
经过数周在社交媒体上的造势、语焉不详的帖子和眨眼表情符号的预热后,谷歌正式发布了其Gemini 3人工智能模型。当日凌晨时分,Pro版本的模型卡(概述了系统关键细节及其基准性能)遭泄露,开发者们随即在X平台上发帖,气氛热烈得如同圣诞老人提前到来。甚至连前OpenAI研究员兼联合创始人安德烈·卡帕西(Andrej Karpathy)也调侃这场预热,他在X平台上写道:“听说Gemini 3能在你提问前就给出答案,还能和你的猫对话。”
当然,该模型能否不负众望,如一位X用户所言“绝对碾压”所有其他尖端模型,包括OpenAI的GPT-5、Anthropic的Claude Sonnet 4.5/Opus 4以及xAI的Grok 4,仍有待观察。但显而易见的是,谷歌此次信心十足地广泛发布Gemini 3的Pro版和“Deep Think”(深度思考)版,与2023年12月初代Gemini模型试探性的首次亮相已相去甚远——彼时,该公司因模型输出“觉醒”内容以及不符合史实或不准确的图像和文本而面临强烈反弹,最终承认“未能达到预期”。其搜索引擎中由Gemini驱动的AI概览功能也曾因系统著名地建议用户食用胶水和石头而在网上引发轩然大波。
谷歌首次在发布首日将Gemini集成至搜索引擎
这一次,Gemini 3将在发布首日全面覆盖谷歌拥有数十亿用户的庞大生态系统——包括其史上最快部署至谷歌搜索引擎。“这是我们首次在搜索服务中推出我们最新的Gemini模型,”谷歌搜索产品副总裁罗比·斯泰因(Robby Stein)在一次媒体预览会上表示。这意味着谷歌的AI Pro和Ultra订阅用户可以在搜索的AI模式下使用Gemini 3,该模式采用了新的可视化布局,包含交互式工具以及图像、表格和网格等元素。
谷歌还受益于一个事实:与过去几次AI发布时不同,OpenAI这次未能抢走它的风头。OpenAI已于8月推出了其备受瞩目的GPT-5模型,但许多观察人士称其发布效果不及预期,令人失望。上周,该公司发布了5.1版本更新,描述其为“更智能”和“更具对话性”,并提供八种不同的“人格”供选择,但这仍然为谷歌掀起Gemini浪潮敞开了大门。
在一篇介绍Gemini 3的博客文章中,Alphabet和谷歌首席执行官桑达尔·皮查伊(Sundar Pichai)宣称,谷歌的AI概览功能现在每月拥有20亿用户,而Gemini应用每月活跃用户超过6.5亿,并且有超过1300万开发者正在使用谷歌“最智能的模型”Gemini进行开发。他写道,如今,“我们正以谷歌的规模来部署Gemini。”
人工智能行业基准测试与安全性提升
谷歌还大肆宣扬该模型在主要AI行业基准测试中的结果,称其在每一项主要的推理测试中都击败了早期的Gemini 2.5 Pro。谷歌表示,该模型在测试逻辑、数学、科学和问题解决的学术风格挑战中表现极其出色,达到了谷歌声称的类似“博士水平”的推理分数。谷歌还表示该模型在事实准确性方面有所改进。
谷歌还宣称该模型在对话中更具思考深度和实用性:它旨在提供更清晰、更直接的见解,而不是给出千篇一律的奉承或充满流行语的回答——这是许多聊天机器人回复中令人非常反感的特点。
此外,谷歌表示Gemini 3“经历了谷歌迄今为止任何AI模型中最全面的安全评估集”,并补充说该模型表现出“减少了奉承倾向,增强了对提示注入攻击的抵抗力,并改进了针对通过网络攻击滥用的防护”。过去一年中,AI安全专家分享了Gemini易受提示注入攻击的许多例子,即攻击者通过在其输入中嵌入恶意指令来操纵模型,以及其他类型的威胁。
安抚出版商
尽管出版商越来越担心谷歌的AI概览功能正在引发一场“流量末日”,扼杀了新闻网站的点击率,谷歌仍继续坚称将保持用户与出版商内容的连接。这一安抚表态的背景是,有研究表明当AI摘要出现时,用户点击结果链接的可能性会降低——而且当AI摘要确实列出了信息来源时,用户也很少点击进入。
“我们每天继续向网络输送数十亿次点击量,并且我们在搜索AI体验中突出展示网页内容,以鼓励用户进一步探索,”一位谷歌发言人在发给《财富》杂志的电子邮件中表示。“一如既往,我们在整个AI模式响应中持续突出显示网页链接,以便人们能够持续学习和探索。”
谷歌还提到了其“查询扇出技术”——本质上,该技术将用户的单个问题分解成许多更小的、在后台进行的搜索,以收集更多相关信息。
“现在,它不仅能执行更多搜索来发现相关的网页内容,而且由于Gemini能更智能地理解您的意图,它可以找到之前可能遗漏的新内容,”该发言人表示。“这意味着搜索能帮助您为特定问题找到相关性更高的网页内容。”
谷歌发挥自身优势
