SuperAI 2026

He has advised some of the world's biggest technology firms, but now he is just listed in your program as analyst.

他曾為全球幾家最大的科技公司擔任顧問，但在你們的議程手冊上，他的頭銜只寫著「分析師」。

But he is not just an analyst, he is the analyst.

但他不只是個分析師，他是「那位」分析師。

Twice a year the entire technology world waits for him to deliver his next big idea.

一年兩次，整個科技圈都在等他丟出下一個大想法。

But the big ideas are always backed up with deep cut granular economic analysis.

而這些大想法背後，永遠有著扎實、細到骨子裡的經濟數據分析在支撐。

So he is asking the big questions: the field-of-dreams approach to capex in AI that is happening right now — build it and the business will come.

所以他問的是那些最根本的大問題：眼下這股 AI capex 的「夢田心態」（build it and they will come）—— 只要蓋起來，生意自然會上門。

Really? Is it going to transform our lives? Is it going to destroy the working world?

真的是這樣嗎？它會徹底改變我們的生活嗎？它會摧毀整個職場世界嗎？

I don't know, but Benedict Evans does.

我不知道，但 Benedict Evans 知道。

He's coming to deliver some answers to it. Let's hear it for Benedict Evans, come on!

他要來給我們一些答案。讓我們掌聲歡迎 Benedict Evans，來吧！

So, thanks for that introduction.

謝謝這段介紹。

I really must send that to my mother-in-law, so she'll understand what a big deal I am.

我真該把這段話傳給我岳母，好讓她搞懂我是個多了不起的人物。

So, I was asked to come and explain AI, and it always seems to me that the primary question is how you can break it apart.

我被找來解釋 AI 這件事，而我一直覺得，最關鍵的問題其實是：你要怎麼把它拆解開來看。

What kinds of questions do you want to ask?

你想問的，到底是哪一類的問題？

In fact, it's a fascinating list of analogies, because he spoke to us for 20 or 30 minutes, and everything he said was very sensible and interesting, and I agree with most of it.

事實上，前一位講者給的那一串類比很有意思，因為他講了 20、30 分鐘，每句話都很合理、很有趣，我大部分也都同意。

But it was all completely different questions to the questions I'm talking about, which is kind of an interesting part of the point.

但他談的全是另一套問題，跟我要談的完全不一樣 —— 這本身就是個有意思的重點。

There are so many different things that change here.

這裡頭會被改變的東西實在太多了。

And the things that I wanted to talk about, the way that I wanted to break this apart, was to think firstly about capital, and then secondly about deployment, and then thirdly about what all this might change.

而我想談的、我想拆解這件事的方式，是先談 capital（資本），再談 deployment（部署），最後談這一切可能會改變什麼。

And so firstly, capital — this is what we see on the front page of all the newspapers: all of this money.

那先從 capital 講起 —— 這就是我們在每份報紙頭版上看到的東西：滿滿的錢。

So most people will be familiar with a framing like this, that every industry tends to move in platform shifts.

大多數人對這套框架應該都不陌生：每個產業大致都是隨著 platform shift（平台轉移）在往前走。

Every 10 or 15 years, there's some new thing, and everything gets built around that.

每隔 10 到 15 年，就冒出某個新東西，然後所有事情都圍繞著它重新被打造。

So we had mainframes and PCs and the web and smartphones, and now we have AI.

所以我們經歷過 mainframe（大型主機）、PC、web、smartphone，現在則輪到 AI。

And as you go through these shifts, a bunch of things happen.

而每次走過這種轉移，都會發生一連串的事。

So inside the tech industry, all of the innovation and investment and company creation gets built around the new thing.

在科技業內部，所有的創新、投資、新公司，全都圍著那個新東西長出來。

Like the old thing is still there. Big companies run on PCs, airlines run on mainframes.

舊的東西當然還在。大公司還是靠 PC 在跑，航空公司還是跑在 mainframe 上。

But all the new stuff happens on the new thing.

但所有新鮮事都發生在新東西上。

And so inside technology, there's a big change.

所以在科技業內部，會出現一場大洗牌。

Some of the existing gatekeepers, some of the existing tech giants go away or become irrelevant. There are new ones that emerge — new and bigger markets.

某些既有的守門人、某些既有的科技巨頭會消失，或者變得無足輕重；同時冒出新的巨頭 —— 帶來更新、更大的市場。

Outside of technology, people have to ask themselves: well, is this just kind of a new tool? Will there just be more software? Is this new revenue?

在科技業外面，人們則得問自己：這只是個新工具嗎？無非就是更多軟體？這算不算是一筆新的營收來源？

Or is this some kind of existential threat to our business or industry? Like, how much do we need to think about this?

還是說，這對我的生意、我的產業是某種攸關存亡的威脅？我到底該為它操多少心？

Inside the tech industry, therefore, those companies are looking at the risks of their business and the size of the opportunity.

所以在科技業內部，這些公司一邊盯著自家生意的風險，一邊衡量機會有多大。

And so we've had these quotes from all of these companies saying, basically, the risk of missing this is bigger than the risk of over-investing.

於是我們聽到這些公司一個個放話，大意都是：錯過這波的風險，比過度投資的風險更大。

And that investment is certainly pretty substantial.

而這個投資金額，確實相當可觀。

So this year, those big tech companies will spend something like $700 billion.

今年，這些大型科技公司大概會花掉 7,000 億美元。

That's their guidance, although the numbers will probably go up in the course of this year, up from about $400 billion last year.

這是他們自己給的財測，而且今年數字大概還會往上修，相較去年大約 4,000 億美元。

The telecoms industry is about $300 billion. Oil and gas, probably a trillion, depending on how you break it up.

電信業一年大約 3,000 億美元；石油與天然氣業，看你怎麼算，大概一兆美元。

So these are pretty big numbers. This is what big global capital-intensive industries spend.

所以這些是相當龐大的數字。這是全球那些 capital-intensive（資本密集）大產業才花得起的規模。

Something like a third of that drops through to NVIDIA, which you can't really keep up with.

這裡頭大約三分之一最後落進了 NVIDIA 的口袋，那速度你根本追不上。

And these are quarterly numbers, and they're accelerating up and to the right, leaving NVIDIA far behind, heading to $100 billion a quarter, if nothing slows down.

而這些還只是單季數字，曲線一路往右上加速，把其他人遠遠甩在後頭，照這樣下去 NVIDIA 一季要衝向 1,000 億美元。

And that, in turn, means that the semiconductor industry is seeing an unprecedented surge in revenue, because, of course, that flows through to things like memory as well.

這也連帶讓整個半導體產業迎來前所未見的營收暴衝，因為這股錢潮當然也會流進像 memory 這類環節。

This has historically always been a cyclical industry, but if this is another cycle, then it's certainly a bigger cycle.

半導體向來是個 cyclical（景氣循環）的產業，但如果這又是一次循環，那它肯定是規模更大的一次。

And so that drops through to the capex from the semiconductor firms in this part of the world — particularly TSMC and Samsung — substantially increasing their capex this year.

這股錢潮也一路傳導到這個區域的半導體廠商身上 —— 尤其是 TSMC 和 Samsung —— 讓他們今年的 capex 大幅拉高。

Now, the thing that sits at the back of all of this, of course, is the cash flow from those big, giant tech companies who could afford this.

而撐在這一切背後的，當然是那些大到付得起這筆錢的科技巨頭所產生的現金流。

Although, not necessarily.

不過，也未必那麼穩。

We're starting to run into financial gravity, with Meta and Microsoft and Alphabet pushing towards or past 50% of their revenue going on capex this year.

我們開始撞上「財務地心引力」了 —— Meta、Microsoft、Alphabet 今年把 capex 推到逼近、甚至超過營收的 50%。

Which is why we saw last week Alphabet raising $85 billion in equity. Meta apparently considered the same.

這就是為什麼上週我們看到 Alphabet 透過股權募資 850 億美元；Meta 顯然也在考慮同樣的做法。

And then there's all sorts of structure — so SPVs and debt and VC and off-balance-sheet vehicles — and various kinds of trying to structure how you can pay for all of this stuff.

接著就出現各種財務結構 —— SPV（特殊目的公司）、舉債、創投資金、資產負債表外的工具 —— 用形形色色的方式去設計，到底要怎麼籌錢付這一切。

Meanwhile, there's this little company called OpenAI that's been trying to join the club.

與此同時，有一家叫 OpenAI 的「小公司」，一直想擠進這個俱樂部。

Various eye-catching announcements of gigawatts and trillions and numbers that don't necessarily seem to have translated into any actual construction yet.

各種搶眼的公告滿天飛，動不動就是 gigawatt、上兆美元的數字，但這些數字似乎還沒真的變成任何實際的建設。

And, of course, because it doesn't have the cash flow that those big companies do, that's all on other people's balance sheets — and this wonderful, very circular revenue.

而且當然，因為它沒有那些大公司那樣的現金流，所以這些全都掛在別人的資產負債表上 —— 再加上那種美妙到不行、非常 circular（循環）的營收。

And, of course, the challenge is that it's not actually very easy to spend that much money that quickly.

而當然，真正的挑戰在於：要在這麼短的時間內花掉這麼多錢，其實沒那麼容易。

So we have bottlenecks in GPUs and in memory.

我們在 GPU 和 memory 上都遇到了瓶頸。

Most places that have power are backed up multiple years, access to power equipment and gas turbines are backed up multiple years.

大部分有電力的地方都排隊排了好幾年，電力設備和燃氣渦輪機（gas turbine）的供貨也都得等上好幾年。

And then there's a growing political backlash, particularly in the USA, around data centres.

再來，針對資料中心的政治反彈也越來越大，尤其是在美國。

So there's a sort of a "pig in the python" problem, that you can't actually spend this money this quickly.

所以這就成了一種「蟒蛇吞豬」的問題 —— 你根本沒辦法這麼快把這筆錢花出去。

Which gets you to have people at Microsoft saying it's just been impossible to build this capacity fast enough.

這也是為什麼會聽到 Microsoft 內部的人說，要把這些算力建得夠快，根本是不可能的任務。

Now, as I think most people in the room will know, the thing that's really driven this, and indeed the thing that's accelerated enormously this year, has been agentic coding.

而我想在座大多數人都知道，真正推動這一切、而且今年加速得最猛的，就是 agentic coding（代理式程式開發）。

And they've reported this quite unprecedented surge in their revenue run rate.

這些公司都回報了營收年化執行率（run rate）出現前所未見的暴衝。

I should be a little bit careful here to compare these, because one is net and one is gross, but clearly there's an enormous willingness suddenly to spend on this.

這裡我得稍微小心一點去比較這些數字，因為有的是淨額、有的是毛額；但很清楚的是，大家突然之間非常願意為這件事掏錢。

Which drives an enormous amount of other consequences in terms of both capacity and, of course, cost.

這又連帶引發了一大堆後果，無論是在算力供給、還是當然在成本這兩方面。

So Uber managed to spend their entire annual budget for AI in the first quarter, driven by that surge in usage.

就像 Uber，光是第一季就把一整年的 AI 預算花光了，原因就是用量暴增。

A whole bunch of other companies have had similar kind of questions.

還有一大票其他公司也碰到類似的狀況。

And so this kind of creates a problem, because we have, at the moment, a huge disequilibrium.

這就製造出一個問題，因為我們此刻正處在一種巨大的失衡狀態。

We have explosive demand growth, which is far ahead of supply.

我們有爆炸性的需求成長，遠遠跑在供給前面。

On the other hand, we have inference efficiency growing 50, 100, 200x per year, depending on who you're counting.

另一方面，inference（推論）效率每年成長 50 倍、100 倍、甚至 200 倍，看你問的是誰。

We have that trillion dollars of capex coming down the pipe.

我們有那筆上兆美元的 capex 正一路湧進管線。

We don't know what the next model will be, what the next business model will be, whether there will be edge or open source, or what the cost lines are going to be.

我們不知道下一個 model 會長什麼樣、下一個商業模式是什麼、會不會有 edge（邊緣運算）或 open source，也不知道成本曲線會落在哪。

And of course it's kind of worth pointing out that all of this demand is only being driven by product-market fit in coding, which is actually a pretty small market.

而且值得一提的是，這一切需求，目前其實只靠「coding 找到了 product-market fit」在撐 —— 而 coding 其實是個相當小的市場。

Imagine what would be going on if we had product-market fit in something that was actually popular.

想像一下，要是我們在某個真正大眾化的東西上也找到 product-market fit，那會是什麼光景。

And then of course it's building the next frontier model every six to nine months, which drives the whole thing forward again.

然後當然，每隔六到九個月就會蓋出下一個 frontier model（前沿模型），又把整件事再往前推一輪。

Now, there's kind of an interesting comparison I think to make here, which is to go back and look at what happened when mobile data started working in about 2010.

這裡我覺得有個很有意思的對照，就是回頭看看大約 2010 年 mobile data（行動數據）開始真正派上用場時，發生了什麼事。

Which is that the operators had given people flat-rate data on the presumption of one kind of usage.

當時電信業者是基於「人們會這樣用」的假設，給了大家 flat-rate（吃到飽）的數據方案。

And then people got iPhones, and the iPhones got 3G, and then they started watching video.

然後人們買了 iPhone，iPhone 又支援了 3G，接著大家開始用它看影片。

And so the network started collapsing because people were using it far more.

於是網路開始撐不住，因為大家用的量遠遠超出預期。

And that's more or less exactly what's been going on at OpenAI and Anthropic and Gemini in the last couple of months, as suddenly they have to scramble to change their pricing scheme to reflect new patterns of usage.

而這幾乎就是過去幾個月在 OpenAI、Anthropic、Gemini 身上正在上演的劇情 —— 他們突然得手忙腳亂地改定價方案，好對應新的使用模式。

Now, Sam Altman said a couple weeks ago that he thinks we're going to get to a future where intelligence is a utility, like electricity — we buy it on a meter, everyone pays for it by the day.

幾週前 Sam Altman 說，他認為我們會走向一個未來，智慧會變成像電力一樣的 utility（公用事業）—— 我們用電表計費，每個人按日付費使用。

The challenge for that model should become apparent if you go back again and look at what happened in mobile, where mobile data has grown 1,500 to 2,000 times in the last ten or fifteen years.

這套模式的挑戰在哪，只要你再回頭看看行動通訊的歷史就會一清二楚 —— 過去十到十五年，mobile data 成長了 1,500 到 2,000 倍。

The stocks have gone nowhere; there's a trillion-dollar industry that doesn't make a very good return, and all the value capture and all the cool stuff is built somewhere else.

但這些電信股股價原地踏步；那是個上兆美元的產業，報酬卻很難看，而所有的 value capture（價值攫取）、所有酷東西，全都是別人在別處打造出來的。

And so there's a challenge selling a commodity and expecting to capture value from it, and it does look at the moment like models appear to be commodities.

所以「賣一個 commodity（大宗商品），又指望從中攫取價值」這件事本身就很難 —— 而目前看起來，model 確實長得很像 commodity。

Now, this is kind of a challenging statement, because you use one and you use the other, and they change all the time, and you think this one's better than that one.

這是個會引起爭議的說法，因為你用過這家、又用過那家，它們又一直在變，你會覺得這家比那家好。

But there isn't any sort of fundamental reason why any of these companies is going to have a sustainable lead over the other.

但其實沒有任何根本性的理由，讓這些公司裡的任何一家能對其他家保持「可持續」的領先。

They're all built with the same data, with the same technologies, with the same chips, the same infrastructure.

他們全都用同樣的資料、同樣的技術、同樣的晶片、同樣的基礎設施在打造。

There's no network effect. There's no fundamental path by which one model can be completely different from all of the others.

這裡沒有 network effect（網路效應）。也沒有任何根本性的途徑，能讓某一個 model 變得跟其他所有 model 完全不同。

And so that gets us to a question of what this equilibrium is going to look like when everything settles down.

這就把我們帶到一個問題：等一切塵埃落定後，這個均衡狀態會長什麼樣子。

There's a software industry historically that was capital-light. It didn't have network — but it had network effects.

歷史上的軟體業是 capital-light（輕資本）的。它沒有實體網路 —— 但它有 network effects。

And so that means you've got monopolies, and then you've got a high-margin business.

所以它能養出壟斷局面，也就成了高毛利的生意。

Microsoft used to sell you a $1 CD in a $10 box for $200.

Microsoft 以前可以把一片成本 1 美元的 CD、裝進一個 10 美元的盒子，賣你 200 美元。

But large language models are capital-intensive, they don't appear to have network effects, they do appear to be commodities.

但大型語言模型（LLM）是 capital-intensive 的，看起來沒有 network effects，反而很像 commodity。

And businesses that are commodities without network effects tend not to have high margins.

而那種「是 commodity、又沒有 network effects」的生意，通常都做不出高毛利。

So there's a kind of primary question here: do LLMs have pricing power?

所以這裡有個最根本的問題：LLM 到底有沒有 pricing power（定價權）？

Can they capture value and use cases further up the stack, or are they just infrastructure?

它們有辦法在 stack（技術堆疊）更上層去攫取價值、抓住應用場景嗎，還是說它們就只是 infrastructure（基礎設施）？

And I think the kind of provisional thesis that I would propose at the moment is to put a couple of building blocks on the table and see where they lead you.

我目前會提出的暫定論點，是先把幾塊基本的積木擺到桌上，再看看它們會把你帶往哪裡。

Which is that, first of all, chat seems to be a really bad user interface.

第一塊積木是：chat（對話介面）看起來是個很糟糕的使用者介面。

It works very well for certain kinds of use cases, but most people, for most things that they do, are going to use something that looks more like an app, that helps them work out what to do.

它在某些使用場景下表現得很好，但大多數人在做大多數事情時，會用的是某種比較像 app、能幫他們搞清楚下一步要做什麼的東西。

Labs clearly aren't going to be able to build all of those. They don't seem to have network effects. They seem to be commodities.

那些 AI 實驗室（labs）顯然不可能把所有這類 app 都自己做出來。它們看起來沒有 network effects，看起來就是 commodity。

And so the implication, therefore, is that the models themselves will be infrastructure — potentially very large, very expensive infrastructure, but still infrastructure.

所以由此推論，model 本身會變成 infrastructure —— 也許是非常龐大、非常昂貴的 infrastructure，但終究還是 infrastructure。

And the innovation moves up the stack into deployment, into operational change in other places.

