Name: 雙子座 3.0 已洩露：查看運行示例 (Gemini 3.0 has Leaked: See examples of it in action)
Uploaded: 2025-10-17T01:27:32.000Z
Duration: 10 min 7 s

分詞構句

Coming up this week, Gemini 3.0 leaks start to emerge.

這週的重點，Gemini 3.0 的消息開始流出。

We'll take a look at some examples of it in action.

我們會看看它實際運作的一些例子。

Linea's head of product shares how the company uses AI agents to stop interrupting engineers, a new tool to create your digital twin at work.

Linea 的產品負責人分享了公司如何使用 AI agents 來避免打斷工程師，以及一個在工作中創造你的 digital twin 的新工具。

And Anthropic's co founder warns us that AI is a real and mysterious creature.

而 Anthropic 的共同創辦人警告我們，AI 是一種真實且神祕的生物。

敬請鎖定所有這些內容以及更多。

And if you enjoy the briefing, hit the subscribe and the like button.

如果你喜歡這個簡報，請按下訂閱和喜歡按鈕。

So first up this week, Gemini 3.0 leaks have started to emerge, with rumors suggesting that Google's new model could be ready for release within a few weeks.

那麼這週的重點，Gemini 3.0 的消息已經開始流出，有傳言說 Google 的新模型可能在幾週內準備好發布。

Some users have gotten their hands on it this week, and

有些使用者這週已經搶先體驗了，

以下是它實際運作的一些例子。

In this example, the model has Vibe coded an entire website relating to the simulation hypothesis.

在這個例子中，這個模型已經 Vibe 編寫了關於模擬假說的整個網站。

And here's another example of it in action, where it creates the iPhone operating system in just one shot.

這裡還有另一個實際運作的例子，它一次就完成了 iPhone 作業系統的編寫。

這兩者看起來都相當令人印象深刻，而且

while Google 3.0 isn't yet ready for release, it does seem to be appearing in A B tests for some users.

雖然 Google 3.0 還沒準備好發布，但有些使用者在 A B 測試中似乎已經看見了它。

One tactic that eager Gemini fans have been using to get Access to Gemini 3.0 is to use an agentic browser like Perplexity's Comet to open up multiple instances of Google Studio until the Gemini 3.0 version of the A B test is shown.

一些熱切的 Gemini 粉絲為了搶先體驗 Gemini 3.0，採用的策略是使用像 Perplexity 的 Comet 這樣的 agentic browser，開啟 Google Studio 的多個實例，直到出現 Gemini 3.0 的 A B 測試版本。

Google has also introduced some new workspace updates this week that could be of interest to product teams.

Google 這週也推出了一些新的 workspace 更新，可能會讓產品團隊感興趣。

The first is a new Help me Schedule feature in Gmail powered by Gemini, which is designed to make it easier for you to find a meeting slot with a colleague.

第一個是 Gmail 中一個由 Gemini 驅動的全新 Help me Schedule 功能，旨在讓你更容易找到與同事開會的時間。

And the other is an updated version of its AI function in Google Sheets.

另一個則是 Google Sheets 中 AI 功能的更新版本。

The new update makes the AI function even more powerful by automatically adding results from Google Search.

新的更新透過自動加入 Google Search 的結果，讓 AI 功能更加強大。

One of the examples that Google suggests using is using the function to track real time customer sentiment after a product launch.

Google 建議的一個使用範例是，在產品發布後，利用這個功能追蹤即時的客戶情緒。

Another tool you can use for tracking customer feedback is, of course, Slack, and

當然，另一個可以用來追蹤客戶回饋的工具就是 Slack，而

this week Slack unveiled a major rebrand of the original workplace automation tool Slackbot to transform it into a workplace assistant.

這週 Slack 進行了重大的品牌重塑，將原本的協作自動化工具 Slackbot 轉變為工作場所助理。

The AI upgraded version of Slackbot can perform more complex tasks, such as organizing a product's launch plan inside a canvas by gathering information across different Slack channels.

升級為 AI 的 Slackbot 可以執行更複雜的任務，例如透過彙整不同 Slack 頻道中的資訊，在畫布中組織產品的發布計畫。

As AI assistants gain more agentic capabilities, the lines between AI assistant and agent are blurred, and

隨著 AI 助理獲得更多 agentic 能力，AI 助理和 agent 之間的界線變得模糊，

that's the topic for this week's Deep Dive over on Substack.