无论Gemini 3反响如何,毫无疑问,谷歌的地位已远高于不到三年前ChatGPT出现引发内部“红色警报”之时。该公司也正在发挥其优势,直接关注其数十亿消费者的需求——包括在Gemini应用中推出首创的生成式购物界面,该界面包含从谷歌拥有500亿条目的购物图谱中提取的商品列表、对比表格和实时定价。
当然,这就是谷歌不那么秘密的制胜法宝:每天流经其产品的海量数据。而Gemini 3再次提醒人们,几乎没有公司(如果有的话)拥有这样的数据基础或全球影响力,能够以如此规模部署AI。
尽管如此,即使是公司首席执行官皮查伊,在谈及人工智能时仍敦促人们保持谨慎。在接受英国广播公司(BBC)的一次新采访中,他表示人们不应“盲目相信”AI工具告诉他们的一切,并补充说它们“容易出错”,敦促人们将其与其他工具结合使用。皮查伊还警告称,如果AI泡沫破裂,没有公司能够幸免。想来,即使是谷歌也不例外。(*)
译者:中慧言-王芳
经过数周在社交媒体上的造势、语焉不详的帖子和眨眼表情符号的预热后,谷歌正式发布了其Gemini 3人工智能模型。当日凌晨时分,Pro版本的模型卡(概述了系统关键细节及其基准性能)遭泄露,开发者们随即在X平台上发帖,气氛热烈得如同圣诞老人提前到来。甚至连前OpenAI研究员兼联合创始人安德烈·卡帕西(Andrej Karpathy)也调侃这场预热,他在X平台上写道:“听说Gemini 3能在你提问前就给出答案,还能和你的猫对话。”
当然,该模型能否不负众望,如一位X用户所言“绝对碾压”所有其他尖端模型,包括OpenAI的GPT-5、Anthropic的Claude Sonnet 4.5/Opus 4以及xAI的Grok 4,仍有待观察。但显而易见的是,谷歌此次信心十足地广泛发布Gemini 3的Pro版和“Deep Think”(深度思考)版,与2023年12月初代Gemini模型试探性的首次亮相已相去甚远——彼时,该公司因模型输出“觉醒”内容以及不符合史实或不准确的图像和文本而面临强烈反弹,最终承认“未能达到预期”。其搜索引擎中由Gemini驱动的AI概览功能也曾因系统著名地建议用户食用胶水和石头而在网上引发轩然大波。
谷歌首次在发布首日将Gemini集成至搜索引擎
这一次,Gemini 3将在发布首日全面覆盖谷歌拥有数十亿用户的庞大生态系统——包括其史上最快部署至谷歌搜索引擎。“这是我们首次在搜索服务中推出我们最新的Gemini模型,”谷歌搜索产品副总裁罗比·斯泰因(Robby Stein)在一次媒体预览会上表示。这意味着谷歌的AI Pro和Ultra订阅用户可以在搜索的AI模式下使用Gemini 3,该模式采用了新的可视化布局,包含交互式工具以及图像、表格和网格等元素。
谷歌还受益于一个事实:与过去几次AI发布时不同,OpenAI这次未能抢走它的风头。OpenAI已于8月推出了其备受瞩目的GPT-5模型,但许多观察人士称其发布效果不及预期,令人失望。上周,该公司发布了5.1版本更新,描述其为“更智能”和“更具对话性”,并提供八种不同的“人格”供选择,但这仍然为谷歌掀起Gemini浪潮敞开了大门。
在一篇介绍Gemini 3的博客文章中,Alphabet和谷歌首席执行官桑达尔·皮查伊(Sundar Pichai)宣称,谷歌的AI概览功能现在每月拥有20亿用户,而Gemini应用每月活跃用户超过6.5亿,并且有超过1300万开发者正在使用谷歌“最智能的模型”Gemini进行开发。他写道,如今,“我们正以谷歌的规模来部署Gemini。”
人工智能行业基准测试与安全性提升
谷歌还大肆宣扬该模型在主要AI行业基准测试中的结果,称其在每一项主要的推理测试中都击败了早期的Gemini 2.5 Pro。谷歌表示,该模型在测试逻辑、数学、科学和问题解决的学术风格挑战中表现极其出色,达到了谷歌声称的类似“博士水平”的推理分数。谷歌还表示该模型在事实准确性方面有所改进。
谷歌还宣称该模型在对话中更具思考深度和实用性:它旨在提供更清晰、更直接的见解,而不是给出千篇一律的奉承或充满流行语的回答——这是许多聊天机器人回复中令人非常反感的特点。
此外,谷歌表示Gemini 3“经历了谷歌迄今为止任何AI模型中最全面的安全评估集”,并补充说该模型表现出“减少了奉承倾向,增强了对提示注入攻击的抵抗力,并改进了针对通过网络攻击滥用的防护”。过去一年中,AI安全专家分享了Gemini易受提示注入攻击的许多例子,即攻击者通过在其输入中嵌入恶意指令来操纵模型,以及其他类型的威胁。
安抚出版商
尽管出版商越来越担心谷歌的AI概览功能正在引发一场“流量末日”,扼杀了新闻网站的点击率,谷歌仍继续坚称将保持用户与出版商内容的连接。这一安抚表态的背景是,有研究表明当AI摘要出现时,用户点击结果链接的可能性会降低——而且当AI摘要确实列出了信息来源时,用户也很少点击进入。
