返回首页

对比阅读

对比阅读:95 Tokens for Cloud Plans, 97% Stay Local: Task Routing Outweighs the Model 与 95个Token做云规划、97%留本地:任务路由比模型本身更重要

AEN
GoogleAntigravityGemma·

95 Tokens for Cloud Plans, 97% Stay Local: Task Routing Outweighs the Model

What This Is

The real story behind this week's Google Antigravity SDK update isn't "models can run offline" — it's the audit Agent division of labor it demonstrates: cloud-side Gemini only receives filenames and task descriptions, and uses 95 Tokens to break out an audit plan; three local Gemma instances then handle exploitation, patch writing, peer review, and regression testing, with 97.2% of Token consumption happening locally.

What's actually worth learning is the control layer behind it — what Google calls "task routing": based on data sensitivity, task difficulty, and failure risk, it decides whether each step runs locally or in the cloud, what the minimum context is, and whether human approval is needed. Google analogizes it to hospital triage: the front desk only collects the necessary symptoms, critical cases get escalated to specialists, and medical records aren't broadcast to every department for the sake of "convenience."

How the Industry Sees It

Supporters view it as a realistic path for enterprise Agent deployment: source code never leaves the machine, the cloud only does planning, and both compliance and cost can be balanced. Internal data processing in finance, healthcare, and manufacturing finally has a way to "use large models without handing over core data." Plenty of architects in the community are already redesigning internal toolchains around this paradigm.

The dissenting view deserves equal airtime. The key to task routing isn't the model — it's data labeling, permission design, audit logs, and the engineering work behind them. Many companies haven't even figured out "which fields count as sensitive," and no routing rule, however clever, will save them. A more measured critique: keeping 97% of Tokens local looks great in the demo, but local models have lower capability ceilings; complex tasks may simply fail to produce a viable plan, and forced routing just swaps the problem from "leakage" to "poor quality." Engineering pragmatists also remind us: if a local model can read the entire disk, access the network, and grab Git credentials, "local" doesn't equal "secure."

Impact on Regular People

  • For enterprise IT: when procuring Agent solutions, "controllable data flow" will gradually matter more than model benchmark scores — routing policies and audit logs will be asked about before model parameters.
  • For individual careers: developers should understand task routing early — not every task should hit a cloud API. Ask first: "does this step touch sensitive data?" Save cost and compliance headaches.
  • For consumer markets: AI on phones and PCs will likely adopt a "local-first, cloud-fallback" default first — ordinary users won't need to perceive what's running under the hood.
来源: juejin.cn
BZH
GoogleAntigravityGemma·

95个Token做云规划、97%留本地:任务路由比模型本身更重要

这是什么

Google Antigravity SDK本周更新的看点不是「断网能跑模型」,而是它示范的一个审计Agent分工:云端Gemini只拿到文件名和任务描述,用95个Token拆出审计计划;三个本地Gemma实例完成漏洞复现、写补丁、互评和回归测试,97.2%的Token消耗在本地。

真正值得学的,是背后那个叫「任务路由」的控制层——根据数据敏感性、任务难度、失败风险,决定每一步走本地还是云端、最小上下文是什么、是否需要人工审批。Google把它类比成医院分诊:前台只收集必要症状,危急情况升级专家,病历不会为了「方便」广播给所有科室。

行业怎么看

支持方视它为企业Agent落地的现实路径:源码不出机器、云端只做规划,合规和成本都能兼顾;金融、医疗、制造业的内部数据处理,终于有了「既用上大模型、又不交出核心数据」的方案。社区里不少架构师已经在按这个范式重新设计内部工具链。

反方意见同样值得听。任务路由的关键不是模型,而是数据标签、权限设计、审计日志这些工程活儿;很多公司连「哪些字段算敏感」都没划清楚,再精巧的路由规则也会失效。更冷静的批评是:演示里97%Token留本地看似漂亮,但本地模型能力上限更低,复杂任务可能根本拆不出可用计划,强行路由只是把问题从「泄密」换成「质量差」。还有工程派提醒:本地模型若能读整个硬盘、访问网络、拿到Git凭据,「本地」并不等于「安全」。

对普通人的影响

  • 对企业IT:采购Agent方案时,「数据流向可控」会逐渐比模型跑分更重要——路由策略、审计日志会比模型参数更先被问。
  • 对个人职场:开发者值得早一步理解任务路由——不是所有活儿都该调云端API,先问「这一步碰不碰敏感数据」,省成本也省合规麻烦。
  • 对消费市场:手机和PC端AI很可能率先采用「本地优先、云端兜底」的默认设置,普通用户不必感知背后机制。
来源: juejin.cn