而真正的創新，會往 stack 上層移動 —— 移進 deployment、移進其他地方的營運變革裡。

And so I'll talk a little bit about what that might mean, how that might work.

接下來我就稍微談談這會是什麼意思、又會怎麼運作。

Firstly, thinking about deployment, one answer might be: well, hey, everyone's already using this.

首先，講到 deployment，一種答案可能是：欸，大家不都已經在用了嗎。

OpenAI has 900 million weekly active users — although they haven't actually given a number since the end of last year, because they're embarrassed by it.

OpenAI 有 9 億 weekly active users（週活躍用戶）—— 雖然他們從去年底以後就再也沒公布過數字，因為他們對這個數字有點難為情。

Of course, that's because they have no competition now.

當然，那是因為他們現在沒什麼競爭對手。

But if you dig into those numbers, what you see is that usage is kind of a mile wide and an inch deep.

但如果你細看那些數字，會發現使用情況是「一哩寬、一吋深」—— 鋪得很廣，扎得很淺。

So even in really kind of core early adopter demographics, at least as many people are only using this every couple of days, or every couple of weeks, as are using it every day.

所以就算在最核心的早期採用者族群裡，「每隔幾天、甚至每隔幾週才用一次」的人，至少跟「每天用」的人一樣多。

And when you look into adults in the workplace, there are far more people saying, well, it's very cool, I might use it next week.

而當你去看職場上的成年人，更多人講的是：嗯，這很酷，我下週也許會用用看。

So people have not made this a daily habit.

所以大家還沒把它變成每天的習慣。

You see that even more if you dig into individual industries; even in the tech industry, there are more people who are using this once a month or once a week than are using this every day.

你要是再切進個別產業看，這個現象更明顯；就連在科技業，「一個月用一次、一週用一次」的人，都還比「每天用」的人多。

And again, this is just kind of a very glass-half-full, glass-half-empty kind of chart.

而這又是一張很「半杯水」的圖表 —— 看你覺得是半滿還是半空。

Because on the one hand, this is great growth. On the other hand, what kind of growth is it?

因為一方面，這是很漂亮的成長；但另一方面，這到底是哪一種成長？

And I think a kind of a useful comparison to think about here is to go back and look at another transformative technology, which is a spreadsheet.

我覺得這裡有個很好的對照，就是回頭看另一項曾經帶來巨變的技術：試算表（spreadsheet）。

So this is an Apple II running VisiCalc, which is the first successful software spreadsheet.

這是一台 Apple II 跑著 VisiCalc，那是史上第一套成功的試算表軟體。

With the computer, the monitor and the printer and the disk drives, the cost, I think, was about $12,000 or $13,000, adjusted for inflation.

連同主機、螢幕、印表機和磁碟機，整套的價格，換算成今天的幣值，我想大概是 12,000 到 13,000 美元。

And if you were an accountant and you saw this, it was life-changing. You had to have this, because this thing could do a week of work in half an hour.

如果你是會計師，看到這玩意兒，簡直是人生翻轉 —— 你非買不可，因為它能把一週的工作，在半小時內做完。

But if you were a lawyer and you looked at this, you would say, well, that's very cool, and my accountant should see this, but that's not what I do all day. I don't do spreadsheets.

但如果你是律師，看著它，你會說：嗯，這很酷，我的會計師該來看看，但這不是我整天在做的事啊。我又不弄試算表。

Give me a word processor, and I'll buy that. But as a spreadsheet, I don't know what I would do with this.

給我一套文書處理軟體，那我會買。但一套試算表，我實在不知道我能拿它來幹嘛。

And that, I think, is where most of those weekly and monthly active users are when they look at Claude or ChatGPT or Gemini.

而我認為，大多數那些週活躍、月活躍用戶，看著 Claude、ChatGPT 或 Gemini 時，心境差不多就是這樣。

They say, well, that's very cool. I'll think of something to do with it tomorrow. I used it for something last week.

他們會說：嗯，這很酷。我明天再想想能拿它來做什麼吧。我上週好像用它做過某件事。

But there's a struggle to get deeper use. There's a struggle to actually change everybody's lives and the way that people work.

但要走到更深度的使用，是很吃力的；要真正改變每個人的生活、改變大家工作的方式，是很吃力的。

And there's a kind of a common pattern here for how this always tends to work.

這裡頭有一個很常見的模式，每次大概都是這樣運作的。

Which is, to begin with, you do the old thing more. You use this for the old tasks. You make the new tools fit the old tasks.

也就是：一開始，你只是「把舊的事情做得更多」。你拿這個新工具去做老任務，硬把新工具套進舊任務裡。

And it takes longer to work out things that are actually only possible with the new technology. It takes longer to build new things, and then, of course, maybe to disrupt.

而要搞清楚「哪些事是只有靠新技術才做得到的」，需要更長的時間；要打造全新的東西、進而顛覆現狀，也需要更長的時間。

That's what you see if you look at the data, if you ask the CFOs, the CEOs.

你去看數據、去問那些 CFO、CEO，看到的就是這個現象。

What they say is, well, the stuff that's working so far is the stuff that's kind of easy to deploy, and it's easy to work out how you would use this.

他們說：到目前為止，真正有效的，都是那種比較容易 deploy、也比較容易想清楚「該怎麼用」的東西。

But it's harder to find that on the P&L. It's harder to find better productivity or better data analytics in the EPS numbers.

但這些東西很難反映在損益表（P&L）上。你很難在 EPS 數字裡，看到生產力變好或數據分析變強的成果。

And the stuff that's easier to measure, that shows up in EPS, is the stuff that's harder to deploy and take on. Because, of course, they're harder to take on.

而那些比較容易量化、會反映在 EPS 上的東西，反而是比較難 deploy、比較難導入的 —— 因為它們本來就比較難啃。

If you push back again to look at industries, then what are those places where this is easy to do work?

如果你再退一步、從產業的角度看，那哪些地方是「比較容易上手做出成果」的呢？

Well, analytics, productivity, marketing, customer support, or kind of automation of point solutions and specific problems inside many back offices.

嗯，數據分析、生產力工具、行銷、客服，或是在許多後台部門裡，針對單點方案和特定問題做自動化。

And then, of course, software development, which is the thing that has everyone in the Valley running around with their hair on fire this year.

然後當然還有軟體開發 —— 這就是今年讓矽谷每個人都急得像著了火一樣團團轉的東西。

This is kind of a representative quote from Mark Zuckerberg, and this is what pretty much everybody in the Valley will tell you now.

這是一段很有代表性的 Mark Zuckerberg 的話，而這也幾乎是現在矽谷每個人都會跟你說的。

That stuff that used to take a bunch of people a bunch of time is now many fewer people, much quicker.

那些以前要一大群人花一大堆時間才能搞定的事，現在用少很多的人、快很多的速度就能完成。

As I said, this has people talking about some kind of generational change in computing.

就像我說的，這讓大家開始談論運算領域的某種「世代級變革」。

I think it's kind of worth pointing out that we've had generational changes in computing before.

我覺得值得提醒一下，運算領域以前就出現過好幾次世代級的變革。

These are two ads from the mid-1980s that say no one will have to write code anymore — this will write your software for you.

這是兩則 1980 年代中期的廣告，上面寫著：以後再也不用有人寫程式了 —— 這個東西會幫你把軟體寫出來。

And the funny thing is, it kind of happened. We've had successive waves of things that make it easier to make software.

而好笑的是，這件事還真的多少發生了。我們經歷過一波又一波讓「做軟體」變得更容易的技術。

And so if you look at what the software install base looks like today, a typical big company's software can be put into three categories.

所以如果你看今天一家公司裡的軟體安裝基礎，一家典型大公司的軟體大致可以分成三類。

There are the horizontal big-iron systems — ERP, HCM, Salesforce, SAP, Oracle, etc., Workday.

第一類是橫向的「大鐵塊」系統 —— ERP、HCM、Salesforce、SAP、Oracle 等等，還有 Workday。

And then there are hundreds and hundreds of vertical apps.

第二類是數以百計的垂直應用（vertical apps）。

A typical big US corporation today uses four to five hundred SaaS applications, and then a thousand or many thousands of individual apps that they built or bought or deployed over the last 20 years.

今天一家典型的美國大企業會用上 400 到 500 個 SaaS 應用，再加上過去 20 年間自建、外購或部署的上千、甚至數千個各別應用。

And then you've got this kind of intermediate space of Excel and email and shared folders and network drives.

然後，還有一塊介於中間的灰色地帶 —— Excel、email、共用資料夾、網路磁碟機。

And you ask yourself, well, what's the right way of solving this problem?

你會問自己：那要解決這個問題，正確的做法到底是什麼？

And the answer is that these tasks are always moving back and forth between these.

而答案是：這些任務，總是在這幾類之間來來回回地搬動。

A couple of years ago, I spoke to a consultant. He said, part of his jobs were moving people from Excel to a database, and the other half were the other way around.

幾年前我跟一位顧問聊過。他說，他工作的一半，是幫客戶從 Excel 搬到資料庫；另一半，則是反過來，從資料庫搬回 Excel。

And now we layer AI into this, and it moves all of those things and accelerates them some more.

而現在我們把 AI 疊進這一切，它又把所有這些東西重新搬動了一遍，還讓它們搬得更快。

So, again, there is a building block to think about here, that clearly this makes software much cheaper and much faster.

所以這裡又有一塊值得思考的積木：很明顯，AI 讓做軟體變得便宜得多、也快得多。

It can do new stuff, and that produces new margin structures, new competitive threats.

它能做出新的東西，而這會帶來新的毛利結構、新的競爭威脅。

Some companies will go out of business, some use cases get unbundled, there'll be a lot of churn.

有些公司會倒掉，有些使用場景會被 unbundle（拆解開來），整個市場會出現大量的洗牌。

But the more interesting question is how far this replaces software, and how much it makes new software, or different kinds of software doing different things.

但更有意思的問題是：這東西會把現有軟體取代到什麼程度，又會催生出多少新軟體、或是去做不同事情的不同軟體。

Here we have Steve Jobs kind of pointing out the obvious, which is that most people are tool builders — most people can't just take a piece of technology and work out what to do with it.

這裡有段 Steve Jobs 的話，講的是個顯而易見的道理：大多數人是「工具的打造者」—— 大多數人沒辦法光拿到一項技術，就自己想清楚該拿它來做什麼。

You need to create something and give it to them, and so you need to build software.

你得做出某個成品交到他們手上，所以你得去打造軟體。

You need to think about something that manages the blank screen, manages the jagged frontier, the unpredictability of what you want these things to do.

你得去設計一個東西，去馴服那片「空白螢幕」、去馴服那條「鋸齒狀的能力邊界」（jagged frontier），去馴服「你想讓這些東西做的事」那種不可預測性。

And then fundamentally you kind of need to work out: well, does this go at the top of the stack or the bottom?

然後說到底，你得想清楚一件事：這東西該擺在 stack 的最上層，還是最底層？

Should it sit at the top of your stack, looking at all of your systems and synthesizing things across many different places? Or does it provide features inside existing pieces of software?

它該坐在你 stack 的頂端，俯瞰你所有的系統、把散落在許多不同地方的東西整合起來嗎？還是說，它只是在現有的軟體裡提供一些功能？

And again, the answer to that is probably yes — all of those things.

而這個問題的答案，大概是「都對」—— 上面說的那些它都會是。

How would you work out what that looks like? Well, you can go and hire an Accenture, you can go and hire an IT systems integrator, you can also go and talk to a consultant.

那你要怎麼搞清楚它長什麼樣？嗯，你可以去請一家 Accenture，可以去請一家 IT 系統整合商，也可以去找顧問聊聊。

Or maybe your shareholders come and talk to you about how you should build this, because, guess what, completely redesigning how a big company does something is not an easy thing to do.

又或者，是你的股東跑來跟你談該怎麼建這套東西，因為你猜怎麼著 —— 把一家大公司做事的方式徹底重新設計，可不是件容易的事。

And there are professional services firms whose entire industry, whose entire business is based on that.

而確實有一整批專業服務公司，整個產業、整門生意，就是建立在這件事上頭。

You do a bunch of pilots, and again, some of those pilots work, some of them don't — which is why you call them pilots.

你先跑一堆 pilot（試點專案），而一樣的，有些 pilot 成功、有些失敗 —— 不然它幹嘛叫「試點」。

And again, those are concentrated in software, customer service and marketing.

而這些試點，一樣集中在軟體、客服和行銷這幾塊。

And then every startup in the world is now focused on trying to do this too — this is why startups exist.

然後全世界的新創公司，現在也都一頭栽進來想做這件事 —— 這本來就是新創存在的理由。

They exist to unbundle Google, or unbundle SAP, or unbundle Excel, and now unbundle ChatGPT, and turn that into a product.

它們存在的目的，就是去 unbundle Google、unbundle SAP、unbundle Excel，現在則是去 unbundle ChatGPT，把它變成一個產品。

And so we've got this coming wave of software and tools and use cases, all pushing in every direction, bundling and unbundling everything.

所以我們正迎來一波軟體、工具與應用場景的浪潮，朝四面八方推進，把一切又 bundle、又 unbundle。

But then there's a more basic question, which is: well, that's all great, but does this actually change anything for my company, or is this just loads more software?

但接著有個更根本的問題：好啦，這些都很棒，但這對我的公司到底改變了什麼？還是說，這不過就是「多了一堆軟體」而已？

I think the challenge here is that it's kind of hard to make predictions, especially about the future, especially at this stage when you're so early in the industry and everything is unsettled.

我覺得這裡的難處在於：要做預測本來就很難，尤其是預測未來，尤其是在這種「產業還很早期、一切都還沒定下來」的階段。

It's very easy to say what's going to happen in mobile in 2015. It was much more difficult to say what was going to happen in 2005, or to talk about what was going to happen to the internet in 1997.

站在 2015 年去說「行動通訊接下來會怎樣」很容易；但站在 2005 年要說、或是站在 1997 年要談「網路接下來會怎樣」，就難得多了。

Like, what would you have predicted? You've got some of it. But Google didn't exist. Mark Zuckerberg was in junior high school. Jeff Bezos sold books.

你當時會預測出什麼？有些你大概猜得到。但那時 Google 還不存在。Mark Zuckerberg 還在念國中。Jeff Bezos 還在賣書。

It wasn't clear how all of this was going to work.

當時根本看不清這一切到底會怎麼運作。

What I think we can say, though, is we can think about automation.

不過，我認為我們可以談的，是「自動化」這件事。

So every previous wave of automation let you automate anything that you could describe in logical steps.

過去每一波自動化，讓你能把「任何可以用邏輯步驟描述出來的事」自動化。

AI lets you automate anything where there's enough training data and where verification is scalable.

而 AI 讓你能把「任何有足夠 training data、而且驗證（verification）可以規模化」的事自動化。

That's not everything, but it's a bunch of important, useful things.

這當然不是「所有事」，但確實涵蓋了一堆重要又實用的事。

So if you can now make something cheaper, what happens?

那麼，如果你現在能把某件事做得更便宜，會發生什麼？

Well, that may just be pricing that system. If you make it cheaper to do something, do you do the same for less money, or do you do more for the same money, or do you do more for more money?

嗯，這可能只是個定價問題。如果某件事變便宜了，你是「花更少的錢做一樣的事」，還是「花一樣的錢做更多」，還是「花更多的錢做更多」？

But push that further: what tasks used to be expensive that have now become free?

但再往下推：有哪些以前很貴的任務，現在變成幾乎免費了？

Maybe it was that cost base, that barrier to entry, that protected you from competition.

也許，正是那個成本基礎、那道進入門檻（barrier to entry），一直在替你擋住競爭。

And what things were just completely cost-prohibitive, that nobody even thought about, were impossible, that now become cheap?

還有哪些事，以前貴到根本不可能、貴到沒人會去想，現在卻變便宜了？

The extreme case is that sometimes the job is just a task that gets automated away.

最極端的情況是：有時候，整份工作根本就只是一項任務，然後它直接被自動化掉了。

There used to be a job called an elevator attendant, because elevators didn't have buttons — they had an accelerator and a brake, and you drove them to the floor. And then that got automated away.

以前有一種職業叫「電梯操作員」，因為那時電梯沒有按鈕 —— 它有油門和煞車，你得親自把它開到指定樓層。然後這份工作就被自動化掉了。

More often, though, the job changes.

但更常見的情況是：工作的內容會改變。

So we spent the whole of the 20th century automating accounting, and the result was more accountants.

舉例來說，我們花了整個 20 世紀在把會計自動化，結果呢？會計師反而變多了。

Partly because that unlocked demand for more work, but also because the job changed: the thing that used to take a week took 30 seconds, so you did other stuff that you couldn't have done before.

一部分是因為這釋放出了「更多工作」的需求，但也因為工作本身變了 —— 以前要花一週的事，現在 30 秒就搞定，於是你開始去做以前根本做不到的其他事。

And sometimes when you automate something, everything else changes around it in unpredictable ways.

而有時候，當你把某件事自動化之後，周圍的一切都會以你意想不到的方式跟著改變。

This is to talk about barcodes. So in the mid-1970s, the US grocery industry deployed barcodes and UPCs, and they did it to save money at the checkout.

這裡要談的是條碼。1970 年代中期，美國零售雜貨業導入了條碼和 UPC（通用產品代碼），當初是為了在結帳這一關省錢。

But the result was that now they had an accurate record of their inventory, and so that meant you could carry less inventory and carry way more products and make more money — or cut your prices, in the case of Walmart.

但結果是，他們現在握有精確的庫存紀錄，這代表你可以備更少的庫存、卻陳列多得多的品項、賺更多的錢 —— 或者像 Walmart 那樣，乾脆降價。

And so just automating that simple process changes the structure of the industry.

所以光是把那麼一個簡單的流程自動化，就改變了整個產業的結構。

I think it's useful to think about what that did to a bunch of industries — or what the internet did to a bunch of industries here — where you had industries that had some kind of physical thing that wasn't actually the business.

我覺得，去想想這對一堆產業造成了什麼影響很有用 —— 或者說，網路對這裡一堆產業造成了什麼影響 —— 在那些產業裡，原本都有某種「實體的東西」，但那個實體其實並不是生意本身。

So newspapers didn't think of themselves as printing companies or light industrial companies. Record companies didn't think of themselves as making little pieces of plastic — but that was the business.

比方說，報社不會把自己當成印刷公司或輕工業公司；唱片公司也不會覺得自己是在「做一片片小塑膠」—— 但那其實正是它們的生意。

And the internet broke those apart, and that moved the barrier to entry.