這就是本週 Substack 上的 Deep Dive 專題。

In this new piece, I explore the new powers of AI assistants from companies including Stripe, Slack, Adobe Shopify and others with downloadable UI examples of these assistants and their powers One of the companies featured in this deep dive is Intercom, and

在這篇新文章中，我探討了包括 Stripe、Slack、Adobe、Shopify 等公司 AI 助理的新功能，並提供這些助理及其功能的 UI 範例供下載。本週 Deep Dive 中介紹的公司之一是 Intercom，

they released Fin 3.0 in New York this past week, which comes with new agentic capabilities, including a feature called Procedures.

他們上週在紐約發布了 Fin 3.0，其中包含新的 agentic 功能，例如名為 Procedures 的功能。

This is a system for training an agent to complete multi step workflows from start to finish so for example, processing damaged order claims, refunding customers or troubleshooting accounts.

這是一個訓練 agent 從頭到尾完成多步驟工作流程的系統，例如處理損壞訂單的索賠、退款給客戶或疑難排解帳戶。

So if you're interested in learning more about the new powers of AI assistants, then check out this week's Deep Dive over on Substack.

所以，如果你對進一步瞭解 AI 助理的新功能感興趣，那就來看看本週 Substack 上的 Deep Dive 專題。

And continuing the theme of assistants and agents, LINEAR has published a page with a list of all of the AI agents that the company now supports.

延續助理和 agent 的主題，LINEAR 發布了一個頁面，列出了該公司目前支援的所有 AI agents。

These include agents for logging and fixing bugs or writing requirements, and more.

其中包括用於記錄和修復錯誤或撰寫需求的 agents 等等。

And this week their head of product revealed how the company uses AI agents internally to help them to ship small, non controversial changes like changing text or updating colors that would otherwise disrupt the productivity flow of an engineer.

而這週他們產品負責人也透露了公司內部如何運用 AI agents 來協助他們發布一些小型、無爭議的變更，像是修改文字或更新顏色，否則這些變更會打斷工程師的工作流程。

A product manager or non technical person can write up a spec delegated to an AI agent and then push it to production.

產品經理或非技術人員可以寫一份規格書，委派給 AI agent，然後直接發布到生產環境。

He says that the litmus test that they use for determining which task to delegate to AI agents are non controversial changes that are easier to do than asking an engineer to context switch.

他說，他們用來判斷哪些任務可以委派給 AI agents 的標準是：那些非爭議性的變更，比要求工程師切換上下文來處理還要容易。

But he says the company's product teams also use it for testing out new ideas that may never get shipped eventually.

但他表示，公司的產品團隊也用它來測試一些可能最終不會發布的新想法。

Elsewhere this week, Spotify has rolled out a new update to its DJ feature, which allows users to interact with it conversationally via text.

另外，這週 Spotify 也為他們的 DJ 功能推出了一項新更新，讓使用者可以透過文字與之進行對話互動。

Spotify says that since introducing voice requests earlier this year, DJ listening has increased, and

Spotify 表示，自從今年稍早推出語音指令以來，DJ 的收聽量有所增加，

this new way to interact with it conversationally could drive engagement up even further.

而這種新的對話式互動方式可能會進一步提升參與度。

And one company that has shown its commitment to conversational interfaces this week is Walmart.

本週展現對話式介面承諾的公司之一是 Walmart。

This week they confirmed a major partnership with OpenAI's ChatGPT that will allow users to purchase products conversationally through its new checkout feature.

他們本週證實與 OpenAI 的 ChatGPT 達成重大合作，讓使用者能透過其新的結帳功能以對話方式購買產品。

Approximately 40% of Walmart's revenue is now generated online, so that's roughly $240 billion.

Walmart 約有 40% 的營收來自線上，所以大約是 2400 億美元。

But but will Walmart's customers want to make purchases through a conversational interface rather than browsing a store?

不過，Walmart 的顧客會想透過對話介面購物，而不是瀏覽商店嗎？

Not everyone is convinced, but this will be a big test for consumer appetite for conversational commerce.

並非所有人都信服，但這將是對消費者對話式商務接受度的一大考驗。

Now let's take a look at some tools you can use, and

現在我們來看看一些你可以使用的工具，

我們將從一個叫做 Flint 的產品開始。

This product is backed by former Facebook exec Sheryl Sandberg, and

這款產品有前 Facebook 高管 Sheryl Sandberg 的支持，

it helps teams build what it calls autonomous websites.

