Research Papers 论文研究 3d ago Updated 3d ago 更新于 3天前 51

Intelligence is Free, Now What? Data Systems for, of, and by Agents 智能免费,接下来怎么办?面向、由及为智能体构建的数据系统

AI inference costs have plummeted by a median factor of 50x annually, reaching levels where GPT-4-class capabilities cost under $1 per million tokens, signaling an era of virtually free intelligence. This cost reduction necessitates a paradigm shift in data systems, categorized into three new challenges: designing systems *for* agents, managing systems *of* agents, and enabling systems synthesized *by* agents. Data systems must evolve to handle "agentic speculation," characterized by high-volume AI推理成本急剧下降,GPT-4级能力成本从2023年初的每百万token约30美元降至不到1美元,部分供应商甚至低于0.10美元。 智能已足以胜任绝大多数知识工作,且成本每月都在降低,标志着“近乎免费智能”时代的到来。 数据系统需重新设计以支持“代理推测”(Agentic Speculation),即处理高并发、异构且大量重复的子查询流。 提出数据系统的三大新挑战:为代理设计系统(For)、由代理管理系统的状态与协调(Of)、以及由代理合成定制数据系统(By)。 建议数据系统从被动执行者转变为主动参与者,通过多查询优化、近似查询处理和提供性能反馈来辅助代理工作。

75
Hot 热度
70
Quality 质量
72
Impact 影响力

Analysis 深度分析

TL;DR

  • AI inference costs have plummeted by a median factor of 50x annually, reaching levels where GPT-4-class capabilities cost under $1 per million tokens, signaling an era of virtually free intelligence.
  • This cost reduction necessitates a paradigm shift in data systems, categorized into three new challenges: designing systems for agents, managing systems of agents, and enabling systems synthesized by agents.
  • Data systems must evolve to handle "agentic speculation," characterized by high-volume, redundant, and heterogeneous query streams, requiring optimizations like multi-query reuse and approximate query processing.
  • Future architectures should move beyond passive query execution to proactive assistance, offering latency estimates, steering agent behavior, and providing pre-computed views to enhance efficiency.

Why It Matters

The drastic drop in AI costs transforms intelligence from a scarce resource into a commodity, fundamentally altering the economic and technical landscape for data infrastructure. For practitioners, this means that traditional data system designs optimized for human-led, low-frequency queries are obsolete; systems must now support massive concurrency and autonomous decision-making loops inherent to AI agents. Researchers and engineers must prioritize new abstractions that allow data layers to actively collaborate with, rather than just serve, intelligent agents.

Technical Details

  • Agentic Speculation: Agents exhibit unique query patterns involving schema introspection, columnar exploration, and combinatorial hypothesis testing, often generating thousands of sub-queries per user request.
  • Redundancy Optimization: Experiments show that 80-90% of sub-queries from multiple agents are redundant; systems can leverage multi-query optimization and shared scans to reuse results and reduce computational waste.
  • Approximate Query Processing (AQP): To accelerate progress, data systems can return approximate answers or stream intermediate results, allowing agents to make decisions without waiting for exact, expensive computations.
  • Proactive System Design: Instead of waiting for explicit SQL, data systems can provide contextual guidance, such as latency estimates or pre-materialized views, effectively steering agents toward more efficient paths.
  • Higher-Level Primitives: Interfaces may evolve to support batch queries with specific approximation requirements or macro-like structures (e.g., DBT-style Jinja macros) to abstract away the complexity of enumerating search spaces.

Industry Insight

  • Architectural Overhaul Required: Organizations must begin redesigning their data stacks to support high-throughput, autonomous workloads. Legacy systems optimized for human interactivity will become bottlenecks for agentic workflows.
  • New Value in Optimization Layers: There is a significant opportunity for vendors to build middleware or extensions that specifically address agentic needs, such as caching layers for speculative queries or interfaces for approximate processing.
  • Shift in Skill Sets: Data engineers and scientists will need to understand agent behavior patterns and probabilistic system interactions, moving beyond deterministic SQL tuning to managing stochastic, high-volume agentic traffic.

TL;DR

  • AI推理成本急剧下降,GPT-4级能力成本从2023年初的每百万token约30美元降至不到1美元,部分供应商甚至低于0.10美元。
  • 智能已足以胜任绝大多数知识工作,且成本每月都在降低,标志着“近乎免费智能”时代的到来。
  • 数据系统需重新设计以支持“代理推测”(Agentic Speculation),即处理高并发、异构且大量重复的子查询流。
  • 提出数据系统的三大新挑战:为代理设计系统(For)、由代理管理系统的状态与协调(Of)、以及由代理合成定制数据系统(By)。
  • 建议数据系统从被动执行者转变为主动参与者,通过多查询优化、近似查询处理和提供性能反馈来辅助代理工作。

为什么值得看

这篇文章深刻揭示了AI成本断崖式下跌对底层数据基础设施的根本性冲击,指出传统的交互式BI工具模式已无法适应Agent主导的工作负载。它提出的“For, Of, By Agents”框架为数据库和数据工程领域的未来研究提供了清晰的战略方向,强调了从人类中心向代理中心范式转变的必要性。

技术解析

  • 成本下降趋势:推理价格每年下降9倍至900倍,中位数约为50倍。这种成本结构的变化使得大规模部署智能体成为经济上可行的选择,即使未达到“诺贝尔奖级别”的智能,当前水平也足以覆盖日常知识工作。
  • 代理推测(Agentic Speculation):代理在查询数据库时并非像人类那样单次提问,而是进行高体积、异构的工作流,包括模式检查、列探索和查询构建。实验显示,多个代理尝试同一任务时,仅10-20%的子计划是独特的,其余80-90%为重复工作,这为系统级优化提供了空间。
  • 系统优化策略
    • 多查询优化与共享扫描:利用传统数据库技术复用重叠子计划的结果。
    • 满意解(Satisficing)与近似查询处理(AQP):返回足够好的近似答案以加速代理决策,或通过流式传输中间结果让代理判断是否需要完整数据。
    • 高级接口原语:引入类似DBT Jinja宏的循环基元,允许代理批量提交带有不同近似要求的查询,而非逐一编写SQL。
  • 主动式数据系统:数据系统可利用其对数据和系统特性的先验知识,主动向代理提供延迟估计、相关查询结果或预计算的物化/虚拟视图,从而引导代理更高效地完成任务。

行业启示

  • 重构数据架构标准:企业和数据团队应重新评估现有数据栈,确保其能够高效处理成千上万个并发代理产生的海量、碎片化且高度冗余的查询请求,而非仅优化人类用户的交互式查询。
  • 投资代理协调与管理基础设施:随着代理承担更多知识工作,需要专门的数据系统来管理长周期任务的状态、协调多代理共识及故障恢复,这将是下一代数据平台的核心竞争力。
  • 探索自动化数据系统合成:鉴于代理有能力从零合成定制数据系统,行业需关注如何验证这些自动生成系统的正确性与安全性,建立信任机制以支持“由代理构建数据系统”的新范式。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Inference 推理 Agent Agent Research 科学研究