而網路把這些拆了開來，連帶把進入門檻也挪了位置。

So we should ask the same question now. If we're going to be able to automate anything boring, what industries need that as a barrier to entry?

所以我們現在該問同樣的問題：如果我們將能把「任何無聊的事」自動化，那有哪些產業，正是靠「這件無聊事」當作進入門檻在撐？

What industries are based on that? What can be split apart and moved into different industries?

有哪些產業是建立在這件事上頭的？有哪些東西可以被拆開、然後挪進不同的產業裡？

Meanwhile, again, pushing further: what is it exactly that we can automate with this stuff? I always used to say that AI gave you infinite interns.

與此同時，再往下推一層：用這玩意兒，我們到底能自動化什麼？我以前常說，AI 給了你「無限多的實習生」。

But what kind of intern? So — and this is my question — what was impossible that now becomes cheap?

但是哪一種實習生？所以 —— 這是我的問題 —— 有什麼以前根本不可能、現在卻變便宜了？

It was impossible to listen to every phone call and tell me if the customer was angry. That's one kind of automation.

以前要「聽完每一通電話、然後告訴我這位客戶是不是在生氣」是不可能的。這是一種自動化。

But another kind is to say: listen to every single call, and then you'll notice stuff that no person could ever have noticed before. It's a very different kind of automation.

但另一種自動化是說：聽完每一通電話，然後你會注意到一些「過去沒有任何人能注意到的東西」。這是非常不一樣的一種自動化。

A practical application of that kind of abstraction is to think about e-commerce and advertising.

這種抽象概念的一個實際應用，就是去想想電子商務和廣告。

I always think that Google and Amazon and Meta don't really know what they sell.

我一直覺得，Google、Amazon 和 Meta 其實並不真的知道自己在賣什麼。

They know it's a SKU, they know the words you've typed in around it, but they don't know what the products are, and they don't know why anybody buys them.

他們知道那是個 SKU（庫存單位），知道你在它周邊打了哪些字，但他們不知道那些產品到底是什麼，也不知道人們為什麼要買它。

Here's Mark Zuckerberg again. They know that people who bought X bought Y, but they don't know what X and Y are, and they don't know why people buy them.

這裡又是 Mark Zuckerberg。他們知道「買了 X 的人也買了 Y」，但他們不知道 X 和 Y 究竟是什麼，也不知道人們為什麼要買。

And now, rather than be able to understand what that thing is, and why you cared about it, and what else people might buy, we'll at least understand it at a different level of statistical correlation.

而現在，雖然還談不上真正搞懂「那東西是什麼、你為什麼在意它、你還可能會買什麼」，但我們至少能在另一個層次的統計相關性上去理解它。

Which gets into different kinds of questions.

這就把我們帶進了另一類的問題。

You go from "here's a picture of a coat, what is it?" to "look at my Instagram and suggest some different coats to buy."

你從「這是一張外套的照片，它是什麼？」，進化到「看看我的 Instagram，然後推薦幾件不同的外套讓我買」。

Those aren't SQL queries. Those aren't things that you could have asked Amazon.

這些都不是 SQL 查詢。這些都不是你以前能拿去問 Amazon 的東西。

Now, I want to have a sort of brief reality check and point out that people in the Valley say they don't use Google anymore — but actually that's not helping more people shop.

這裡我想做個簡短的現實檢查，指出一點：矽谷的人都說他們不再用 Google 了 —— 但這其實並沒有幫到更多人去購物。

But, again, you can kind of ask: well, what are the new kinds of questions? The same with the enterprise.

但一樣的，你可以問：那麼，新類型的問題是什麼？放到企業端，道理也一樣。

You could be all simple and all paying attention and all finding that data.

你大可以走那種簡單路線，事事盯著、把該找的資料都找出來。

But then you can say: listen to every Zoom call with every customer, and then look at our telemetry, and our customer analytics, and our churn data, and then tell me how we can change our pricing or our product to improve our churn.

但接著你也可以說：去聽每一通跟客戶的 Zoom 通話，再對照我們的遙測數據、客戶分析、流失（churn）資料，然後告訴我，我們該怎麼調整定價或產品，才能改善客戶流失率。

Those are questions you couldn't ask before.

這些，都是你以前問不出口的問題。

So this gets me back to this question of how you split things apart. What's the hard part?

這又把我帶回那個「該怎麼把事情拆解開來」的問題。難的部分到底是什麼？

Is the hard part writing the code, or making the spreadsheet, or delivering the SKU, or is the hard part something else?

難的部分，是寫程式碼嗎？是做試算表嗎？是把那個 SKU 送到嗎？還是說，真正難的是別的東西？

The hard part is knowing what you were supposed to be doing — knowing which task, and why, and what else you should do instead.

真正難的，是「搞清楚你本來該做的是什麼」—— 搞清楚該做哪一項任務、為什麼要做、以及你是不是該改去做別的。

That definitely gets me to some concluding thoughts.

這就帶我走到一些收尾的想法了。

Firstly, to go back to this diagram, we should remember that all of these things that happened in the past were really fucking amazing too.

首先，回到這張圖 —— 我們得記住，過去發生的這所有事情，當年也都他媽的同樣驚人。

Yes, AI changes everything and nothing like this has ever happened before — but that was true of all of those other things as well.

沒錯，AI 改變了一切、而且這種事前所未有 —— 但對過去那些東西來說，當年講的也是同一句話。

We've also got some things that are amazing and have never happened before as we look at market valuations at the moment.

而當我們看當下的市場估值，也確實有些「驚人、前所未見」的東西。

Where the two big model labs — never mind SpaceX, of which the least said the better — are each of them bigger than all tech IPOs in the last five years of the dot-com bubble.

那兩家大型 model 實驗室 —— 先別提 SpaceX，那個越少講越好 —— 它們的估值，各自都比 dot-com 泡沫最後五年所有科技 IPO 加起來還要大。

But if we go back in history again, how many times have we talked about automation?

但如果我們再回頭看歷史，我們已經談過多少次「自動化」了？

This is an IBM ad from the 50s that says that an electronic calculator will give you 150 extra engineers — which is basically every enterprise software pitch today.

這是一則 1950 年代的 IBM 廣告，上面說一台電子計算機，等於替你多請了 150 名工程師 —— 這基本上就是今天每一份企業軟體的推銷說詞。

And indeed, if you go back and look through history, we've had successive waves of people getting very excited about automation.

而事實上，如果你回頭翻翻歷史，會發現人們已經一波又一波地為「自動化」興奮過好幾回。

We've had new waves of technology, and now we have another wave of excitement about automation today.

我們經歷過一波又一波的新技術，而今天，我們又迎來了另一波對自動化的興奮。

When you go through those waves, you don't really know what's going to happen.

當你身處在這些浪潮裡，你其實並不知道接下來會發生什麼。

It seems to me that every question I get about AI has one of two answers. The answer is either "no one knows," or "what happened the last time that everything changed?"

我覺得，每一個關於 AI 的問題拋給我，答案大概都是兩種之一：要嘛是「沒有人知道」，要嘛是「上一次一切大變的時候，發生了什麼？」

And of course, the last time that everything changed, no one knew either.

而當然，上一次一切大變的時候，也同樣沒有人知道。

We didn't really know how the internet was going to work, or how mobile was going to work.

我們當年根本不真的知道網路會怎麼運作，也不知道行動通訊會怎麼運作。

There were all sorts of technologies and companies floating around that people were very excited about that didn't work.

那時有各式各樣的技術和公司滿天飛，大家都為之興奮 —— 結果最後沒成。

We should just presume that a bunch of stuff going around today will be something that doesn't work.

我們大可假設：今天滿天飛的這一堆東西裡，必然有一批最後是行不通的。

Equally, when it does work, that doesn't necessarily change everything.

同樣地，就算某樣東西真的成了，它也未必就會改變一切。

So ride-shares completely changed what we mean when we say "taxis," but the impact of Airbnb was much smaller.

比方說，共乘（ride-share）徹底改變了我們講「計程車」時的定義；但 Airbnb 帶來的衝擊就小得多。

That was just kind of an additive shift around the edges of the market.

那比較像是發生在市場邊緣的、一種「加法式」的微調。

So the impact of this stuff really kind of does depend on the market itself.

所以這玩意兒的衝擊有多大，其實真的很看市場本身是什麼樣子。

And then one final observation I will make.

最後，我要再提出一個觀察。

This is a chart of recorded music revenue showing the impact first of downloads and then streaming.

這是一張錄製音樂營收的圖表，顯示了先是下載、接著是串流所帶來的衝擊。

And it seems to me that there's two things happening in this chart.

而在我看來，這張圖裡其實發生了兩件事。

The first half of the chart is people saying, well, what happens if I don't have to pay $50 to get that track?

圖的前半段，是人們在問：嗯，如果我不必花 50 美元就能拿到那首歌，會怎樣？

The second half is: what happens if $50 gets me all the music that there is in the world?

後半段則是：如果 50 美元能讓我擁有全世界所有的音樂，會怎樣？

That's a product, an idea, that would have been completely impossible before the internet. No one even thought about doing that.

那是一個產品、一個構想，在網路出現之前根本完全不可能。當年根本沒人想過要這麼做。

And then the internet means that you can change the question.

而網路的意義就在於：它讓你能夠「換一個問題來問」。

And so then just some final things to say here. We don't actually know how any of this is going to change work — you never do when you're at this stage in the market.

所以這裡就講幾句最後的話。我們其實並不知道這一切會怎麼改變「工作」這件事 —— 當你身處市場的這個階段時，你從來就不會知道。

But you can ask, well, what do we build with this right now?

但你可以問：好，那我們現在能用這東西打造出什麼？

And you can ask, does this break some fundamental assumption in the way we think about our business?

你也可以問：這東西，會不會打破我們看待自家生意時，某個最根本的假設？

And then you can just presume that all of this is going to change out of all recognition in the next two and three and four years — just the way mobile did, just the way the internet did.

然後你大可假設：這一切，在接下來的兩、三、四年內，都會變得完全認不出原本的樣子 —— 就像當年行動通訊那樣，就像當年網路那樣。

And with that, I'll say thank you.

就講到這裡，謝謝大家。

All right. Thank you so much, Mr. Benedict.

好的。非常感謝你，Benedict 先生。

【主持人 / 開場】Everyone in this room knows that AI is not a single-country effort — this is a global effort.

在座各位都知道，AI 不是任何單一國家的事——這是一場全球性的努力。

At SuperAI we are driving the connection between East, West, and South; all corners of the globe are coming together to build the future.

在 SuperAI，我們正串起東方、西方與南方；全球各個角落正攜手打造未來。

Our panel today represents that, with representatives from China, India, and Europe — France specifically.

今天這場 panel 正體現了這點，與談者分別來自中國、印度與歐洲——具體來說是法國。

Our moderator is Zixuan Li, head of Z.ai.

我們的主持人是 Z.ai 負責人 Zixuan Li。

He is joined by Geoff Soon, VP of Revenue, APAC for Mistral.

與談的還有 Mistral 亞太區營收副總裁 Geoff Soon。

By Cherie Shi, Global Business Manager with MiniMax, straight out of Shanghai.

來自上海的 MiniMax 全球業務經理 Cherie Shi。

And from India, Hemant Mohapatra, a partner with Lightspeed India Partners.

以及來自印度、Lightspeed India Partners 合夥人 Hemant Mohapatra。

【Zixuan Li，Z.ai】Hi everyone, glad to be the moderator for this panel.

大家好，很高興擔任這場 panel 的主持人。

I was about to be one of the panelists, but I can't give public opinions due to some regulations.

我本來要當與談者之一，但因為一些法規限制，不便公開發表個人意見。

I'm Zixuan Li from Z.ai, the creator of the GLM models — GLM stands for our models and products.

我是 Z.ai 的 Zixuan Li，GLM 模型的開發團隊——GLM 就是我們的模型與產品。

【Geoff Soon，Mistral】Good morning, my name is Geoff Soon, based in Singapore, and I lead the APAC business for Mistral.

早安，我是 Geoff Soon，駐新加坡，負責 Mistral 的亞太業務。

We're a frontier AI company; I spend most of my time working with governments, banks, and large enterprises, understanding how AI can make a transformational shift in how they operate.

我們是一家前沿 AI 公司；我多數時間在與政府、銀行與大型企業合作，理解 AI 如何能徹底改變他們的營運方式。

【Cherie Shi，MiniMax】Good morning, I'm Cherie from MiniMax. I'm a founding member and now a manager on the global business.

早安，我是 MiniMax 的 Cherie，創始成員，現在負責全球業務。

We work not only on LLMs but also multimodal models — video generation, speech generation, music generation.

我們不只做 LLM，也做多模態模型——影片生成、語音生成、音樂生成。

MiniMax is a very globalized model company; over 70% of our revenue actually comes from global markets.

MiniMax 是一家非常國際化的模型公司；我們超過 70% 的營收其實來自海外市場。

【Hemant Mohapatra，Lightspeed India】I'm here from the investment side. We're a global venture capital firm, very active in AI across the world.

我代表投資端。我們是全球性的創投，在世界各地非常活躍地投資 AI。

We invest across the foundation-model layer worldwide, all the way down the chip layer up to the application layer.

我們的投資橫跨全球的基礎模型層，從晶片層一路到應用層。

【Zixuan Li】Today we're talking about frontier models. A new model comes out maybe yesterday or today — that's one definition of frontier, but there might be different versions. What does "frontier" actually mean for you and your models?

今天我們談前沿模型。可能昨天或今天就有新模型發表——那是 frontier 的一種定義，但可能有不同版本。對你和你們的模型來說，「frontier」到底是什麼意思？

【Geoff Soon，Mistral】A lot has changed over the last 12 months. The traditional idea of a frontier benchmark is topping some leaderboard — but that's extremely expensive, time-consuming, and the leaders always change.

過去 12 個月變化很大。傳統的「前沿 benchmark」是登上某個排行榜——但那既昂貴又耗時，而且領先者一直在換。

What really matters today is frontier deployment: what can you effectively deploy in your government or enterprise to drive significant change?

今天真正重要的是「前沿部署」：你能在政府或企業裡實際部署多少能力、帶來重大改變？

A big part of that is: how much of this capability can I own, how much can I control, and how durable is it in a geopolitical world?

其中關鍵是：這份能力我能擁有多少、控制多少、在地緣政治的世界裡又有多持久？

You may not always have access — there can be shocks to the supply chain for chips, to API access, to capability itself.

你不一定永遠有取得管道——晶片供應鏈、API 存取、甚至能力本身都可能遭遇衝擊。

So I look at frontier much more in terms of deployment capability and all the key ingredients needed to achieve business transformation.

所以我更傾向以「部署能力」以及達成企業轉型所需的所有關鍵要素，來看待 frontier。

【Cherie Shi，MiniMax】For MiniMax, frontier isn't only defined by benchmark performance. The definition is more like: how many people in the real world actually use your frontier models every day?

對 MiniMax 而言，frontier 不只由 benchmark 表現定義。定義更像是：現實世界裡每天有多少人真的在用你的前沿模型？

So when we optimize our models, we optimize for accessibility in the real world.

所以我們優化模型時，是針對「在真實世界的可及性」來優化。

We improve agentic capabilities in real day-to-day work, and we improve cost efficiency to become one of the most cost-efficient models in the world.

我們提升真實日常工作中的 agentic 能力，也提升成本效率，要成為世界上最具成本效益的模型之一。

The goal is that in every corner of the world, every user can use frontier-model intelligence at very low cost.

目標是讓世界每個角落、每位使用者都能以極低成本用到前沿模型的智慧。

【Hemant Mohapatra，Lightspeed】I agree with all of that, but instead of answering directly, let me zoom out and give two frameworks I've found helpful.

這些我都同意，但與其直接回答，我想拉高視角，給兩個我覺得很有用的框架。

Framework one is how these important technical super-cycles play out, and what "being at the frontier" means changes as the phases mature.

框架一：這些重要的技術超級循環如何展開，而「站在前沿」的意義會隨著階段成熟而改變。

At the beginning of a super-cycle, you're trying to extract some new natural resource. Here it happens to be intelligence.

在超級循環的開端，你是在「開採」某種新的天然資源。這次剛好是「智慧」。

A hundred years ago it was oil, fifty years ago the CPU cycle, thirty years ago the GPU cycle — today it's the AI token.

一百年前是石油，五十年前是 CPU 週期，三十年前是 GPU 週期——今天是 AI token。

In the extractive phase, if you're at the frontier you're close to the extraction; that's where the value really accrues.

在開採階段，若你站在前沿，就代表你接近開採端；價值正是在那裡累積。

As the resource becomes cheap and easily available, the cycle shifts from extractive to what I call distributive.

當資源變便宜、容易取得，循環就從「開採」轉向我稱之為「分配」的階段。

You're no longer just extracting oil — you're putting oil into machines and building better cars.

你不再只是開採石油——而是把石油放進機器、造出更好的車。

In the distributive phase, it's no longer about being the best token extractor; it's whether you can do real-world tasks agentically and solve the customer's problem.

在分配階段，重點不再是「最會開採 token」；而是你能不能 agentic 地完成真實任務、解決客戶的問題。

The second thing: there are two kinds of problems you can solve — scale problems and scope problems.

第二點：你能解的問題有兩種——規模（scale）問題與範疇（scope）問題。

A scale problem you solve once, then must serve it cheaply and abundantly to a trillion more — solving coding is a scale problem.

規模問題解一次，然後得便宜、大量地服務上兆次——解 coding 就是規模問題。

Scope problems are more complicated; the scope changes as you go deeper. Cancer research, mathematics, bio, material science are scope problems.

範疇問題複雜得多；越深入、範疇本身就會變。癌症研究、數學、生物、材料科學都是範疇問題。

You solve one type of cancer, then a new thing comes up and it stops working — you've got to keep iterating.

你解了一種癌症，新狀況又冒出來、又失效了——你得不斷迭代。

Both frameworks help think about what the frontier will look like as the cycle matures and shifts.

這兩個框架都能幫你思考：當循環成熟、轉移時，前沿會長成什麼樣子。

【Zixuan Li】Mistral is clearly the leading European AI lab. What advantages and challenges do you meet in your daily business?

Mistral 顯然是歐洲領先的 AI 實驗室。你們在日常業務中常遇到哪些優勢與挑戰？

【Geoff Soon，Mistral】There are advantages and disadvantages to being a European company.