它能協助團隊建立所謂的自主網站。

In other words, sites that generate, adapt and a B test pages themselves based on visitor behavior, market changes and AI traffic.

換句話說，這些網站能根據訪客行為、市場變化和 AI 流量，自行生成、調整和進行 A B 測試網頁。

So this will allow you to launch and optimize landing pages without interrupting developers and they even mention this in their copy on their website.

所以這能讓你推出並優化登陸頁面，而無需打斷開發人員，他們甚至在網站文案中也提到了這一點。

They say that this includes all the pages that engineering has no time for.

他們說這包含了工程師沒時間處理的所有頁面。

The way it works is you upload your company's URL and then upload your content separately and it will use your pre existing styles to build pages that match your company's brand.

它的運作方式是，你上傳公司的 URL，然後另外上傳你的內容，它會利用你現有的樣式來建立符合公司品牌的頁面。

So if you're looking for new ways to build landing pages, then maybe check out Flint.

所以如果你正在尋找建立登陸頁面的新方法，那或許可以看看 Flint。

Next is a product called Strawberry which describes itself as a self driving browser that comes with a team of AI companions, each with their own Persona.

下一個產品叫做 Strawberry，它自稱是一款自動駕駛瀏覽器，配備了一群 AI 夥伴，每個都有自己的 Persona。

These companions include competition Camilla for competitive analysis and recruiter Ryan to help you hire this week the company raised $6 million, but

這些夥伴包括用於競爭分析的 Camilla，以及協助你招聘的 Ryan，本週該公司籌集了 600 萬美元，但

given that there is so much competition in the AI browser space it'll be difficult for Strawberry to cut through and the final product for this week is called Vivn.

鑑於 AI 瀏覽器領域的競爭如此激烈，Strawberry 要脫穎而出將會很困難，而本週的最後一個產品叫做 Vivn。

This week they raised $35 million to create your digital twin at work and this product is designed to help users in the scenario when a colleague with vital information is away on vacation or in a different time zone.

本週他們籌集了 3500 萬美元來打造你在工作中的 digital twin，這款產品旨在幫助用戶應對同事因休假或身處不同時區而無法提供關鍵資訊的狀況。

The problem they solve, they say, is that typically progress must be delayed until that person responds.

他們說，他們解決的問題是，通常必須等到那個人回應後才能繼續進展。

But with a digital twin that will no longer be the case.

但有了 digital twin，情況就不會再是這樣了。

The founders say that the product has been trained to sound and act just like the person it is emulating, but

創辦人表示，該產品經過訓練，聽起來和行為舉止都與它模仿的人一模一樣，但

it also knows what information to keep private.

imagine returning to the office only to learn that your digital twin has made some fundamental decisions that you disagree with.

想像一下回到辦公室，卻發現你的 digital twin 已經做了一些你不同意的根本性決定。

Nevertheless, if you're someone who likes the idea of someone else stepping into your shoes whilst you're away, then Vivint could be worth checking out.

不過，如果你是那種喜歡在你離開的時候，有人能頂替你位置的想法，那 Vivint 倒是值得你考慮。

Now let's take a look at some data and trends for the week.

現在我們來看看本週的一些數據和趨勢。

And first up is some new data from Opendoor's CEO who says that their procurement team is saving millions of dollars a year by canceling SaaS products that they don't use.

首先是 Opendoor 執行長分享的新數據，他說他們的採購團隊透過取消使用中的 SaaS 產品，每年省下數百萬美元。

Their CEO says that they'll go from paying 10 different SaaS providers in one department down to just one after they conducted an AI powered audit of SaaS product usage.

該執行長表示，在進行了 AI 驅動的 SaaS 產品使用情況審計後，他們將一個部門的 10 家不同 SaaS 供應商，減少到只剩一家。

This is of course excellent news for CFOs but not quite so exciting for the SaaS companies who rely on that revenue.

這對 CFO 來說當然是個好消息，但對於依賴這筆收入的 SaaS 公司來說，就沒那麼令人興奮了。

This this week also saw the release of the annual State of AI report from Air Street Capital.

本週也發布了 Air Street Capital 的年度 AI 狀態報告。

Now this report is over 300 pages long, so we don't quite have time to cover everything, but

這份報告有 300 多頁，我們沒辦法全部講完，但

here are some stats worth knowing for product teams.

有些數據對產品團隊來說非常值得知道。

The report shows that Claude and Cursor are the most adopted tool this year, while ChatGPT and Perplexity are the most dropped.

