2026-W42 · 2026.10.12—10.18

从候选包到可执行变更

From candidate pack to executable change

这一周,演示、并行代理、实时语音动作、图像编辑与 checkout 配置都更容易越过“建议”进入协作与外部系统。真正可扩大的不是某个生成结果,而是能说明它基于哪个版本、谁批准、何时停止和如何撤回的交接过程。

This week, slides, parallel agents, real-time voice actions, image edits and checkout configuration became easier to move from suggestion into collaboration and external systems. What scales is not a single generated result, but a handoff that can state its version, approver, stop point and rollback path.

一、协作画布扩大了“可编辑”,没有扩大“已批准”

1. A collaborative canvas expands what is editable, not what is approved

AI Creation

OpenAI 10 月 16 日在 ChatGPT release notes 中称,Slides 可从对话、文件、提纲或示例演示开始,支持实时共同编辑,并能在 Web 导出 PowerPoint、PDF 或 Google Slides;但它只向有 Space access 的合资格 Pro、Business、Enterprise 用户逐步推出。协作编辑使一页的排版更快变动,却不会自动确认图表数值、照片权利、品牌版本、替代文本或导出时的字体替换。

On October 16, OpenAI said in its ChatGPT release notes that Slides can start from a chat, file, outline or example deck, supports real-time collaboration, and can export on the web to PowerPoint, PDF or Google Slides; it is rolling out to eligible Pro, Business and Enterprise users with Space access. Collaborative editing makes a layout move faster, but does not automatically validate chart numbers, photo rights, brand versions, alt text or font substitution on export.

ChatGPT release notes ↗

二、并行不是共识:先合并证据,再合并文案

2. Parallel is not consensus: merge evidence before copy

AI Marketing

同一条 OpenAI 更新称,ChatGPT chat tab 的 GPT‑6 Pro 可让多个代理并行处理任务不同部分,用户可跟随进度并打开完成任务。这个结构适合把研究、素材盘点、文案草案和反例检查拆开,但并不表示各子任务使用同一份简报、权限或事实版本。应要求每个任务交回来源、输入版本、未解决项和负责人;一名编辑负责处理冲突,才形成可发布的候选包。

The same OpenAI update says GPT-6 Pro in ChatGPT’s chat tab can let multiple agents work on different parts of a task in parallel, with users following progress and opening completed tasks. This suits splitting research, asset inventory, copy drafting and counterexample checks, but does not mean each sub-task used the same brief, permissions or fact version. Require every task to return sources, input version, open issues and an owner; one editor resolves conflicts before there is a publishable candidate pack.

GPT‑6 Pro update ↗

三、实时语音动作需要更窄的写权限

3. Real-time voice action needs narrower write permission

AI Commerce

OpenAI API changelog 在 10 月 15 日加入 GPT‑Live 对 Decisions 的原生集成:应用能在用户说话中行动,或随着输入/输出 transcript delta 触发动作。次日,prompt-to-voice 向全部 API 客户开放,并可在 GPT‑Live session 中复用。把语音接到动作的价值在于减少等待;风险是转录可能未完成、身份未核验、用户意图被误解。订单、退款、改地址、扣款不能因一段语音或一次工具调用就自动完成。

On October 15, OpenAI’s API changelog added native Decisions integration for GPT-Live: an application can act while a user speaks or trigger actions from input/output transcript deltas. The next day, prompt-to-voice opened to all API customers for reuse in GPT-Live sessions. Connecting voice to action reduces waiting, but transcripts may be incomplete, identity may be unverified and intent may be misunderstood. Orders, refunds, address changes and charges must not complete merely because of a voice turn or tool call.

OpenAI API changelog ↗

四、生成控制需要同成本一起验收

4. Generation controls must be accepted alongside cost

AI Creation

Runway Dev 在 10 月 12 日加入 Ideogram 4.5:最多五张参考图、黑白 mask 局部编辑、1–8 张输出、1K/2K 或 auto 比例,以及按质量和输出数量相乘的 credits。更多控制并不自动带来可读文字、正确商品事实、干净边缘或商用许可;自动比例和批量出图也会放大审批面与成本面。先在一张授权图和一个固定 mask 上比较不同输出数、质量档与真实 credits,再选择可交付组合。

On October 12, Runway Dev added Ideogram 4.5 with up to five reference images, black-and-white-mask local edits, one to eight outputs, 1K/2K or auto ratios, and credits multiplied by quality and output count. More control does not automatically yield legible text, correct product facts, clean edges or commercial clearance; automatic ratios and batches also expand approval and cost surfaces. Compare output count, quality and actual credits on one authorized image and fixed mask before choosing a deliverable combination.

Runway Dev changelog ↗

五、把配置帮助留在测试环境,把交易事实留在系统记录

5. Keep configuration help in test; keep transaction facts in systems of record

AI Commerce

Shopify 10 月 13 日将 Sidekick 放进 desktop checkout/customer accounts editor,并允许它解释设置、建议 app、搜索 Help Center,以及在 Checkout Blocks 中帮助规划 payment、delivery 和 validation function。ElevenLabs 10 月 16 日的 CRM/工单指南则描述语音代理通过原生接入或 webhook 读取记录、写入工单与发送通话后摘要。两者都应从低风险测试开始:配置建议不是已执行变更,读取事实不是写回权限,连接成功也不是已解决客户问题。

On October 13, Shopify put Sidekick in the desktop checkout/customer-accounts editor to explain settings, recommend apps, search Help Center, and help plan payment, delivery and validation functions in Checkout Blocks. ElevenLabs’ October 16 CRM/help-desk guide describes voice agents reading records, writing tickets and sending post-call summaries via native connections or webhooks. Start both at low risk: a configuration suggestion is not an executed change, reading facts is not permission to write, and a successful connection is not a resolved customer problem.

Shopify changelog ↗ · ElevenLabs guide ↗

本周可复用交接:三包、一表、一次撤销演练

Reusable handoff this week: three packs, one matrix, one revocation rehearsal

AI World workbench

A 候选包只读资料和待审草稿;B 批准包含来源、事实版本、冲突、预算、双人签字与回退;C 执行包只有一个系统、一个时间窗和一个撤销人。再配一张权限矩阵:绿色查询、黄色草拟写入、红色交易/外发。先演练拒绝 B 时删除 C、撤销连接、保留日志并回放原因;做不到就不扩大。

Pack A—candidate holds read-only material and reviewable drafts; Pack B—approval holds sources, fact version, conflicts, budget, two-person sign-off and rollback; Pack C—execution has one system, one time window and one revoker. Add a permission matrix: green lookup, yellow draft write, red transaction/external send. First rehearse rejecting B, deleting C, revoking the connection, preserving logs and replaying why; do not scale if this fails.