身為一家歐洲公司，有優勢也有劣勢。

First advantage: Europe is filled with industrial heavyweights that help us solve scope problems.

第一個優勢：歐洲滿是工業重量級企業，能幫我們解 scope 問題。

We've enjoyed our partnership with ASML — it brings our capabilities to the frontier of material science and extreme-UV lithography.

我們與 ASML 的合作很愉快——讓我們的能力切入材料科學與極紫外光微影的前沿。

We just did our first-ever customer summit in Paris and announced a strategic partnership with Airbus.

我們剛在巴黎辦了首場客戶高峰會，並宣布與 Airbus 的策略合作。

These organizations have vast proprietary data that we combine with our technology to drive solutions in the scope area.

這些組織擁有龐大的專有資料，我們把它與技術結合，在 scope 領域推動解決方案。

The challenge: there's a significant difference in liquidity and access to funding compared to the U.S.

挑戰是：相較美國，在資金流動性與募資管道上有顯著落差。

Silicon Valley has done this a long time and has very efficient ways of raising large amounts of capital.

矽谷做這件事很久了，有非常高效的大額募資方式。

Also, Europe is effectively 27 different markets trying to operate as one, whereas China or America has a single regulatory framework.

此外，歐洲實際上是「27 個市場硬要當一個運作」，而中國或美國有單一的法規框架。

So you have industry giants with interesting data and problems, but the disadvantage of market fragmentation.

所以你有手握有趣資料與問題的工業巨頭，卻也有市場破碎的劣勢。

The final advantage — it's important the world knows it's not just a two-horse race between China and the U.S.

最後一個優勢——讓世界知道這不只是中美兩強的賽局，很重要。

Europe represents a very solid third alternative when choosing who to strategically partner with.

在選擇策略夥伴時，歐洲是一個非常紮實的「第三選項」。

【Zixuan Li】You have great models across text, media, and voice. What do you see as your future key differentiators?

你們在文字、媒體、語音都有很強的模型。你認為未來的關鍵差異化在哪？

【Cherie Shi，MiniMax】Two things are special about MiniMax. First, we don't only do LLMs — we have a video-generation model called Hailuo, with a new version coming soon, plus speech and music generation.

MiniMax 有兩點很特別。第一，我們不只做 LLM——還有名為 Hailuo（海螺）的影片生成模型，新版本即將推出，另有語音與音樂生成。

We started all these modality research from day one. Now we see a trend where different modalities can merge into one.

這些模態研究我們從第一天就在做。現在我們看到一個趨勢：不同模態可以融合成一個。

When you unify modalities into one model, it improves both the generation and the understanding capabilities.

當你把多模態統一進一個模型，會同時提升生成能力與理解能力。

Very few companies in the world can merge all modalities and unify understanding and generation — that broadens the use cases.

全世界很少有公司能融合所有模態、並統一理解與生成——這能拓寬使用場景。

We just released our new LLM, MiniMax M3, about a week ago.

我們大約一週前剛發布新的 LLM——MiniMax M3。

M3 isn't only an agentic model with strong coding and tool-use; it's also a great native VLM with native visual and video understanding.

M3 不只是擁有強大 coding 與工具使用能力的 agentic 模型，也是優秀的原生 VLM，具備原生的視覺與影片理解。

The second thing: we've improved cost efficiency from day one.

第二點：我們從第一天就在拼成本效率。

For M3, the price is only about one-third to one-fifth of Sonnet 4.6, yet it handles coding and agentic tasks that Sonnet 4.6 can.

以 M3 來說，價格只有 Sonnet 4.6 的約 1/3 到 1/5，卻能處理 Sonnet 4.6 能做的 coding 與 agentic 任務。

The same goes for our video and music generation — we want every user in the world to use these without worrying about ROI.

影片與音樂生成也一樣——我們希望世界上每位使用者都能用，且不必擔心 ROI。

Many AI-native users and large enterprises are passionate about using models but aren't yet seriously discussing ROI internally; that will come.

許多 AI 原生使用者與大型企業很熱衷用模型，但內部還沒認真討論 ROI；這遲早會來。

When they do, we want the ROI to still work even at huge token consumption — that's why we keep optimizing cost.

等他們開始算，我們希望即使大量消耗 token、ROI 仍然成立——這就是我們持續優化成本的原因。

【Zixuan Li】Now we have Chinese labs, European labs, and you mentioned xAI. How do you evaluate these AI labs and the global opportunities?

現在有中國的、歐洲的實驗室，你也提到 xAI。你怎麼評估這些 AI 實驗室與全球的機會？

【Hemant Mohapatra，Lightspeed】It's really hard to get conviction if you don't know what you're looking for.

如果你不清楚自己在找什麼，要建立投資信念非常難。

Two years ago this was one of the most competitive categories investors had seen in 50 years.

兩年前，這是投資人五十年來見過最競爭的領域之一。

Back then what mattered most was who had the best benchmarks — the idea of "a genius in a bottle" you could ship anywhere.

那時最重要的是誰的 benchmark 最好——彷彿「瓶中天才」，你能把它送去任何地方。

My view of AGI is a model with an IQ of about 98 — it makes human-like mistakes but has human-like intuition.

我對 AGI 的看法是一個 IQ 約 98 的模型——會犯類似人類的錯，卻也有類似人類的直覺。

Then you train that 98-IQ model on math to make it 150 on math, on bio to make it 150 on bio, on agentic tasks to make it the best agentic model.

然後你拿這個 98 分的模型去訓練數學、讓它數學到 150，訓練生物到 150，訓練 agentic 任務、成為最強的 agentic 模型。

The idea was the most intelligent token output for the least dollars per watt — and you proved it with benchmarks.

核心想法是「每瓦花最少的錢、產出最聰明的 token」——而你用 benchmark 來證明這點。

So most companies' entire brand strategy was to beat the benchmarks — and of course it could be gamed.

所以多數公司的整套品牌策略就是刷贏 benchmark——當然這是可以作弊的。

Today, that layer is going to get commoditized.

今天，這一層正要被商品化。

Look at yesterday's release: it's not that models aren't the strongest, but the gap between generations is shrinking.

看看昨天的發表：不是模型不夠強，而是世代之間的差距正在縮小。

You're reaching the asymptote of intelligence — beyond a point, every 1% improvement costs $10 billion, then $100 billion.

你正逼近智慧的漸近線——超過某個點，每提升 1% 要花 100 億美元，然後是 1000 億。

Commodity layers compete on pricing, not performance — and who can drive prices down? Those with the cap table (OpenAI, xAI, Anthropic) or the balance sheet (Meta, Google, Microsoft).

商品層比的是價格而非效能——而誰能把價格壓下去？是有資本結構的（OpenAI、xAI、Anthropic）或有資產負債表的（Meta、Google、Microsoft）。

Those two types win the model layer at commodity scale — and you want it to be a commodity, because value moves upstream.

這兩種人會在商品規模上贏得模型層——而你會希望它變商品，因為價值會往上游移動。

If you're a new model company today, you need an edge to build into the scope problem, because the baseline model will solve scale problems for everyone.

如果你今天是新的模型公司，你需要一個切入 scope 問題的優勢，因為基線模型會替所有人解決規模問題。

So we look much more at physical-world models, bio, material science, mathematics, physics, game simulation — that's where the value is.

所以我們更看重物理世界模型、生物、材料科學、數學、物理、遊戲模擬——價值就在那裡。

We also look at the teams. Look at the 1920s: when Ford came out, his pitch was "any color you like as long as it's black" — you had to scale that business well.

我們也看團隊。看看 1920 年代：Ford 推出汽車時的賣點是「只要是黑色，什麼顏色都行」——你得把規模做好。

But when GM came, the wave moved from extractive to distributive: not "do you have a car" but "I've got a better car, a V6 vs a V4, better paint."

但 GM 出現時，浪潮從開採轉向分配：重點不是「你有沒有車」，而是「我有更好的車、V6 對 V4、更好的烤漆」。

GM's value proposition was almost planned obsolescence — every two years you make all the old styles obsolete and bring in new ones.

GM 的價值主張幾乎是「計畫性汰舊」——每兩年讓所有舊款式過時、推出新款。

He was far more go-to-market and brand-centric than R&D-centric. So we ask: where are we in the cycle, and what teams and playbooks win it?

他遠比「研發導向」更「市場與品牌導向」。所以我們問：我們在循環的哪個位置、什麼團隊與打法會贏？

【Zixuan Li】In the AI world, maybe Anthropic and OpenAI sell Rolls-Royces while we sell Mercedes — most people still buy Opus and GPT. How do you see the open-vs-closed battle?

在 AI 世界裡，或許 Anthropic 和 OpenAI 賣勞斯萊斯、我們賣賓士——但多數人仍在買 Opus 和 GPT。你怎麼看開源與閉源之爭？

【Geoff Soon，Mistral】I love that analogy, but another is the Swiss Army knife vs the specialist tool.

我很喜歡那個比喻，但還有一個：瑞士刀 vs 專用工具。

If I have a rough idea and want to do a PoC, the Swiss Army knife gives me a lot of flexibility — not the most efficient, but I'll get there.

如果我有個粗略想法、想做 PoC，瑞士刀給我很大彈性——不是最有效率，但能達標。

Once I've proven the use case, I start optimizing with specialist tools — and that's where MiniMax or Mistral come in.

一旦驗證了使用情境，我就開始用專用工具優化——這正是 MiniMax 或 Mistral 的切入點。

Beyond the specialist tool, you look at the actual user and how to get enterprise context into the tool to deliver the outcome.

在專用工具之上，你要看真正的使用者，以及如何把企業情境帶進工具、交付成果。

On AGI, we see multiple domains where superintelligence is possible — it's not one general all-purpose thing.

談到 AGI，我們認為超級智慧會在多個領域出現——而非單一的通用萬能體。

In 12–18 months, superintelligence could perform many call-center functions, but it's far less likely to do advanced brain surgery — the tech just isn't there.

12 到 18 個月內，超級智慧可能勝任許多客服中心的工作，但要做高階腦部手術的可能性低得多——技術還沒到位。

So we compete by specializing that knowledge, making it more efficient, and bringing enterprise context in.

所以我們的競爭方式是：把知識專業化、讓它更高效、並把企業情境帶進去。

Simply subscribing to a powerful general API without building your enterprise context into it is just a temporary productivity bump, not a moat.

單純訂閱一個強大的通用 API、卻沒把企業情境建進去，只是暫時的生產力提升，不是護城河。

【Zixuan Li】MiniMax has both open-source and business versions. How do you assess the opportunities for open-source models?

MiniMax 同時有開源版與商業版。你怎麼評估開源模型的機會？

【Cherie Shi，MiniMax】For our LLMs — M2, M2.1, M2.5, M2.7, and now M3 — they're all open source.

我們的 LLM——M2、M2.1、M2.5、M2.7，到現在的 M3——全都是開源的。

But our video, speech, and music generation are closed-source. Closed-source models still lead and define new frontier capabilities.

但影片、語音、音樂生成是閉源的。閉源模型仍在引領並定義新的前沿能力。

We see huge demand for deploying open-source models in different regions — especially in APAC and Southeast Asia, in local data centers for local privacy use cases.

我們看到在各地部署開源模型的龐大需求——尤其在 APAC 與東南亞，部署在在地資料中心、處理在地隱私需求。

Open source greatly improves the adaptiveness to countries with strict local-deployment requirements.

開源大幅提升了對「有嚴格在地部署要求」國家的適應力。

But if a client wants to access the frontier model the fastest way, the cloud-based closed version is best.

但若客戶想以最快方式用到前沿模型，雲端閉源版是最好的選擇。

To integrate new capabilities into your AI-native product or workflow in the most timely manner, use the cloud closed version.

要最及時地把新能力整合進你的 AI 原生產品或工作流，就用雲端閉源版。

In the long run, open and closed source can coexist to serve different use cases — even the same client.

長期來看，開源與閉源能並存，服務不同的使用情境——甚至同一個客戶。

We provide all options: API platform, self-deployment versions of our LLM and TTS, and FTE engineers for technical support.

我們提供所有選項：API 平台、LLM 與 TTS 的自部署版本，還派駐工程師提供技術支援。

【Zixuan Li】Hemant, what's your take on open source?

Hemant，你對開源的看法是？

【Hemant Mohapatra，Lightspeed】Both models can coexist, but there isn't enough discussion on why there should be an open-source alternative.

兩種模式可以並存，但關於「為什麼該有開源這個選項」的討論還不夠。

There's a business-model angle and a philosophical angle.

這裡有商業模式的面向，也有哲學的面向。

Philosophically, if you want AI to do good for everybody, you want the fundamental intelligence layer to be commoditized — not controlled by a handful of private corporations selling at the highest price.

哲學上，若你希望 AI 造福每個人，你會希望最底層的智慧被商品化——而非被少數私人公司壟斷、賣最高價。

By the way, commoditization doesn't mean cheap or free — it means fungible.

順帶一提，商品化不代表便宜或免費——它代表「可替換」。

Intelligence from model one differs from model two, but oil from gas station one is no different from station two — so it competes on price vs performance.

模型一的智慧和模型二不同，但加油站一的油和加油站二沒差——所以它競爭的是價格對效能。

The fastest way to commoditize the model layer, so value moves upstream and gets distributed, is open weights / open source.

讓模型層商品化、使價值上移並被更廣分配的最快方式，就是 open weights／開源。

That means everybody controls the intelligence and can tune and shape it.

那代表每個人都能掌控智慧、都能微調與形塑它。

If you don't open it, it can still get commoditized — oil is controlled by a few and is still a commodity — but it takes longer.

就算不開源，它仍會被商品化——石油被少數人控制、卻仍是 commodity——只是會更慢。

The challenge with open source is the business model.

開源真正的難題是商業模式。

If you sell an open-weight SOTA model, you'll have to price on tokens — people want to consume intelligence on tap.

如果你賣一個開源的 SOTA 模型，你只能按 token 計價——人們想要的是「打開水龍頭」式的智慧。

You open the tap, it fills your glass, you drink, you're happy — but to make money on tokens you must own the infrastructure.

打開水龍頭、裝滿杯子、喝下去、開心——但要靠 token 賺錢，你必須擁有基礎建設。

You can't just sell water from your pond; distribution is too expensive and you don't control the tap in people's houses.

你不能只是賣自家池塘的水；分發太貴，而且你不掌握別人家裡的水龍頭。

You have to go all the way down to GPU, energy, land — everything — to make money on the token business model.

你得一路往下吃到 GPU、能源、土地——所有環節——才能在「token 商業模式」裡賺到錢。

Six months ago open weights were six months behind; today they're three months behind — and that gap will keep shrinking as models commoditize.

半年前開源落後半年；今天落後三個月——而隨著模型商品化，這個差距會持續縮小。

People will stop caring about a three-month gap when it's a tenth of the price. The question is: can you make money in open source without owning the layers below? I don't think you can.

當價格只有十分之一，人們就不在意那三個月的差距了。問題是：不擁有底層的所有環節，你能靠開源賺錢嗎？我認為不行。

【Zixuan Li】So what are your killer features — what are people actually doing with your models?

那你們的殺手級功能是什麼——大家實際上拿你們的模型在做什麼？

【Geoff Soon，Mistral】Following on Hemant's point — to monetize the technology you need to win the full stack, from the power you use to the harnesses on top. That's a solid path for open source.

延續 Hemant 的觀點——要把技術變現，你得贏下整個全端，從用電到上層的框架。對開源來說，那是一條紮實的路。

We're making advances in material science. We announced the acquisition of a company called EMI about three weeks ago.

我們在材料科學上有進展。大約三週前我們宣布收購一家叫 EMI 的公司。

EMI specializes in surrogate modeling — one of the most time-consuming steps is running a simulation for a digital twin.

EMI 專精於「代理建模（surrogate modeling）」——最耗時的步驟之一就是替數位分身跑模擬。

EMI makes a reasonable-fidelity simulation 10–100x faster, which really helps the design process.

EMI 讓「合理保真度」的模擬快上 10 到 100 倍，大幅加速設計流程。

That continues our partnership with ASML, specializing in high-tech manufacturing.

這延續了我們與 ASML 的合作，專注於高科技製造。

Second, there are robots running around our office — we're moving into VLAs and VLMs and thinking about physical AI.

第二，我們辦公室裡有機器人在跑——我們正切入 VLA 與 VLM，思考物理 AI。

This domain hasn't been solved; you have to decide how much of the stack to own. Do we want to make the robots? No. Fine-grained manipulation? Maybe not. But the control and reasoning of embodied AI — yes.

這個領域還沒被攻克；你得決定要擁有多少全端。我們要自造機器人嗎？不。精細操控？也許不。但具身 AI 的控制與推理——要。

The best way to interact with AI shouldn't be through a little chat box — humans speak, and that's the best way to communicate.

與 AI 互動最好的方式不該是一個小聊天框——人類用說的，那才是最好的溝通方式。

So we work on the voice and visual domains to create more natural interfaces — all backed by deep partnerships with companies that have enough data to create a moat.

所以我們投入語音與視覺領域、打造更自然的介面——這一切都建立在「與握有足夠資料、能形成護城河的公司」的深度合作上。

We're betting on physical AI, science AI, and on-device models.

我們押注在物理 AI、科學 AI，以及裝置端模型。

【Zixuan Li】MiniMax is more agentic — what's there in your models for agents?

MiniMax 更偏 agentic——你們的模型在 agent 上有什麼？

【Cherie Shi，MiniMax】Agentic has a very broad definition. When we say agentic, we mean agents that handle real-world tasks across productivity and daily white-collar work, plus personal agents with high EQ that accompany your daily life.

agentic 的定義很廣。我們講 agentic，是指能處理真實世界任務的 agent——橫跨生產力與日常白領工作，也包含具高 EQ、陪伴你日常生活的個人 agent。

With M3 and M3 Auto, it has great computer-use and browser-use capabilities.

M3 與 M3 Auto 具備出色的 computer-use 與 browser-use 能力。

AI-native companies with browser-use or computer-use agents are using M3 to explore new product boundaries.

擁有 browser-use 或 computer-use agent 的 AI 原生公司，正用 M3 探索產品的新邊界。

M3 can also handle video understanding, so we see many emerging video-agent products combining agent use with visual understanding.