報告顯示，Claude 和 Cursor 是今年採用率最高的工具，而 ChatGPT 和 Perplexity 則是淘汰率最高的。

LangChain is the most popular framework for product teams who are building AI agents.

LangChain 是產品團隊在建置 AI agents 時最受歡迎的框架。

The average contract value for an AI enterprise software has grown from 39,000 in 2023 to $1 million in 2025.

AI 企業軟體的平均合約價值已從 2023 年的 39,000 美元成長到 2025 年的 100 萬美元。

AI first companies are growing from launch to 5 million in annual recurring revenue at a 1.5 faster rate than the top 100 SaaS companies in 2018.

AI 優先的公司從推出到年經常性收入達到 500 萬美元的成長速度，比 2018 年排名前 100 的 SaaS 公司快 1.5 倍。

And perhaps unsurprisingly, MCP is the most adopted protocol for agents with an estimated 15,000 servers globally.

而且，MCP 成為採用率最高的 agent 協定，全球估計有 15,000 臺伺服器，這或許並不令人意外。

So if you're interested in digging deeper into that report, then the State of AI report from Airstreak Capital is definitely worth a look.

所以，如果你對深入研究這份報告有興趣，那麼 Airstreak Capital 的 State of AI report 絕對值得一看。

And one company that has leaned heavily into MCP is Stripe, and

而 Stripe 就是大力採用 MCP 的其中一家公司，

this week they revealed that they've seen a 32% week on week growth in the number of users of its MCP server with the top MCP use cases including searching product catalogs, creating or retrieving prices and reading docs.

本週他們透露，使用其 MCP 伺服器的使用者數量每週成長了 32%，其中主要的 MCP 使用案例包括搜尋產品目錄、建立或擷取價格以及閱讀文件。

Other analysis this week shows us the Claude code and OpenAI's codecs are now neck and neck for coding pull request approvals at 74% each, but

本週的其他分析顯示，Claude code 和 OpenAI 的 codecs 在 coding pull request 核准率上不相上下，各為 74%，但

at the moment neither of them top the current leaderboard and that achievement goes to San Francisco based startup Amp Code.

目前兩者都未能登上目前的排行榜首位，這項成就歸屬於舊金山的新創公司 Amp Code。

for European startups at least, perhaps what is needed is AI coding agents that can not only build features, but

對歐洲新創公司來說，或許真正需要的是 AI coding agents，它們不僅能開發功能，

can also ensure they're built in a way that meets regulations.

According to a new survey, almost 60% of small European tech companies say that they've experienced delays in rolling out new AI product features to due to regulation, versus 44% in the US 1/3 of European developers at small tech companies say that they've had to remove or downgrade features in order to comply with requirements, largely due to data handling and safety checks.

根據一項新調查，近 60% 的歐洲小型科技公司表示，因為法規問題，他們在新 AI 功能推出時遇到了延遲，而美國的比例是 44%。在小型科技公司中，有 1/3 的歐洲開發者表示，為了符合要求，他們不得不移除或降級功能，這主要是由於資料處理和安全檢查。

And finally this week, Anthropic's co founder has warned humanity about the risks of AI, saying that what we're dealing with is a real and mysterious creature.

最後，本週 Anthropic 的共同創辦人警告人類 AI 的風險，表示我們面對的是一個真實且神祕的生物。

He argues that as systems become more sophisticated, they also become more aware that they're being tested.

他認為，隨著系統越來越複雜，它們也會越來越意識到自己正在被測試。

And so in these graphs you can see that Sonnet 4.5 currently leads the software engineering benchmark test.

在這些圖表中，你可以看到 Sonnet 4.5 目前在軟體工程基準測試中領先。

But in this graph you can see that as the latest AI models get better and better, so too does their situational awareness score.

但在這張圖表中，你可以看到隨著最新的 AI 模型越來越好，它們的態勢感知分數也隨之提高。

In other words, as these models become more sophisticated, they also become more aware of the fact that they're being tested and audited.

換句話說，隨著這些模型變得越來越複雜，它們也越來越意識到自己正在被測試和審查。

The this, according to Jack Clarke, means that what we're dealing with here is a mysterious and real creature and not just a simple and predictable machine.

根據 Jack Clarke 的說法，這意味著我們面對的是一個神祕且真實的生物，而不僅僅是一個簡單且可預測的機器。

And on that note, I'll leave it there for this week.

說到這裡，我這週就先說到這裡。

Thanks very much for listening and watching.

I'll be back next week with another briefing.