“我们每天继续向网络输送数十亿次点击量,并且我们在搜索AI体验中突出展示网页内容,以鼓励用户进一步探索,”一位谷歌发言人在发给《财富》杂志的电子邮件中表示。“一如既往,我们在整个AI模式响应中持续突出显示网页链接,以便人们能够持续学习和探索。”
谷歌还提到了其“查询扇出技术”——本质上,该技术将用户的单个问题分解成许多更小的、在后台进行的搜索,以收集更多相关信息。
“现在,它不仅能执行更多搜索来发现相关的网页内容,而且由于Gemini能更智能地理解您的意图,它可以找到之前可能遗漏的新内容,”该发言人表示。“这意味着搜索能帮助您为特定问题找到相关性更高的网页内容。”
谷歌发挥自身优势
无论Gemini 3反响如何,毫无疑问,谷歌的地位已远高于不到三年前ChatGPT出现引发内部“红色警报”之时。该公司也正在发挥其优势,直接关注其数十亿消费者的需求——包括在Gemini应用中推出首创的生成式购物界面,该界面包含从谷歌拥有500亿条目的购物图谱中提取的商品列表、对比表格和实时定价。
当然,这就是谷歌不那么秘密的制胜法宝:每天流经其产品的海量数据。而Gemini 3再次提醒人们,几乎没有公司(如果有的话)拥有这样的数据基础或全球影响力,能够以如此规模部署AI。
尽管如此,即使是公司首席执行官皮查伊,在谈及人工智能时仍敦促人们保持谨慎。在接受英国广播公司(BBC)的一次新采访中,他表示人们不应“盲目相信”AI工具告诉他们的一切,并补充说它们“容易出错”,敦促人们将其与其他工具结合使用。皮查伊还警告称,如果AI泡沫破裂,没有公司能够幸免。想来,即使是谷歌也不例外。(*)
译者:中慧言-王芳
Google released its Gemini 3 AI model today after weeks of social media hype, vague posting, and wink emojis. An early-morning leak of the Pro version's model card---which outlines key details about the system and its benchmark performance---had developers posting on X as though Santa had arrived early. Even former OpenAI researcher and cofounder Andrej Karpathy joked about the buildup: “I heard Gemini 3 answers questions before you ask them. And that it can talk to your cat,” he wrote on X.
It remains to be seen, of course, whether the model lives up to the hype that it would, as one X user put it, “absolutely crush” all other state-of-the-art models, including OpenAI's GPT-5 and Anthropic's Claude Sonnet 4.5/Opus 4 and xAI's Grok 4. But what is clear is that Google's confident, widespread release of Gemini 3, in Pro and “Deep Think” versions, is a long way from its tentative debut of the first Gemini model in December 2023---after which the company faced intense backlash over “woke” outputs and ahistorical or inaccurate images and text, ultimately admitting it had “missed the mark.” Its Gemini-powered AI Overviews in Search also triggered an online furor after the system famously told users to eat glue and rocks.