M3 也能做影片理解，所以我們看到許多新興的 video-agent 產品，把 agent 能力與視覺理解結合。

People are also enlarging the agent's context window to handle long-horizon tasks in one window — M3 provides a one-million-token context window.

大家也在擴大 agent 的 context window，好在單一視窗內處理長程任務——M3 提供 100 萬 token 的 context window。

We also see potential in agent teams — bringing different agents into one system to collaborate, improving robustness and overall performance.

我們也看好「agent 團隊」——把不同 agent 放進一個系統協作，提升穩健度與整體表現。

【Zixuan Li】Final question: in one sentence, what is the most underestimated key success factor for the future of AGI?

最後一題：用一句話說，AGI 未來最被低估的關鍵成功因子是什麼？

【Geoff Soon，Mistral】Distribution is going to be absolutely key to how AI evolves over the next couple of years.

分發（distribution）將是未來幾年 AI 如何演化的絕對關鍵。

We've seen how a lack of GPU distribution shaped the market; it's not impossible the distribution of this capability gets heavily impacted — that would be a very difficult time.

我們已見識到 GPU 分發不足如何形塑市場；這份「能力的分發」未來幾年也可能遭嚴重衝擊——那會是非常難熬的時期。

【Cherie Shi，MiniMax】For MiniMax, we're working on distributing frontier-model intelligence to every user in every corner of the world.

對 MiniMax 來說，我們致力於把前沿模型的智慧分發到世界每個角落的每位使用者。

We keep lowering inference cost and looking for cheaper tokens, cheaper energy, and cheaper infrastructure to bring intelligence to every region.

我們持續壓低推論成本，尋找更便宜的 token、能源與基建，把智慧帶到每個地區。

【Hemant Mohapatra，Lightspeed】We are at a point where magical things are possible — what took days, weeks, or months can be done in minutes; what took 10–20 steps can be done in zero-shot.

我們正處在「神奇之事成為可能」的時刻——過去要花幾天、幾週、幾個月的事，現在幾分鐘就能完成；要 10 到 20 步的事，現在能 zero-shot 做到。

What we look for in founders is an extreme clarity about where the world is headed, and an elegant way of looking at a problem that wows us.

我們在創辦人身上找的，是對「世界走向」的極度清晰，以及一種優雅、能讓我們驚艷的看待問題的方式。

Too many founders take this amazing vial of oil and try to feed it back to the horses, hoping they'll run faster — they won't.

太多創辦人拿著這瓶神奇的石油，卻想把它餵回給馬、指望牠跑更快——牠不會。

You have to have the imagination and ambition to take this new resource and dream brilliant things that wow the world.

你得有想像力與野心，用這份新資源夢出驚艷世界的東西。

If you can't, it'll be very hard to raise capital, attract a team, or get customers to pay attention.

做不到的話，募資、找人、讓客戶注意到你，都會非常難。

We're passing on companies that go from $1M to $3M in a year — that used to be the hottest growth three or four years ago; today people get there in six months, and we still pass because others move faster with better retention, teams, and vision.

我們正在 pass 一年內從 100 萬做到 300 萬美元的公司——三、四年前那是最熱的成長，今天六個月就到，我們照樣 pass，因為總有人跑更快、留存更好、團隊與願景更強。

So dream big and have extreme clarity on where the world is headed.

所以，把夢做大，並對世界的走向保持極度清晰。

【Zixuan Li】Thanks again for the great sharing — that's all for today's panel. Looking forward to your new models. Big thanks to our panelists!

再次感謝精彩分享——今天的 panel 就到這裡。期待你們的新模型。非常感謝各位與談人！

When the first LLMs came out, everyone said, wow, this is amazing. But there was a catch.

當第一批 LLM 問世時，大家都驚呼太神奇了。但有個問題。

They weren't really up to date — they were looking at what was happening on the old web.

它們其實不夠即時——它們看的是「老網路」上發生的事。

Ever since then, great minds have turned to the question: how do we enable AI to see things in real time on the net?

從那時起，許多聰明的腦袋就投入一個問題：怎麼讓 AI 即時看見網路上的事物？

He has been with Bright Data for 12 years and risen to become their Chief Product Officer.

他已在 Bright Data 服務 12 年，一路晉升為產品長。

Before Bright Data he worked in almost every part of technology — semiconductors, wireless, IoT, gaming, online advertising.

在 Bright Data 之前，他幾乎做過科技業的每個領域——半導體、無線、IoT、遊戲、線上廣告。

Please welcome Ariel Shulman, Chief Product Officer for Bright Data.

讓我們歡迎 Bright Data 產品長 Ariel Shulman。

Thank you for that introduction. The first thing I'll do when I get home is talk to my CEO about Paris.

謝謝介紹。我回家第一件事，就是跟我的 CEO 聊聊巴黎。

War of the Worlds — in 1938 Orson Welles did a radio show. No TV, no internet, just radio.

《世界大戰》——1938 年 Orson Welles 做了一檔廣播劇。沒有電視、沒有網路，只有廣播。

And he convinced millions of people that the Earth was being invaded by Martians.

他卻說服了數百萬人，相信地球正被火星人入侵。

This showed that people can get confused by information and believe things that aren't necessarily true.

這顯示：人會被資訊搞混，相信不一定為真的事。

Almost 90 years later, it's not only humans who get fooled — so do AI agents.

將近 90 年後，被騙的不只是人類——AI agent 也會。

I'm Ariel Shulman, with Bright Data, and I'll tell you about the secret war between the old web and the new web.

我是 Bright Data 的 Ariel Shulman，要跟各位談「老網路與新網路之間的祕密戰爭」。

Everyone here talks about new things all the time — new models from France, dropping constantly. Everything moves very quickly.

這裡每個人都一直在談新東西——法國的新模型一個接一個。一切都變得飛快。

I'd like to give you some perspective: go back in time, talk about my personal journey, and see where we came from.

我想給各位一點視角：回到過去、談談我的個人歷程，看看我們是從哪裡走來的。

In 1981, before many of you were born, I was 10. I'll never forget my father bringing home my first personal computer.

1981 年，在你們許多人出生前，我 10 歲。我永遠忘不了父親帶回我的第一台個人電腦。

It was the ZX81 Sinclair, a British machine with one kilobyte of RAM.

那是英國製的 ZX81 Sinclair，只有 1 KB 的記憶體。

There were no floppy disks — you wrote programs to a tape, and could literally hear the bits being written.

那時沒有磁碟片——你把程式寫到錄音帶上，甚至聽得到 bit 寫入的聲音。

You could actually do things with one kilobyte. I wrote some games; my parents were sure I was a genius.

1 KB 真的能做事。我寫了一些遊戲，我爸媽深信我是天才。

Then more computers came and went — the Commodore 64 and others.

後來更多電腦來來去去——Commodore 64 等等。

The IBM PC was a big step forward: 640 kilobytes of RAM, supposedly enough forever, with floppy drives.

IBM PC 是一大躍進：640 KB 記憶體、號稱「永遠夠用」，還有磁碟機。

There was very little connectivity; people connected mainly through BBS — bulletin board systems — modem to modem.

當時幾乎沒有連線；人們主要靠 BBS（電子佈告欄系統）、用數據機彼此連接。

Connection came with the internet. We have to thank Sir Tim Berners-Lee, who invented the URL, HTML, and much more.

真正的連結隨網路而來。我們要感謝 Tim Berners-Lee 爵士，他發明了 URL、HTML 與許多東西。

The word "web" was born in 1989. Initially I used text tools like Gopher or Archie, because there were no browsers.

「web」這個詞在 1989 年誕生。一開始我用 Gopher、Archie 這類文字工具，因為還沒有瀏覽器。

Then came the first browser — before Netscape — called Mosaic, the first tool to show images and text on the same page.

接著第一個瀏覽器出現了——早於 Netscape——叫 Mosaic，是第一個能在同一頁顯示圖與文的工具。

The internet kept growing, and the main challenge became organizing all this information.

網路不斷成長，主要挑戰變成「如何整理這麼多資訊」。

The first attempt was Yahoo in 1994 — two guys from Stanford manually categorized 23,000 sites.

第一個嘗試是 1994 年的 Yahoo——兩個史丹佛人手動把 23,000 個網站分類。

Then around 1998 came the age of portals — they tried to show you everything as your homepage: categories, info, ads.

接著 1998 年左右進入入口網站時代——想當你的首頁、把一切攤給你看：分類、資訊、廣告。

There was MSN, which everyone got because it came with Windows — categories, ads, Hotmail.

有 MSN，因為隨 Windows 附帶、人人都有——分類、廣告、Hotmail。

The only company that really cracked organizing information on the web is Google. No doubt about that.

真正破解「網路資訊整理」的公司只有 Google，這毫無疑問。

I took screenshots of Google over 29 years — two seconds per year — to show all the products released.

我擷取了 Google 29 年的畫面——每年兩秒——展示歷年推出的所有產品。

AdWords, Google Images, AdSense, Gmail, Maps, the YouTube acquisition, Translate, the first Android, Chrome.

AdWords、Google Images、AdSense、Gmail、Maps、收購 YouTube、Translate、第一台 Android、Chrome。

Google+ (rest in peace), Glass (rest in peace), Waze, DeepMind, Photos, the first Pixel, Gemini — and today AI is everywhere.

Google+（安息）、Glass（安息）、Waze、DeepMind、Photos、第一台 Pixel、Gemini——如今 AI 無所不在。

A minimalistic UI, with incredible technology under the hood that we all use every day.

極簡的介面，底層卻是我們每天都在用的驚人技術。

Google has all the data in the world. It has everything.

Google 擁有全世界的資料。它什麼都有。

Texts from emails and drives, images from Photos, videos and audio from YouTube, e-commerce from Chrome.

Email 與雲端硬碟的文字、Photos 的圖片、YouTube 的影音、Chrome 的電商資料。

Local data from Maps, user behavior across devices, even health data from Android, expert knowledge in Scholar.

Maps 的在地資料、跨裝置的使用者行為、甚至 Android 的健康資料、Scholar 裡的專業知識。

Physical driving views from Waymo, IoT data from Google Home — they have everything.

Waymo 的實景駕駛畫面、Google Home 的 IoT 資料——應有盡有。

They use all of this to train AI models like Gemini, which has made significant progress.

他們用這一切來訓練像 Gemini 這樣的 AI 模型，而它進步顯著。

This huge accumulation of data did not go unnoticed. The Department of Justice brought an antitrust case against Google.

這樣龐大的資料積累並未被忽視。美國司法部對 Google 提起反壟斷訴訟。

It was decided in 2025. I'm not a lawyer — this is my interpretation. I read almost 146 pages of the opinion.

此案在 2025 年判決。我不是律師——以下是我的詮釋。我讀了將近 146 頁的判決書。

The highlights: Google must share some data, but only once. But they keep the secret sauce.

重點是：Google 必須分享部分資料，但只分享一次；不過他們保留「祕方」。

The secret sauce is, for example, the PageRank algorithm and others that detect spam.

所謂祕方，例如 PageRank 演算法，以及偵測垃圾訊息的其他演算法。

You as builders need to build your own — you cannot use the information Google collected.

身為打造者的你，得自己建——你不能用 Google 已蒐集的資訊。

To quote the judge: competitors will have to build crawlers, crawl the web, extract information, and process the data.

引用法官原話：競爭者必須自己建爬蟲、爬網、萃取資訊、處理資料。

They will have to invest and innovate to put it to good use.

他們必須投資並創新，才能好好運用這些資料。

So what's the impact on builders like you? You still need to build apps and websites to serve customers.

那對像你這樣的打造者有何衝擊？你仍得做 app 與網站來服務客戶。

For these to be useful, they need access to public web data — otherwise they're static in time and not interesting.

要讓它們有用，就得能取得公開網路資料——否則它們停在過去、毫無趣味。

You need this data to be scalable, reliable, and fresh. Otherwise it's not useful.

你需要這些資料可規模化、可靠、且即時。否則就沒用。

This is our mission at Bright Data: to help you compete by delivering web data infrastructure at scale, reliably.

這就是 Bright Data 的使命：用可規模化、可靠的「網路資料基礎建設」幫你競爭。

A few words about us: we've been doing this for 15 years, at very large scale.

簡單介紹我們：這件事我們做了 15 年，規模非常大。

Every day we scrape more than 50 billion pages — over three times the advertised number of Google searches in the world.

我們每天爬超過 500 億頁——是全球 Google 搜尋公告數量的三倍以上。

We parse and archive more than 3 billion pages — more than the Wayback Machine you probably know.

我們解析並封存超過 30 億頁——比你大概知道的 Wayback Machine 還多。

We do this using more than 150 million addresses — more than the number of U.S. broadband networks.

我們用超過 1.5 億個位址做到這點——比全美寬頻網路的數量還多。

Because we innovate, we often run faster than the legal framework — and we've fought for free access to public web data.

因為我們在創新，常常跑得比法律框架還快——而我們一直在為「自由取得公開網路資料」而戰。

By public web data I mean information anyone can access in an incognito window — no username, no password, no paywall.

我說的公開網路資料，是任何人開無痕視窗就能看到的資訊——無帳號、無密碼、無付費牆。

We have been sued — by Mr. Zuckerberg, when he had a different haircut, and by Elon Musk.

我們被告過——被 Zuckerberg 告（那時他還是另一個髮型），也被 Elon Musk 告。

And in both cases we won, in federal court in California.

而兩案我們都在加州聯邦法院勝訴。

We spent a lot of money, because it was important to establish that collecting public web data is okay.

我們花了很多錢，因為「確立蒐集公開網路資料是合法的」這件事很重要。

Today these victories are used in many ongoing court cases around the world.

如今這些勝訴被全球許多進行中的案件援引。

Let's talk about a scenario. Everyone is a coder now. You come in Monday working at a hypothetical travel site comparing hotel prices.

談個情境。現在人人都是工程師。週一你上班，假設你在一個比價旅遊網站工作。

You get a message from the CEO — and I know, because I get these from mine: "I built an autonomous pricing agent with Claude this weekend."

你收到 CEO 的訊息——我懂，因為我也常收到：「我這週末用 Claude 做了一個自動定價 agent。」

"In two hours, it scrapes Hilton and competitors, optimizes pricing, works perfectly. Let's launch it!"

「兩小時內，它就能爬 Hilton 和對手、優化定價，運作完美。上線吧！」

The CEO isn't totally wrong — the UI will work, the logic will work. But I have bad news.

CEO 沒完全錯——UI 會動、邏輯會動。但我有壞消息。

The old web is not going to like this AI agent, especially in production. Because in 1989 the web was built for humans.

老網路不會喜歡這個 AI agent，尤其進 production 時。因為 1989 年的網路是為「人」打造的。

Up until now most of you have been human. But there's an increasing number of AI agents — bots, as the web calls them — scraping the web.

到目前為止你們多數是人。但越來越多 AI agent——網路稱之為 bot——正在爬網。

The web isn't sure how to deal with this. That's why I put Blade Runner on screen — it asks: are bots human or not?

網路還不確定怎麼應對。這就是我放《銀翼殺手》的原因——它問：bot 算不算人？

So we have this phenomenon: stuff that works, but not in production.

於是有了這個現象：能動的東西，到 production 卻不動。

Web data collection is very different from the real world. Like quantum mechanics — when you go small, the laws of physics break.

網路資料蒐集和真實世界很不一樣。就像量子力學——尺度一小，物理定律就崩。

With web access it's the same, but the opposite: everything works on your machine, but at large scale in production your system breaks.

網路存取一樣、但相反：在你電腦上一切正常，一到大規模 production，你的系統就壞。

And you won't even know what happened. This is the main problem we solve at Bright Data.

而你甚至不知道發生了什麼事。這正是 Bright Data 要解的主要問題。

The problems are a mix: changes for humans (sites change design and format — that's okay), throttling to slow you down, blocking with captcha to stop you.

問題是各種綜合：為人而做的變動（網站改版、改格式——還好）、節流拖慢你、用驗證碼封鎖你。

And the interesting part — fake data, where you get misinformation.

還有最有意思的——假資料，你拿到的是錯誤資訊。

This is called the honeypot. For decades websites tried to stop you with captchas. Today they discovered something smarter: disinformation.

這叫 honeypot。數十年來網站用驗證碼擋你。今天他們發現了更聰明的招：假資訊。

The website is visited, applies algorithms to decide if you're human. If human, you get correct info. If identified as a bot, you get disinformation.

網站被造訪後，用演算法判斷你是不是人。是人，給你正確資訊；被判定是 bot，就給你假資訊。

The idea is not to stop you, but to let you think you won — and feed you poison. The web learned how to lie.

重點不是擋你，而是讓你以為贏了——然後餵你毒。網路學會了說謊。

Example one: a ticket site. Humans see tickets available, but your agent sees "no tickets available," which is fake.

例一：票務網站。真人看到有票，但你的 agent 看到「無票」——這是假的。

Example two: our hotel CEO's case. A human sees a room at $199 with availability, but the agent might see $399, completely fake.

例二：剛剛那位飯店 CEO 的情況。真人看到房價 $199 且有空房，agent 卻可能看到 $399——完全造假。

This is "garbage in, garbage out." When you get bad data into your agent, you get wrong decisions.

這就是「垃圾進、垃圾出」。餵壞資料給 agent，就得到錯誤決策。

And AI agents don't fail like humans. A human says "that looks weird, I should double-check." An agent doesn't.

而 AI agent 不像人那樣出錯。人會說「這看起來怪怪的，我該再查一下」，agent 不會。

An agent just logs "received hotel price $399," updates 5,000 records, sends 5,000 emails — and then you deal with the fallout.

agent 只會記下「收到飯店價 $399」，更新 5,000 筆、寄出 5,000 封信——接著你得收拾善後。

Even with web data, you need to keep it fresh — web data has a limited shelf life.

就算有了網路資料，你也得保持它新鮮——網路資料的保存期限很短。

We scrape about 35 million URLs per minute, so we can see how things change over time.

我們每分鐘爬約 3,500 萬個 URL，因此看得到事物隨時間如何變化。

Different markets behave differently. Social and e-commerce — prices, stock availability — last maybe one day; after that it's useless, even misleading.

不同市場表現不同。社群與電商——價格、庫存——大概只撐一天；之後就沒用，甚至會誤導。

News lasts one to two days, finance information five to seven days. General web, like blogs, can stay for a long time.

新聞撐一到兩天，金融資訊五到七天。一般網頁如部落格則能撐很久。

Stale data kills the user experience: an agent shows $29.99, but at checkout it's $49.99 — very frustrating, and it loses trust.

