demonstrate its ability to autonomously complete ten-day coding projects via self-evolving feedback loops.