First time Google adds Gemini to Search on day one
This time around, Gemini 3 is getting a sweeping day-one rollout across a large swath of Google's ecosystem with its billions of users---including its fastest-ever deployment into Google Search. “This is the very first time we're shipping our latest Gemini model in search,” said Robby Stein, vice president of product for Google Search, in a press preview. That includes Google's AI Pro and Ultra subscribers getting access to Gemini 3 in Search's AI Mode, with new visual layouts featuring interactive tools and elements like images, tables, and grids.
Google also benefits from the fact that, unlike during past AI rollouts, OpenAI didn't manage to steal its thunder this time. OpenAI already debuted its massively hyped GPT-5 model in August, a release many observers said fell short and was underwhelming. Last week the company released a 5.1 update it described as “smarter” and “more conversational,” with eight different “personalities” to choose from, but that still left the door wide open for Google to make Gemini waves.
In a blog post introducing Gemini 3, Alphabet and Google CEO Sundar Pichai boasted that Google's AI Overviews now has 2 billion users per month, while the Gemini app has more than 650 million active monthly users, and more than 13 million developers are building with Gemini, Google's “most intelligent model.” Today, he wrote, “we're shipping Gemini at the scale of Google.”
Improvements on AI industry benchmarks and safety
Google also crowed about the model's results on major AI industry benchmarks, saying that it beat the earlier Gemini 2.5 Pro on every major test of reasoning. It said the model performs extremely well on academic-style challenges testing logic, math, science, and problem-solving, reaching scores that Google claimed resemble “PhD-level” reasoning. It also said the model improved on factual accuracy.
Google also claimed the model is more thoughtful and useful in conversation: Instead of giving generic flattery or buzzword-filled answers---much-disliked features of many chatbot responses---it's supposed to offer clearer, more direct insight.
In addition, Google said Gemini 3 has “undergone the most comprehensive set of safety evaluations of any Google AI model to date,” adding that the model shows “reduced sycophancy, increased resistance to prompt injections and improved protection against misuse via cyberattacks.” Over the past year, AI security experts have shared many examples of Gemini's vulnerability to prompt injections, in which attackers manipulate the model by embedding malicious instructions into its input, and other types of threats.
Reassuring publishers
Amid rising publisher fears that Google's AI Overviews are causing a “traffic apocalypse” that kills click-throughs to news sites, Google continued to insist that it will keep connecting users to publisher content. That reassurance comes despite research showing that users are less likely to click on result links when an AI summary appears---and that when AI summaries do surface sources, users rarely click through to them.
“We continue to send billions of clicks to the web every day, and we're prominently highlighting the web in our Search AI experiences in a way that encourages onward exploration,” a Google spokesperson told Fortune by email. “As always, we continue to prominently display links to the web throughout the AI Mode response, so people can continue learning and exploring.”
Google also pointed to its “query fan-out technique”---essentially, taking a user's single question and breaking it into many smaller, behind-the-scenes searches to gather more relevant information.
“Now, not only can it perform even more searches to uncover relevant web content, but because Gemini more intelligently understands your intent, it can find new content that it may have previously missed,” the spokesperson said. “This means Search can help you find even more highly relevant web content for your specific question.”
Google is playing to its strengths
No matter how Gemini 3 is received, there's little question that Google is far ahead of where it stood less than three years ago, when ChatGPT's arrival sparked an internal “code red.” The company is also playing to its strengths, looking directly at what its billions of consumers want---including unveiling first-of-its-kind generative shopping interfaces in the Gemini app with product listings, comparison tables, and live pricing pulled from Google's 50-billion-item Shopping Graph.
That, of course, is Google's not-so-secret sauce: the massive amounts of data that flows through its products every day. And Gemini 3 is yet another reminder that few companies, if any, have the data foundation or the global reach to ship AI at this scale.
Still, even Pichai, the company's CEO, is still urging caution when it comes to AI. In a new interview with the BBC, he said people should not “blindly trust” everything AI tools tell them, adding they are “prone to errors” and urging people to use them alongside other tools. Pichai also warned that no company would be immune if the AI bubble burst. Presumably even Google.