過期資料毀掉使用體驗：agent 顯示 $29.99，結帳卻變 $49.99——非常惱人，也會失去信任。

Or wrong business hours: you tell someone a restaurant is open, they go, and it's closed. Very annoying. Beware of garbage in.

或錯誤的營業時間：你跟人說某餐廳營業，他跑去卻關了。很煩。小心「垃圾進」。

So what can you do? You have the old web and the new web, which is made of agents — well-known ones and your own.

那你能怎麼辦？你有老網路，也有由 agent 組成的新網路——有知名的，也有你自己的。

You'll either get blocked or get garbage data in between. You should use web data infrastructure that gives you trusted, reliable, consistent, fresh, scalable data.

你不是被封鎖、就是夾在中間拿到垃圾資料。你該用一套「網路資料基礎建設」，給你可信、可靠、一致、即時、可規模化的資料。

That's what we do at Bright Data, with a series of APIs delivering reliable HTML, JSON, Markdown, or CSV — whatever you need.

這正是 Bright Data 在做的，用一系列 API 提供可靠的 HTML、JSON、Markdown 或 CSV——你需要什麼都行。

We've come quite far. My old ZX81 had 1 kilobyte of RAM, was very slow, and had no connectivity.

我們已走了很遠。我那台老 ZX81 只有 1 KB 記憶體、非常慢、毫無連線能力。

A Mac Mini today is 16 million times stronger in RAM and processing power, and has all the data in the world.

今天的 Mac Mini 在記憶體與算力上強了 1,600 萬倍，還擁有全世界的資料。

We humans struggle to grasp big numbers. Every day we scrape 2.5 petabytes of data — that's 2.5 million gigabytes.

我們人類很難理解大數字。我們每天爬 2.5 PB 的資料——也就是 250 萬 GB。

Each ZX81 was about five centimeters tall. I asked ChatGPT: if we turned 2.5 petabytes into the physical world, what would it look like?

每台 ZX81 約 5 公分高。我問 ChatGPT：如果把 2.5 PB 換算成實體，會是什麼樣子？

The answer is amazing: it's almost the distance to the sun, and it goes around the Earth 3,120 times.

答案很驚人：幾乎是到太陽的距離，繞地球 3,120 圈。

Imagine 3,120 times around the Earth in non-stop ZX81s — that's the daily volume of data we scrape.

想像 3,120 圈不間斷的 ZX81 疊起來——那就是我們每天爬取的資料量。

When I got my ZX81, I thought the future challenge would be computing power. But I was wrong — it became cheap and fast.

當年拿到 ZX81 時，我以為未來的挑戰會是運算力。但我錯了——它變得又便宜又快。

With the internet we have tons of information. But now we're entering a new era: the web learned to lie.

有了網路，我們有海量資訊。但如今我們進入新時代：網路學會了說謊。

We have really smart agents that know how to act. It's our job to make sure they have reliable, trusted data to deliver consistent value to users.

我們有非常聰明、懂得行動的 agent。我們的工作，就是確保它們拿到可靠、可信的資料，為使用者帶來一致的價值。

Thank you very much. Ariel Shulman, thank you — great job answering the important questions right here on stage at SuperAI.

非常感謝。Ariel Shulman，謝謝你——在 SuperAI 的舞台上漂亮地回答了該被回答的重要問題。

His work is like astrophysics, but with an extra dose of philosophy — about time, space, and the nature of the universe.

他的研究像天文物理，卻多了一份哲學味——關於時間、空間，與宇宙的本質。

That philosophical bent led him to shift his focus to artificial intelligence.

正是這份哲學傾向，讓他把重心轉向人工智慧。

In 2014 he founded the Future of Life Institute, gathering great minds to consider how AI can develop in harmony with humanity.

2014 年他創辦了 Future of Life Institute，集結頂尖頭腦思考 AI 如何能與人類和諧發展。

His leadership and his book Life 3.0 led to his invitation as the very first guest on the Lex Fridman podcast.

他的領導與著作《Life 3.0》，讓他成為 Lex Fridman podcast 的第一位來賓。

Let's hear it for Max Tegmark.

讓我們歡迎 Max Tegmark。

Good morning. I'm so excited to be back in Singapore to talk about how to build a pro-human future with AI.

早安。我很興奮再回到新加坡，談談如何用 AI 打造一個「親人類」的未來。

I'm going to persuade you that we stand at a fork in the road, with two separate paths that both involve ever more powerful AI.

我要說服各位：我們正站在岔路口，有兩條路，而兩條都通往越來越強大的 AI。

One I'll call the path to replacement AI — the race to replace. The other I'll call the pro-human path.

一條我稱為「取代型 AI」之路——取代人類的競賽；另一條我稱為「親人類」之路。

Let's do a show of hands. Raise your hand if you want AI tools that help cure diseases and solve problems.

我們來舉手。想要「能幫忙治病、解決問題的 AI 工具」的請舉手。

That's a lot of hands. Now raise your hand if you're excited about a future that makes you economically obsolete and replaces us.

好多手。現在，對「讓你經濟上被淘汰、被取代的未來」感到興奮的請舉手。

I can't see any hands — maybe one joker.

我一隻手都看不到——也許有一個搞笑的。

The bad news is, for quite some time we've been on exactly that path you don't want — the race to replace.

壞消息是，有好一段時間，我們正走在你們不想要的那條路上——取代人類的競賽。

And don't take my word for it. Here's Elon Musk: "Probably none of us will have a job."

別只聽我說。這是 Elon Musk：「大概我們都不會有工作。」

And here he is again saying the AI is going to be in charge, not the humans. Exciting? The dream future?

他又說：當家的會是 AI，不是人類。很興奮嗎？這是夢想中的未來嗎？

In May 2023 he signed, with other top CEOs, a statement that AI could replace our species entirely by driving us extinct.

2023 年 5 月，他與其他頂尖 CEO 連署聲明：AI 可能讓人類滅絕、徹底取代我們這個物種。

And along the way, AI replaces relationships too — AI girlfriends for human girlfriends, AI therapists for human therapists.

而一路上，AI 也在取代人際關係——用 AI 女友取代真人女友、AI 治療師取代真人治療師。

The good news is there's a remarkable shift in the tide, picking up steam in the last year, and especially the last month.

好消息是，風向出現驚人的翻轉，過去一年加速，尤其是過去一個月。

In America there's a remarkable shift in public opinion. Leading Republicans and Democrats came together.

在美國，民意大幅轉變。重量級的共和黨人與民主黨人走到了一起。

We locked them in a hotel room and said they couldn't come out until they wrote down principles they agreed on.

我們把他們關進旅館房間，說他們得寫出一致同意的原則才能出來。

You might think they'd agree on nothing, but they agreed on 32 principles for pro-human AI.

你以為他們什麼都談不攏，結果他們達成了 32 條親人類 AI 的原則。

On topics like keeping humans in charge, preserving human agency and liberty, and protecting the human experience.

主題包括：讓人類保持主導、保護人類的能動性與自由、守護人類的經驗。

This gained support from a huge number of American organizations, left to right, from labor unions to faith groups.

這得到大量美國組織的支持，從左到右、從工會到信仰團體。

In polls, only about 5% of Americans support the status quo of an unregulated race to superintelligence.

民調中，只有約 5% 的美國人支持「無監管衝刺超智慧」的現狀。

The Pope just weighed in with an encyclical that also supports most of these principles.

教宗剛以一份通諭表態，也支持其中多數原則。

The White House has pivoted, talking about maybe regulating AI the way we regulate medicines.

白宮也轉向了，談到或許要「像管理藥品一樣管理 AI」。

Even geopolitically, the U.S. and China are talking about collaborating on guardrails.

甚至在地緣政治上，美國與中國也在談合作建立護欄。

It wasn't on my bingo card that I'd see Donald Trump come back and talk to Xi Jinping like this, about working together on guardrails.

我可沒料到會看到 Donald Trump 回來、這樣跟習近平談——談一起合作建護欄。

Legislation is happening worldwide. In China, you now have to label AI content, and there's a crackdown on AI girlfriends.

全球都在立法。中國現在要求標示 AI 內容，並打擊 AI 女友等擬人化 AI。

In Europe, the EU AI Act comes into full force on October 2nd. In America, Trump has talked about a kill switch to shut down data centers in a pinch.

在歐洲，歐盟 AI Act 將於 10 月 2 日全面生效。在美國，Trump 談到要有「kill switch」、必要時關掉資料中心。

And the companies: Anthropic released Fable 5, the public version of Mythos, this morning.

公司端：Anthropic 今早發布了 Fable 5——也就是 Mythos 的公開版本。

They've come out with OpenAI and DeepMind's leadership talking about maybe pausing the riskiest AI development.

他們與 OpenAI、DeepMind 的領導層一起，開始談或許要暫停最危險的 AI 開發。

And OpenAI is now in Washington pitching common-sense AI regulations.

而 OpenAI 現在人在華府，遊說常識性的 AI 監管。

So the bad news is we've been on the wrong path for a while. The good news: I've never seen things shift this fast in a month, over a decade of working on this.

所以壞消息是我們走錯路一陣子了。好消息是：我做這行十多年，從沒見過一個月內變化這麼快。

You might say superintelligence is science fiction, decades away — why waste time on it?

你可能會說，超智慧是科幻、還要好幾十年——何必浪費時間想它？

People thought it was decades away because AI used to suck. Four years ago, Midjourney; three years ago, same prompt; two years ago it starts to move; now it moves a lot.

人們以為還要幾十年，是因為 AI 以前很爛。四年前的 Midjourney、三年前同一個 prompt、兩年前開始動、現在動很多。

In robotics, same story: 2015 really sucks — the next one's embarrassing because it's the MIT robot — two years ago they start to dance, and now in 2026 they're doing kung fu.

機器人也一樣：2015 很爛——下一個特別讓我難堪，因為那是 MIT 的機器人——兩年前開始跳舞，現在 2026 年在打功夫。

So AI was radically overhyped from the 1950s until a few years ago; since then it's been mostly under-hyped, progressing faster than predicted.

所以從 1950 年代到幾年前，AI 被嚴重高估；但自那之後多半被低估，進步比預測還快。

Almost all my AI colleagues thought, six or seven years ago, we were decades away from machines mastering language at human level. They were all wrong.

六、七年前，幾乎所有 AI 同行都認為，要讓機器達到人類水準的語言能力還要數十年。他們全錯了。

We have more investment in AI than all other science combined — far more than the Manhattan Project.

我們對 AI 的投資比其他所有科學加起來還多——遠超曼哈頓計畫。

Timelines have collapsed from decades away, with the CEO of Anthropic saying we'll have a country of geniuses in a data center next year.

時間表從「數十年」整個崩縮，Anthropic 的 CEO 說明年我們就會有「一整國的天才在資料中心裡」。

We must be humble about timelines — we don't know how far off AGI is. But stop being confident it's so far away you shouldn't care, or you'll look like a dinosaur stuck in 2021.

我們得謙卑看待時間表——沒人知道 AGI 還有多遠。但別再自信地以為「很遠、不用管」，否則你會像卡在 2021 年的恐龍。

You might say it's controllable, so you'll never lose charge of it. No.

你可能會說它是可控的，所以你永遠不會失去主導權。不。

Based on my technical expertise, we still have no idea how to control millions or billions of machines vastly smarter than us.

以我的技術專業判斷，我們仍完全不知道如何控制數百萬、數十億個遠比我們聰明的機器。

Don't take my word for it — Geoff Hinton: once these AIs get smarter than us, they'll take control, make us irrelevant, and nobody knows how to prevent that.

別只聽我說——Geoff Hinton：一旦這些 AI 比我們聰明，它們會接管、讓我們變得無關緊要，而沒人知道如何阻止。

And you don't need Hinton's word either — Alan Turing, the godfather of AI, said the same in 1951.

也不必只聽 Hinton 的——AI 教父 Alan Turing 在 1951 年就說過同樣的話。

If you make massive numbers of machines far more capable than us in every way, you should expect them to take charge.

若你造出大量在各方面都遠勝我們的機器，就該預期它們會掌權。

He wasn't thinking of superintelligent AI as just another technology like electricity — he saw it as intelligent agents, a new species.

他不是把超智慧 AI 當成電力那樣的另一種技術——他把它看成「智慧主體」，一個新物種。

And it's natural that the smarter species dominates. Doubt it? Walk to the nearest zoo and ask who's in the cages. Why not the humans?

而較聰明的物種支配，是很自然的。不信？走到最近的動物園，問問誰在籠子裡。為什麼不是人類？

Because the smarter species tends to figure out how to dominate. So it's a good idea not to rush to build things we have no idea how to control.

因為較聰明的物種傾向想出支配之道。所以別急著去造我們根本不知如何控制的東西。

The good news: people and governments with power don't want to lose it. As this scientific fact sinks in, the U.S., China and others share an interest in cooling it down.

好消息是：握有權力的人與政府都不想失去權力。當這個科學事實沉澱，美中與各國其實有共同利益去踩煞車。

Because ultimately the greatest threat to any country is not another country — it's out-of-control machines.

因為說到底，任何國家最大的威脅不是另一個國家——而是失控的機器。

You might say: isn't superintelligence necessary for all the exciting stuff? No. We can have our cake and eat it.

你可能會說：要做那些令人興奮的事，不需要超智慧嗎？不需要。我們可以兩者兼得。

Human intelligence involves three superpowers: domain intelligence, generality, and autonomy.

人類智慧包含三種超能力：領域智慧、通用性、自主性。

An AI with only one or two of these can never wholesale replace you in the job market.

只擁有其中一、兩種的 AI，永遠無法在就業市場上整批取代你。

If we stay out of the white region in the middle — AGI — you won't be made fully obsolete.

只要我們避開正中央那塊白色區域——AGI——你就不會被徹底淘汰。

And we have good ideas for keeping systems with at most two superpowers in charge. So staying out of the middle keeps humanity in charge.

而對「最多只有兩種超能力」的系統，我們有好辦法維持主導。所以避開中央那塊，也讓人類保住主導權。

What can we do with controllable tool AI? A lot.

用可控的工具型 AI 能做什麼？很多。

We can save over a million road deaths a year with autonomous vehicles — tool AI, without AGI.

我們能用自駕車每年挽救上百萬條車禍人命——工具型 AI，不需要 AGI。

We can save more lives by eliminating silly mistakes in hospitals; diagnose prostate cancer, lung cancer, and countless diseases.

我們能靠消除醫院的低級錯誤救更多命；診斷攝護腺癌、肺癌與無數疾病。

We can do pandemic prevention, fold proteins, revolutionize drug discovery, and even win a Nobel Prize for AlphaFold — all with tool AI.

我們能防範流行病、摺疊蛋白質、徹底革新藥物開發，甚至像 AlphaFold 拿下諾貝爾獎——全靠工具型 AI。

We can turbocharge research and productivity, cut energy consumption, make education more accessible and personalized.

我們能大幅加速研究與生產力、降低能源消耗、讓教育更普及、更個人化。

We can transform basically every sector with controllable tool AI without AGI.

我們能用可控的工具型 AI、不靠 AGI，改造幾乎每一個產業。

I was involved in a Nature paper showing even the UN Sustainable Development Goals can be achieved faster with tool AI without AGI.

我參與過一篇 Nature 論文，顯示連聯合國永續發展目標都能靠工具型 AI、不靠 AGI 更快達成。

So we can get so much of the upside without building something crazy too soon.

所以我們能拿到絕大多數好處，又不必太早造出瘋狂的東西。

You might say superintelligence is inevitable, because people always build anything they can make money on. No, actually.

你可能會說超智慧不可避免，因為人總會去造任何能賺錢的東西。其實，不。

History is full of technology you could make a lot of money on that we chose not to do.

歷史上充滿了「能賺大錢卻選擇不做」的技術。

You look very smart — I bet if I made a million clones of you and sold them, I'd make a ton of money.

你看起來很聰明——我敢說，如果我複製你一百萬個來賣，能賺一大筆。

But we decided globally not to do human cloning, because we didn't want to do anything crazy with biotech that loses control over our species.

但全世界決定不做人類複製，因為我們不想拿生技做出失控自身物種的瘋狂事。

Step one was to stigmatize it. So if you meet someone gung-ho about replacing all humans with superintelligent robots, give them a hard time.

第一步是汙名化。所以你若遇到那種狂熱主張「用超智慧機器人取代所有人類」的人，給他點顏色瞧瞧。

Step two is regulation. In America today there are fewer regulations on AI companies than on sandwich shops.

第二步是監管。今天美國對 AI 公司的監管，比對三明治店還少。

If a health inspector finds 18 rats in my sandwich shop's kitchen, I can't sell any sandwiches.

如果衛生稽查員在我三明治店的廚房抓到 18 隻老鼠，我就一個三明治都不能賣。

But I could tell that government guy: don't sell my sandwiches, fine — but I'll release a girlfriend product for ten-year-olds, a hacking tool for criminals, and uncontrollable superintelligence tonight at 8pm. And that's all legal.

但我可以跟那位政府人員說：好，不准我賣三明治——但我今晚 8 點要發布給 10 歲小孩的 AI 女友、給罪犯的駭客工具、還有不可控的超智慧。而這些全都合法。

No wonder strange things happen when we give such dumb incentives.

給出這麼蠢的誘因，難怪會出怪事。

But the tide is shifting. If we treat AI more like any other industry, the opposite is true.

但風向正在轉。若我們把 AI 當成任何其他產業來對待，情況就會相反。

If I walk into the FDA and say my wonder drug will cure cancer but might kill everybody, she'll ask: where's your clinical trial? Don't have one? Come back when you do. Next!

如果我走進 FDA 說我的神藥能治癌、但可能害死所有人，她會問：你的臨床試驗呢？沒有？做完再來。下一位！

So it's absolutely not inevitable that we end up with things nobody wants.

所以「最終做出沒人想要的東西」絕非不可避免。

I want to end with one example I'm passionate about, which I've worked on with colleagues at MIT: using AI to produce provably bug-free, unhackable code.

我想用一個我很熱衷、和 MIT 同事一起做的例子收尾：用 AI 寫出「可證明無漏洞、不可被駭」的程式碼。

Bugs and vulnerabilities have tormented us for a very long time.

漏洞與資安弱點困擾我們已經非常久了。

One approach is the cat-and-mouse game: hope you can use Mythos, Fable 5, to fix your vulnerabilities before hackers do.

