Google 发布 Gemini 2.0
从「聊天」走向「Agent 时代」的宣言式升级
Google 发布 Gemini 2.0 Flash,主打原生多模态与 Agentic 能力,并演示 Project Mariner 等浏览器代理。它把 Google 的战略重心从对话模型转向「能替你干活」的智能体。
2024 年底的大模型行业,出现了一个新的关键词:Agent。OpenAI 有能自主推理的 o 系列,Anthropic 有能操作电脑的 computer use,人人都说「下一个时代是模型替人干活」。但 Google 作为搜索巨头,这个叙事对它来说既诱人又危险——它必须证明自己不只是「聊天做得最好」,还要能「把事情办成」。
2024 年 12 月 11 日,Google 给出了答案:Gemini 2.0 Flash。这个版本原生支持图像、视频、音频输出,内置工具调用和多模态推理。但发布会的重头戏是演示——Project Mariner 是一个能像人一样操作浏览器的代理,Deep Research 能自主检索并整理报告,甚至还有一个能玩游戏的代理。整场发布会传递的核心信息只有一个:信息时代,正在走向智能体时代。
Gemini 2.0 的战略姿态非常明确。Google 没有把 Agent 当做一个附加功能,而是把它写进了旗舰模型的路线图里。它传递的信号是:未来的模型不是回答问题的,而是完成任务的。这种「Agent 优先」的宣告,让整个行业的产品叙事都跟着变了调。
当然,演示的惊艳与落地的困难总是相伴而行。浏览器代理面对复杂页面依然会出错,自主执行长任务还远谈不上稳定。Gemini 2.0 更多是宣告方向,而非交付终局。但方向的宣告,本身就很有分量。
回看 Gemini 2.0,它是 Google 在 Agent 时代的一次主动转身。它没有发明 Agent,但它让「模型替人干活」从一个模糊的愿景,变成了巨头路线图上的白纸黑字。此后几乎所有大厂的新品发布,都绕不开「Agent」这个词——而 Google 选择在 2024 年底的那个冬天,第一个把话说满。
By late 2024 a new keyword dominated the large-model industry: agent. OpenAI had its self-reasoning o-series, Anthropic had computer use that operated a machine, and everyone said "the next era is models doing work for people." For Google, as a search giant, this narrative was both tempting and dangerous—it had to prove it was not just "best at chat" but could also "get things done."
On December 11, 2024, Google gave its answer: Gemini 2.0 Flash. This version natively supported image, video, and audio output, with built-in tool use and multimodal reasoning. But the centerpiece of the launch was the demos—Project Mariner was an agent that operated a browser like a human, Deep Research could autonomously search and compile reports, and there was even an agent that played games. The core message of the entire event: the information age is moving into the agentic age.
Gemini 2.0's strategic posture was unmistakable. Google did not treat agents as an add-on feature but wrote them into the flagship model's roadmap. The signal: future models are not for answering questions but for completing tasks. This "agent-first" declaration shifted the product narrative across the industry.
Of course, impressive demos and difficult deployment always walk together. Browser agents still make mistakes on complex pages, and autonomous long-horizon execution is far from stable. Gemini 2.0 was more a declaration of direction than a delivery of the end state. But the declaration itself carried weight.
Looking back, Gemini 2.0 was Google's proactive turn in the agent era. It did not invent the agent, but it turned "models doing work for people" from a vague vision into black and white on a giant's roadmap. Since then, nearly every major vendor's launch has revolved around the word "agent"—and Google chose that winter of 2024 to be the first to say it out loud.
展开完整事件档案人物、主题、模型与产品
- 人物
- —
- 模型
- —
- 产品
- —