Google 领导层重组，四位元老离职创业

2026 年 8 月 6 日，Google 用一份人事公告回答了一个技术问题。Gemini 3.5 Pro 原计划 6 月发布，一再延期，旗舰空窗近半年；同一天，Alphabet 股价单日跌约 4%。市场用脚投票之后，Google 的选择不是发布新模型，而是重组领导层。

Demis Hassabis 卸任 DeepMind 日常管理，转任 Alphabet 首席科学家兼 DeepMind 主席；Koray Kavukcuoglu 升任 SVP，独立接管 Gemini 产品线。管理权从此分离：研究机构与旗舰产品不再由同一人同时掌舵。对一家以「研究驱动」为荣的公司，这是罕见的组织信号——旗舰的发布节奏，需要有人专门负责，而不是等研究团队「准备好了再说」。

更震动行业的是四位元老的离开。Jeff Dean、Sanjay Ghemawat、Oriol Vinyals、Quoc Le——这串名字几乎就是 Google 深度学习史的前半本目录。他们离职创办 Discovery Loop，谷歌作为创始投资方。这不是一次出走，而是一次「内部创业」：巨头出钱，元老出力，新实验室与老东家保持资本关系。过去两年，大厂 AI 元老离职创业已成常态，但一次走四位基础设施级研究者，仍然刷新了纪录。

把时间线摆开，这次重组有清晰的因果链：6 月计划发布未兑现 → 7 月 Flash 系列守住生产基本盘但旗舰缺席 → 8 月股价承压 → 领导层重组。Google 用组织调整回应延期，等于承认问题不在模型能力，而在节奏管理。对开发者，Gemini API 的路线图与负责人更替，旗舰模型的发布预期需要重新校准。

Discovery Loop 会做什么，四位元老与 Google 之间有没有摩擦，这些细节外界无从得知，也不该编造。可以确认的是结果：Google 的 AI 领导层完成代际交接，Gemini 由 Kavukcuoglu 的新班子主导，而一代基础设施研究者进入了创业周期。旗舰延期没有杀死 Google 的 AI 野心，但它证明了一件事——在 2026 年，模型发布节奏本身就是公司治理问题。

On August 6, 2026, Google answered a technical problem with a personnel announcement. Gemini 3.5 Pro had been planned for June, slipped repeatedly, and left a flagship gap of nearly half a year; the same day, Alphabet shares fell about 4%. After the market voted with its feet, Google's answer was not a new model but a leadership reshuffle.

Demis Hassabis stepped back from DeepMind's day-to-day management to become Alphabet chief scientist and DeepMind chairman; Koray Kavukcuoglu was promoted to SVP with independent charge of the Gemini line. Management was split: the research institution and the flagship product no longer shared one captain. For a company proud of being research-driven, this is a rare organizational signal—flagship cadence needs an owner, not a "we'll ship when ready" research team.

What shook the industry more was the departure of four veterans. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, Quoc Le—the list reads like the first half of Google's deep-learning history. They left to found Discovery Loop, with Google as founding investor. This was not an escape but an internal startup: the giant funds, the veterans build, and the new lab keeps a capital tie to the old house. Veteran departures from big labs had become routine over the past two years, but four infrastructure-level researchers at once still set a record.

Laid out on a timeline, the reshuffle has a clear causal chain: the June launch missed → July's Flash models held the production base while the flagship stayed absent → August's share pressure → leadership reshuffle. Google answered the delay with organization, effectively admitting the problem was not model capability but cadence management. For developers, Gemini's roadmap and leadership changed, and expectations for flagship release timing need recalibration.

What Discovery Loop will build, and whether friction exists between the four veterans and Google, are details outsiders cannot know and should not invent. What is confirmable is the outcome: Google's AI leadership completed a generational handover, Gemini is steered by Kavukcuoglu's new team, and a generation of infrastructure researchers entered a founding cycle. The flagship delay did not kill Google's AI ambitions, but it proved one thing—in 2026, model release cadence is itself a corporate-governance problem.