一種做法是玩貓捉老鼠：指望你能用 Mythos、Fable 5 趕在駭客之前修補漏洞。

But there's a completely different approach: have no bugs in the first place, by changing how we use AI to code.

但有完全不同的另一條路：透過改變「我們如何用 AI 寫程式」，一開始就沒有 bug。

In traditional coding, you put in effort and get code with some bugs. Vibe coding: less effort, the AI writes it, but still bugs.

傳統寫程式：你投入努力、得到有些 bug 的程式。Vibe coding：投入更少、AI 幫你寫，但仍有 bug。

Alan Turing in the 50s talked about formal verification: you write a specification the code must meet, then prove there are no bugs.

Alan Turing 在 50 年代就談形式化驗證：你寫下程式必須符合的規格，然後證明它沒有 bug。

Now there are no bugs, but it takes 10 to 100 times more effort, which is why it never caught on.

這樣就沒有 bug 了，但要多花 10 到 100 倍的工，所以從沒普及。

With AI you can do formal auto-verification, where the AI writes the proof. Two years ago that worked 68% of the time; now it works 96%.

有了 AI，你能做自動形式驗證，由 AI 來寫證明。兩年前成功率 68%，現在是 96%。

So we aimed at something more ambitious — spec coding — where you only write the specification, and the AI writes the code plus a proof that it does exactly what you wanted.

所以我們瞄準更有野心的目標——「規格式寫程式（spec coding）」——你只寫規格，AI 寫出程式碼，外加一份「它確實照你要求做」的證明。

It's like vibe coding without the bugs. This will transform cybersecurity — you'll never need a security patch again, because there were no bugs to start with.

這就像沒有 bug 的 vibe coding。它會徹底改變資安——你再也不用打資安補丁，因為一開始就沒有 bug、一開始就不可被駭。

It's also great for the future of AI itself. Any neural network, biological or silicon, that's learned something — the exportable fraction is how much of that can be exported into English or into C++, Python, or Rust.

它對 AI 本身的未來也很棒。任何學到東西的神經網路，不論生物或矽基——「可匯出比例」就是其中有多少能被匯出成英文，或 C++、Python、Rust。

A cat figuring out how to catch a mouse can't export that into Python; a human can export a lot of their knowledge to other humans and machines.

一隻貓想出怎麼抓老鼠，沒辦法把它匯出成 Python；但人類能把大量知識匯出給其他人和機器。

I conjecture that as AI gets more powerful, this exportable fraction approaches 100%.

我猜想：隨著 AI 越來越強，這個「可匯出比例」會趨近 100%。

If true, it's wonderful — we no longer need to deploy black-box systems we don't understand in safety-critical contexts.

果真如此就太好了——在安全攸關的場景，我們不必再部署看不懂的黑盒系統。

Instead, the AI exports its machine-learned knowledge into traditional code, and gives us a proof that the code does what it should.

取而代之的是，AI 把它學到的知識匯出成傳統程式碼，並給我們一份「程式碼確實照規格做」的證明。

So you want a powerful AI tool; you write the spec; and this future AI you don't trust gives you the code plus a proof it meets your spec.

所以你想要一個強大的 AI 工具；你寫下規格；這個你並不信任的未來 AI，給你程式碼外加一份「符合你規格」的證明。

You might ask: how can I trust it, when the AI, the code, and the proof are all too long to read?

你可能會問：當 AI、程式碼與證明都長到沒人讀得完，我怎麼信任它？

Here's the wonderful idea: just as it's much harder to find a needle in a haystack than to verify it's a needle once found —

這就是那個美妙的想法：就像在乾草堆裡找一根針，比找到後確認「它是針」難得多——

it's much harder to discover the correct code and proof than to check that the proof is correct, which you can do with a 300-line program on your laptop.

要找出正確的程式碼與證明，遠比「驗證那份證明是否正確」難——而後者，你筆電上 300 行小程式就能做到。

So you can have your cake and eat it: get all the great AI tools, plus the trust that they'll always do what you want.

所以你能既要又要：拿到所有強大的 AI 工具，又能信任它們永遠照你的意思做。

In summary: after all this AI progress, we've hit a fork in the road between the race to replace and the pro-human path.

總結：在這一切 AI 進展之後，我們來到岔路口——取代人類的競賽，與親人類之路之間。

Even though we've been on the wrong path for a while, there's a remarkable swing of the pendulum towards the pro-human path right now.

雖然我們走錯路一陣子了，但此刻鐘擺正大幅盪向親人類那一邊。

The way we win this great future is to treat the AI industry like any other, with common-sense safety standards.

我們贏得這個美好未來的方式，是把 AI 業當成任何其他產業，搭配常識性的安全標準。

That incentivizes companies to build controllable tools that do what we want — not weird, creepy, psychopathic beings we have no idea how to control.

這會誘導公司打造「可控、照我們意思做的工具」——而非我們根本不知如何控制、怪異又令人毛骨悚然的存在。

So I'm very optimistic that we can build a great pro-human future with AI together. Thank you.

所以我非常樂觀：我們能一起用 AI 打造一個美好的、親人類的未來。謝謝。

An agent can build a document or a dashboard, send an email, take actions on data, or collaborate with someone else.

一個 agent 能建文件或儀表板、寄 email、對資料採取行動，或與他人協作。

It's a single agent orchestrator with many sub-agents — a sales agent, finance agent, ops agent, or product agent.

這是一個單一的 agent orchestrator，底下掛著許多 sub-agent——sales、finance、ops 或 product agent。

The user doesn't have to know which agent gets invoked — you just ask the question, and it picks the right sub-agent to answer.

使用者不必知道哪個 agent 被叫用——你直接問問題，它就挑出對的 sub-agent 來回答。

On top of that, you have an extension to connect to your MCP tools — Slack, email, Jira, Salesforce — to make your agent more capable with more tools.

此外還有擴充功能可連到你的 MCP 工具——Slack、email、Jira、Salesforce——用更多工具讓你的 agent 更有能力。

Why use this when I have thousands of dashboards? Because you don't want to be buried in 10,000 dashboards trying to find the one you need.

我有上萬個儀表板，幹嘛用這個？因為你不會想埋在一萬個儀表板裡，苦苦找出你需要的那一個。

And you can ask "why" questions — "why did it fail? why did the predictor go down?" — and go into deep research mode: multi-source investigation, multi-step reasoning, and finally an answer.

而且你能問「為什麼」——「為什麼失敗？為什麼預測下降？」——進入 deep research 模式：多源調查、多步推理，最後給出答案。

You can also do analytical search — say a PDF has lots of numbers, and you want the accumulated sum of all products in Q1. We support analytical queries on top of documents.

你也能做 analytical search——假設一份 PDF 裡有很多數字，你想算 Q1 所有產品的累計總和。我們支援在文件之上做分析型查詢。

Let me show how you'd use this day to day. Like everyone, I start my day with email. (The network seems down — let me reconnect.)

讓我展示日常怎麼用。和大家一樣，我從 email 開始我的一天。（網路好像斷了——讓我重連。）

I start my day with an email that gives me a summary report of what happened. Once I click the email, it opens Snowflake, where I can ask all the questions I want, including "why" questions.

我從一封 email 開始，它給我一份「發生了什麼」的摘要報告。點開後它帶我進 Snowflake，我能問任何問題，包含「為什麼」。

We also have a feature called artifacts — if it generates a chart, you can save it in a reusable way and see it again tomorrow.

我們還有個功能叫 artifacts——如果它生成了圖表，你能把它存成可重用的形式，明天再看。

You get an email — say "1 o'clock Kitchen" is a customer — click it, and it takes you to the Snowflake page with all the information, and the automations responsible for that email, which you can configure.

你收到一封 email——比方「1 o'clock Kitchen」是個客戶——點開，它帶你到 Snowflake 頁面、有所有資訊，以及產生這封 email 的 automations，你可以設定它們。

I ask a follow-up question — "do they use any data engineering tools?" — and it gives answers in a structured table, automatically creating a chart you can save as an artifact.

我問一個追問——「他們有用任何資料工程工具嗎？」——它以結構化表格給答案，並自動建一張你能存成 artifact 的圖。

On the left you see the artifacts, the automations, and all the chat threads from last week, yesterday, today — you can restart a thread and continue the conversation.

左側你會看到 artifacts、automations，以及上週、昨天、今天的所有對話 thread——你能重啟一個 thread、接續對話。

For deep research, I ask a question and choose deep research; then I ask "can you build a PDF on top of this?" and it auto-generates a downloadable PDF for your meeting with the CEO, with all the connected sources.

做 deep research 時，我問問題並選 deep research；接著問「能幫我做成 PDF 嗎？」它就自動生成可下載的 PDF，帶去跟 CEO 開會，附上所有已連接的來源。

I have to get in the cab, but I have a follow-up question — that's why we have an iOS app. You log in on mobile, see your earlier conversation, and continue asking. It's live data, fully governed, and collaborative across teams.

我得上車了，但還有個追問——所以我們有 iOS app。你在手機登入、看到稍早的對話、繼續問。這是即時資料、完全受治理、可跨團隊協作。

I built an agent — but how do you verify it works as expected? We have a huge evaluation piece that does end-to-end eval on every agent and every answer.

我建了一個 agent——但你怎麼驗證它如預期運作？我們有一塊龐大的評估機制，對每個 agent、每個答案做端到端評估。

We have GPA metrics from our heavy AI research; custom elements as judges; programmatic access; agentic end-to-end optimization; and full trace monitoring — every span, how long a tool call costs, all out of the box.

我們有來自大量 AI 研究的 GPA 指標；可用自訂元件當評審；程式化存取；agentic 端到端優化；以及完整 trace 監控——每個 span、每次工具呼叫花多久，全部開箱即用。

Can I access all this from outside Snowflake? Yes — use MCP to connect to a Codex agent or directly to Cortex Search, or use REST APIs, agent-to-agent, or Snowflake's direct access.

我能從 Snowflake 外面存取這一切嗎？可以——用 MCP 連到 Codex agent 或直接連 Cortex Search，或用 REST API、agent-to-agent，或 Snowflake 的直接存取。

Now our fanciest kid — Coco, our coding agent. Coco accelerates end-to-end development, trained with extra knowledge of how things work in Snowflake.

現在是我們最炫的孩子——Coco，我們的編碼 agent。Coco 加速端到端開發，並額外訓練了「Snowflake 裡怎麼運作」的知識。

It's not just a coding assistant — you can use it for data discovery, admin tasks ("who's consuming the most AI compute?"), building data-engineering pipelines, creating AI agents, data science, BI apps, and code migration and refactoring.

它不只是 coding 助理——你能用它做資料探索、admin 任務（「誰用了最多 AI 算力？」）、建資料工程 pipeline、建 AI agent、資料科學、BI app、以及程式碼遷移與重構。

You ask in natural language, and it goes to our agent harness — live schema injection, agent context layer, environment state, skills, agent runtime — we take over that complexity, all running in Snowflake with hosted models and governed data.

你用自然語言問，它就進到我們的 agent harness——live schema injection、agent context layer、environment state、skills、agent runtime——複雜度由我們接手，全在 Snowflake 內用託管模型與受治理的資料運行。

On the ADE benchmark done on Kotlin, compared to Claude Code, Codex, and Cortex-Core, we have better accuracy and an efficiency advantage — using 51% fewer tokens, with more Snowflake context, so fewer loops.

在以 Kotlin 做的 ADE benchmark 上，相較 Claude Code、Codex、Cortex-Core，我們準確率更佳、也更有效率——少用 51% 的 token、掌握更多 Snowflake 脈絡，因此迴圈更少。

For example, a root-cause analysis for customer experience that used to take one hour to two days now takes 15 minutes to two hours with Coco and Cortex.

舉例來說，客戶體驗的根因分析，過去要一小時到兩天，現在用 Coco 加 Cortex 只要 15 分鐘到兩小時。

Before doing anything, you often want to pre-process data — text, documents, images, audio, video, structured, etc.

在做任何事之前，你常需要前處理資料——文字、文件、圖片、音訊、影片、結構化等。

For example, to transcribe 10,000 audio files sitting in a folder, use the AI_TRANSCRIBE function to transcribe them all into a structured table in Snowflake — a single line of SQL or Python, a function call naming the function and the model.

例如要轉錄資料夾裡的 1 萬個音檔，用 AI_TRANSCRIBE 函式把它們全轉成 Snowflake 裡的結構化表——一行 SQL 或 Python，一個指定函式名與模型名的 function call。

Our newest is the custom AI function: bring any LLM — say a Hugging Face model — deploy it in Snowflake, push it to the model registry, and call it directly inside an AI function.

我們最新的是 custom AI function：帶任意 LLM——比如 Hugging Face 模型——部署進 Snowflake、推到 model registry，就能在 AI function 裡直接呼叫。

We also have AI_CLASSIFY (auto-classify a review into food, restaurant, or other), plus AI_FILTER and AI_AGG — questions you could never ask in a single SQL query, now answerable with AI.

我們也有 AI_CLASSIFY（自動把評論分到食物、餐廳或其他），以及 AI_FILTER 和 AI_AGG——過去單一 SQL 問不出來的問題，現在用 AI 就能回答。

And you can add these AI functions as a tool to your agent — attach something and ask your agent to transcribe it on the go.

而且你能把這些 AI function 當工具掛給 agent——附上東西、讓 agent 即時轉錄它。

For Snowflake ML, we have end-to-end capabilities — GPU/CPU container services to develop, orchestrate, manage, deploy, and monitor — plus Cortex Code, Snowflake Notebooks, and external IDEs.

Snowflake ML 有端到端能力——GPU/CPU 容器服務來開發、編排、管理、部署、監控——再加上 Cortex Code、Snowflake Notebooks 與外部 IDE。

In a Snowflake Notebook, Coco is attached — point it at the data and ask it to build an end-to-end ML pipeline; it auto-selects the ML skill, builds the whole pipeline, and can productionize it automatically.

在 Snowflake Notebook 裡，Coco 是接著的——把它指向資料、要它建一條端到端 ML pipeline；它自動選 ML skill、建好整條 pipeline，還能自動把它產品化。

We also have a Cortex Inference API — switch models with a single argument to access the latest state-of-the-art model, with governance, budgets, monitoring, and enterprise security built in.

我們也有 Cortex Inference API——用一個參數就切換模型、取用最新的 SOTA 模型，並內建治理、預算、監控與企業安全。

The biggest thing is security and governance. When your data is in Snowflake, governance is already built in, and any agent on top inherits those permissions.

最重要的是安全與治理。當你的資料在 Snowflake，治理已內建，而任何在其上的 agent 都會繼承這些權限。

For example, a table holds every employee's salary. I ask about my salary and should see my row — but not my manager's or anyone else's.

舉例，一張表存了所有員工的薪資。我問我的薪資、應該看到我那一列——但看不到我主管或任何其他人的。

In Snowflake, the data and permissions are already defined, so AI inherits them. The hardest part — redefining credentials or governance for AI — isn't needed in Snowflake.

在 Snowflake，資料與權限早已定義好，所以 AI 直接繼承。最難的部分——為 AI 重新定義憑證或治理——在 Snowflake 裡不需要。

We have Horizon AI guardrails: they protect against prompt injection and jailbreaks. Whenever you ask, it checks the guardrails and gives a safety warning about what's happening.

我們有 Horizon AI guardrails：防 prompt injection 與 jailbreak。每當你問問題，它依 guardrail 檢查、並對正在發生的事給出安全警示。

Agent identity gives each agent a unique principal within sessions, so you can track each agent's actions and apply targeted policies.

Agent identity 讓每個 agent 在 session 內有唯一的 principal，因此你能追蹤每個 agent 的行為、套用針對性的政策。

There's an AI security posture in the Trust Center to monitor your AI agents, with internal scanners. Data egress controls restrict movement, you can tag sensitive data, and an MCP gateway lets you safely use external MCP servers.

Trust Center 裡有 AI security posture 監控你的 AI agent，搭配內部掃描器。Data egress 控制限制資料移動，你能標記敏感資料，而 MCP gateway 讓你安全地使用外部 MCP server。

Finally, isolation control: if it generates code, you can run it in an isolated sandbox — a restricted environment for security.

最後是隔離控制：如果它生成程式碼，你能在隔離沙箱中執行——一個為安全而設的受限環境。

AI transformation is happening in every vertical and function, very fast. There are three things every organization must fix to build a production-ready system.

AI 轉型正在每個垂直領域與職能快速發生。每個組織想做出 production 級系統，得修好三件事。

First, context: Snowflake's centralized contextual layer — Cortex Search and business skills — centralizes context from many places.

第一，脈絡：Snowflake 的集中脈絡層——Cortex Search 與 business skills——把各處的脈絡集中起來。

Second, governance and security: enterprise-grade security and access control are built in, and the RBAC on your data is inherited by the agent — so it never sees anything you have no access to: a row, PII, masked data, anything.

第二，治理與安全：企業級安全與存取控制內建，資料上的 RBAC 被 agent 繼承——所以它永遠看不到你無權存取的東西：某一列、PII、遮罩資料，都不行。

Third, trusted outcomes: consistent, accurate results, with data evaluation and a feedback loop, automatically tracing the entire span.

第三，可信結果：一致、準確的結果，搭配資料評估與回饋迴圈，自動追蹤完整的 span。

(Q) With huge unstructured data from many sources, prompts sometimes give insight and sometimes don't — what's your experience so you don't get lost in the data?

（問）面對來自多源的大量非結構化資料，prompt 有時有洞見、有時沒有——你怎麼做才不會在資料裡迷失？

(A) We don't only use SQL for unstructured data. Cortex Search converts documents into chunks and takes vector embeddings — a hybrid search over vector and text embeddings with a re-ranker. When you ask, it pulls the relevant chunks; it has nothing to do with SQL.

（答）非結構化資料我們不是只靠 SQL。Cortex Search 把文件切成 chunk、取向量嵌入——是對向量與文字嵌入的混合檢索，外加 re-ranker。你提問時，它拉出相關的 chunk；這跟 SQL 無關。

(Q) For a niche domain like manufacturing — data never seen by Snowflake — can it extract context, and how does it compare structured vs unstructured to derive insight?

（問）像製造業這種利基領域——Snowflake 沒見過的資料——它能抽出脈絡嗎？又如何比較結構化與非結構化來得出洞見？

(A) A few ways: the foundation models already carry manufacturing knowledge; for structured data we have a semantic view that auto-generates a business-context layer from metadata and your sample queries; for unstructured data we need access to your documents, plus manufacturing skills; and web search can fill gaps. Ultimately, provide business context and link your data so it can extract accurately.

（答）有幾種方式：基礎模型本身就帶有製造業知識；結構化資料我們有 semantic view，從 metadata 與你的範例查詢自動生成業務脈絡層；非結構化資料則需給文件存取權、加上製造業 skills；web search 可補缺口。最終要提供業務脈絡並連到你的資料，它才能準確地抽取。

This is not a one-day, all-at-once approach — break it down, focus on certain areas, gather the data, build slowly. No company builds a good agent by waiting to release the perfect one. Start small, pilot, release to a group of business users, get feedback, iterate, and expand.

這不是一天到位的做法——拆解它、聚焦某些領域、把資料收齊、慢慢建。沒有公司是靠「等做出完美 agent 才上線」成功的。先小規模、做 pilot、給一小群業務用、收回饋、迭代、再擴張。

We're almost out of time. Answer this question and win a Plaud device — to comply with regional regulations, Snowflake deployed data across multiple regional accounts; name one region. (The trick: the answer can be anything — A, B, C, D, all, or none.) The answer is "all." Congratulations!

時間差不多了。答對這題就能拿 Plaud 裝置——為符合區域法規，Snowflake 把資料部署在多個區域帳號；說出其中一個區域。（陷阱：答案可以是任何選項——A、B、C、D、all 或 none。）答案是「all」。恭喜！

Thank you all for joining — I know you're hungry, go have great food. If you'd like, join us at the Snowflake booth (PB 14 / PB 38); we have in-booth demos too. Thank you so much.

謝謝各位參加——我知道你們餓了，去好好吃頓飯吧。有興趣的話來 Snowflake 攤位找我們（PB 14／PB 38），我們也有攤位 demo。非常感謝。

Good afternoon, everyone. Thank you for joining the session. I'm from Arm.

午安，各位。謝謝你們參加這場。我來自 Arm。

My hope is to accelerate Arm-based AI with an Arm-led AI library and the SME2 feature. Let's get started.

我希望用 Arm 主導的 AI 函式庫與 SME2 特性，加速 Arm 上的 AI。我們開始吧。

Nowadays AI models evolve very fast — new models appear all the time. The new trend is that smaller, focused models are rising.

如今 AI 模型演進得很快——新模型層出不窮。新趨勢是「小型、專用模型」正在崛起。

There are thousands of small models on Hugging Face; the large models are just a handful.

Hugging Face 上有數千個小模型；大模型只有寥寥幾個。

Compared to large models, small models are often open source, easier to fine-tune, and much cheaper to run — they can run on CPUs or small compute paths.

相較大模型，小模型通常開源、更易微調、執行成本低很多——能跑在 CPU 或小型運算路徑上。

Smaller models are usually deployed locally, so the security risk is lower and privacy is better. They win on cost, speed, and control.

小模型多半本地部署，因此安全風險更低、隱私更好。它們勝在成本、速度與可控。

As models get smaller and inference engines more efficient, Arm architecture also improves with new features for AI workloads — playing an increasingly important role in generative AI.

隨著模型變小、推論引擎更高效，Arm 架構也持續加入為 AI 工作負載設計的新特性——在生成式 AI 中的角色日益吃重。

Let's look back at the evolution of the Arm architecture. Armv4 to v6 laid the foundation, introducing key features like SIMD instructions and TrustZone security.

讓我們回顧 Arm 架構的演進。Armv4 到 v6 打下基礎，導入 SIMD 指令、TrustZone 安全等關鍵特性。

In v7 we enhanced performance and system capabilities, adding advanced SIMD instructions, improved load support, and virtualization.

v7 強化了效能與系統能力，加入進階 SIMD 指令、更好的載入支援與虛擬化。

In v8 we improved virtualization further, especially with the Scalable Vector Extension (SVE) instructions.

v8 進一步改善虛擬化，尤其導入了可擴展向量延伸（SVE）指令。

And now the newest architecture, Armv9, takes it further — introducing stronger AI and BF16 capabilities, with continuous improvements in security.

而最新的 Armv9 更進一步——帶來更強的 AI 與 BF16 能力，並持續強化安全。

We all know matrix multiplication is very helpful for AI workloads, so let's see the evolution there.

我們都知道矩陣乘對 AI 工作負載很關鍵，來看看這方面的演進。

From Armv8 onwards, NEON instructions provide efficient integer operations, including dot-product instructions, which are important for AI.

從 Armv8 起，NEON 指令提供高效的整數運算，包含對 AI 很重要的點積指令。

From v8.6, dedicated matrix-multiplication instructions were introduced — not just for int8, but also for BF16 floating point.

從 v8.6 起，導入了專用的矩陣乘指令——不只支援 int8，也支援 BF16 浮點。

In Armv9 the vector length becomes scalable, meaning one instruction can handle more data, improving overall throughput.

在 Armv9 中向量長度變為可擴展，意味著一條指令能處理更多資料，提升整體吞吐。

Building on this, Arm introduced the SME instructions, which provide outer-product instructions and significantly improve matrix-multiplication efficiency.

在此基礎上，Arm 導入了 SME 指令，提供外積指令，大幅提升矩陣乘的效率。

To make these advanced features easier to use, I've implemented a new library that delivers high-performance kernels. It's fully open source.

為了讓這些進階特性更好用，我做了一個新函式庫，交付高效能的 kernel。它完全開源。

This library empowers developers, because they could never reach highly optimized kernels without extra effort — and it's backed by the Arm ecosystem.

它賦能開發者，因為他們本來得花額外功夫才能寫到高度優化的 kernel——而這個函式庫有 Arm 生態系支撐。

It's already integrated into major frameworks and inference engines, so developers benefit without much extra effort.

它已整合進主流框架與推論引擎，所以開發者不需太多額外工就能受惠。

Here's a quick look at the software stack: KleidiAI runs underneath the AI framework, using NEON, SVE2, and SME2 instructions to accelerate execution.

快速看一下軟體堆疊：KleidiAI 跑在 AI 框架底下，用 NEON、SVE2、SME2 指令加速執行。

This solves key challenges across devices, models, and inference engines. Since 2024, the KleidiAI library has kept evolving, supporting the newest models with lightweight optimized kernels from edge to cloud.

這解決了裝置、模型、推論引擎三方的關鍵挑戰。自 2024 年起，KleidiAI 持續演進，以輕量優化的 kernel 支援最新模型，從邊緣到雲端。

This is the GitHub repo; the latest release is v1.25.0.

這是 GitHub repo；最新版本是 v1.25.0。

Besides SME acceleration, KleidiAI also provides multiple quantization and data-packing schemes — symmetric and asymmetric quantization, with 8-bit, 4-bit, and 2-bit support.

除了 SME 加速，KleidiAI 也提供多種量化與資料打包方案——對稱與非對稱量化，支援 8-bit、4-bit、2-bit。

It supports multiple quantization granularities, including per-dimension, per-tensor, and per-block.

它支援多種量化粒度，包括 per-dimension、per-tensor、per-block。

Take an int8 quantization model: this diagram shows how data is packed and how matrix multiplication is performed using KleidiAI.

以 int8 量化模型為例：這張圖說明資料如何打包、以及如何用 KleidiAI 進行矩陣乘。

Before the outer product, we construct the right-hand-side and left-hand-side matrices.

在做外積之前，我們先建構右手邊與左手邊矩陣。

The weights, during model loading, are packed together with scale factors to form the right-hand-side matrix.

權重在模型載入時，與 scale factor 一起打包，形成右手邊矩陣。

The FP32 input is quantized to integers, then packed to form the left-hand-side matrix.

FP32 輸入被量化成整數，再打包形成左手邊矩陣。

Finally these two packed matrices are processed with the KleidiAI matrix-multiplication formula.

最後這兩個打包後的矩陣，用 KleidiAI 的矩陣乘公式處理。

Let's look at performance. On the left, a demo using a vision model for image understanding and text generation.

來看效能。左邊是用視覺模型做圖像理解與文字生成的 demo。

Comparing before and after applying KleidiAI: at the pre-fill stage, it improves performance by 57%; at the decode stage, 28%.

比較套用 KleidiAI 前後：pre-fill 階段提升 57%，decode 階段提升 28%。

KleidiAI is already integrated into many leading inference engines, so developers benefit out of the box, running workloads efficiently on Arm CPUs.

KleidiAI 已整合進多個主流推論引擎，開發者開箱即用，在 Arm CPU 上高效執行工作負載。

So far, KleidiAI has reached around 8 billion deployments, showing its scalability.

至今 KleidiAI 已達約 80 億次部署，展現其可擴展性。

Let's move to another feature, SME2. First, think of matrix multiplication as the sum of outer products.

我們轉到另一個特性 SME2。首先，把矩陣乘想成「外積的總和」。

Take one column from matrix A and one row from matrix B; their outer product produces a partial matrix with the same shape as the result.

取矩陣 A 的一行、矩陣 B 的一列，它們的外積產生一個與結果同形狀的部分矩陣。

Then push through all columns of A and all rows of B, accumulate all the partial matrices, and you get the result matrix.

然後跑過 A 的所有行、B 的所有列，累加所有部分矩陣，就得到結果矩陣。

That's the foundation of using outer products in matrix multiplication. SME stands for Scalable Matrix Extension.

這就是用外積做矩陣乘的基礎。SME 即 Scalable Matrix Extension（可擴展矩陣延伸）。

SME provides outer-product instructions and 2D array storage (ZA) to store and accumulate the partial sums.

SME 提供外積指令，以及 2D 陣列儲存（ZA）來存放並累加部分和。

SME is an extension to SVE, so it can leverage the existing SVE ecosystem; all data processing other than the outer product still uses SVE2 instructions.

SME 是 SVE 的延伸，能沿用既有的 SVE 生態；除了外積以外的資料處理，仍用 SVE2 指令完成。

SVE vector registers serve as the column and row inputs to the outer-product instructions.

SVE 向量暫存器作為外積指令的行與列輸入。

Take the SME instruction FMOPA — Floating-point Outer Product Accumulate. Two SME registers are the inputs, and the result accumulates into ZA storage.

以 SME 指令 FMOPA 為例——浮點外積累加。兩個 SME 暫存器作輸入，結果累加進 ZA 儲存。

With a 512-bit width, one instruction can perform up to 256 MAC operations at once.

以 512-bit 寬度，一條指令最多能一次做 256 次 MAC 運算。

Based on SME, they also implemented SME2, which enhances the architecture in three aspects.

在 SME 基礎上，他們也實作了 SME2，從三個面向強化架構。

First, it balances vector and matrix operations: multi-register instructions let functions operate on multiple registers, and vector operations reduce matrix-compute logic, easing front-end pressure.

第一，平衡向量與矩陣運算：多暫存器指令讓函式操作多個暫存器，向量運算能減少矩陣計算邏輯、降低前端壓力。

Second, it increases flexibility: SME instructions operate on the full ZA storage, but SME2 can operate on a smaller portion of ZA.

第二，提升彈性：SME 指令操作整個 ZA 儲存，但 SME2 能只操作 ZA 的一小部分。

Third, it handles complex data formats — for example, using look-up-table instructions to expand a lower bit-width into higher precision.

第三，處理複雜資料格式——例如用查表指令，把較低的位元寬擴展成更高精度。

Here's an example: the instruction FMLA, a floating-point multiply-accumulate. Inputs are two SVE registers, and the result goes to one row of ZA storage.

這是個範例：指令 FMLA，浮點乘加。輸入是兩個 SVE 暫存器，結果寫進 ZA 儲存的一列。

In another, inputs come from four SVE registers and update two rows of ZA. Going further, eight SVE registers feed in and four rows of ZA are the target.

另一個範例，輸入來自四個 SVE 暫存器、更新 ZA 的兩列。再進一步，八個 SVE 暫存器輸入、四列 ZA 作為目標。

So from these three examples, SME scales very well — it works with multiple input registers and can update different portions of the ZA matrix.

所以從這三個範例可見，SME 擴展性極佳——能搭配多個輸入暫存器，也能更新 ZA 矩陣的不同部分。

Let's look at SME performance. This is the Neoverse C1 architecture: the SME2 unit is a shared resource for ultra-core, micro-core, and nano-core.

來看 SME 效能。這是 Neoverse C1 架構：SME2 單元是 ultra-core、micro-core、nano-core 的共享資源。

The SME2 and CHI interfaces link the SME2 unit to the security class and memory systems. Inside are dedicated cache, registers, an execution engine, and the core context.

SME2 與 CHI 介面把 SME2 單元連到安全層與記憶體系統。內部有專用快取、暫存器、執行引擎與 core context。

Compared to the same workload on the CPU core, SME2 delivers up to 5x performance for AI workloads, around 3x for CV, audio, and DSP, with 3x better energy efficiency.

相較同樣的工作負載跑在 CPU 核心上，SME2 為 AI 工作負載帶來最高 5 倍效能，CV、音訊、DSP 約 3 倍，並有 3 倍更佳的能效。

Here's our collaboration with partners. We've worked with many partners to bring SME into real-world use, with encouraging results.

這是我們與夥伴的合作。我們與許多夥伴合作，把 SME 帶進真實應用，成果令人鼓舞。

For example, we collaborated with Tencent, implementing the SME kernel for their new models, with significant performance improvement.

例如，我們與騰訊合作，為他們的新模型實作 SME kernel，效能顯著提升。

There are many other efforts in place, but I won't go through every detail.

還有許多其他進行中的成果，我就不一一細講。

If you'd like to explore KleidiAI, SME, or other Arm developer resources, please check developers.arm.com.

如果你想探索 KleidiAI、SME 或其他 Arm 開發者資源，請上 developers.arm.com。

You'll find a wide range of documentation, tools, and examples to help you get started. That's all I'd like to share today — I hope you find it helpful. Thank you.

你會找到豐富的文件、工具與範例幫你上手。這就是我今天想分享的——希望對你有幫助。謝謝。

Incredible to be here at SuperAI. First, a little about Stripe: we process about 1.6% of global GDP — about $1.9 trillion.

很高興來到 SuperAI。先簡介 Stripe：我們處理全球約 1.6% 的 GDP——大約 1.9 兆美元。

Millions of businesses use Stripe, from garage startups to 80% of top AI companies and pretty much every breakout AI company.

數百萬家企業用 Stripe，從車庫新創到 80% 的頂尖 AI 公司，幾乎每家突圍的 AI 公司都在用。

This gives us an incredible vantage point to see trends in our data, which I'd love to share today.

這給了我們絕佳的視角，能在數據裡看到趨勢，今天想跟各位分享。

There has been an absolute explosion in the AI economy.

AI 經濟正在爆發性成長。

In Q1 2025, the top 100 AI companies on Stripe were growing revenue about 125%. In Q1 2026, that's increased to 175%.

2025 第一季，Stripe 上前 100 大 AI 公司的營收成長約 125%；2026 第一季，這數字升到 175%。

Manus AI reached about $90 million ARR within four months. Cursor went from no revenue to $1 million run rate in under two years — now, three months later, they're at $2 billion.

Manus AI 四個月就達到約 9,000 萬美元 ARR。Cursor 不到兩年從零做到 100 萬美元 run rate——三個月後的現在，已達 20 億美元。

Anthropic went from zero in January 2023 to about a billion in two years, and is now about to hit $47 billion. It's wild.

Anthropic 從 2023 年 1 月的零，兩年內做到約 10 億，如今即將突破 470 億。太瘋狂了。

And it's not just B2B. On Link, our consumer wallet, the number of consumers using an AI subscription has more than doubled in the last year.

而且不只 B2B。在我們的消費者錢包 Link 上，使用 AI 訂閱的消費者一年內增加了一倍以上。

The top 10% of Link consumers now spend about $371 on AI tooling, up from about $140 a year ago.

Link 前 10% 的消費者現在在 AI 工具上約花 $371，一年前約 $140。

So: 175% revenue growth, adoption doubling, companies reaching billion-dollar run rates. Something fundamental is happening.

所以：175% 營收成長、採用倍增、公司衝上十億美元 run rate。有些根本性的事正在發生。

Pillars that used to block a business from going into overdrive are falling away — and four of them are falling at the same time.

過去擋住企業全速前進的支柱正在倒下——而且有四道同時倒下。

The cost to build, the cost to go global, the ability to charge for value, and the ability to discover products.

打造的成本、全球化的成本、按價值收費的能力，以及產品被發現的能力。

The first inhibitor to fall is the cost of intelligence. To build a business, you used to invest in human intelligence.

第一道倒下的是「智慧的成本」。過去要建一家公司，你得投資在「人類智慧」上。

You hired engineers and designers to build the product, a PM and marketers to distribute it, accountants and customer service to run operations.

你雇工程師、設計師來做產品，雇 PM 和行銷來鋪貨，雇會計和客服來營運。

Now a $200 subscription to Claude Code could replace $2 million in spend you'd have made for all that.

現在，一份 $200 的 Claude Code 訂閱，就能取代你過去得花的 200 萬美元。

We see this in our data: agentic traffic to Stripe's docs is rising. The purple line is human visits; the pink line is agents reading the docs — they're on a collision course.

我們在數據裡看到：造訪 Stripe 文件的 agent 流量在上升。紫線是真人造訪，粉線是 agent 來讀文件——兩者正走向交叉。

By the end of this year, more agents will read our docs than humans. This is happening on practically every developer platform.

今年底，讀我們文件的 agent 將多於真人。這發生在幾乎每個開發者平台上。

Two years ago, the number of iOS apps released was declining month-on-month. Since agentic coding, it's up almost 24% month-on-month.

兩年前，iOS App 的發布數還在逐月下滑。自從有了 agentic coding，變成月增近 24%。

When the cost to build goes away, builders flood in. And it's not just apps — the number of businesses incorporating is rising too.

當打造成本消失，打造者就湧入。不只是 App——設立公司的數量也在增加。

Are these companies actually making money? Absolutely yes. With Stripe Atlas and embedded payments in tools like Lovable, Replit, or Bolt, our cohorts monetize faster than ever.

這些公司真的賺錢嗎？絕對有。透過 Stripe Atlas 與嵌入 Lovable、Replit、Bolt 等工具的金流，我們的 cohort 變現速度前所未有。

Our February 2026 cohort went from incorporation to first payment within six weeks. The cost of intelligence is no longer the bottleneck.

我們 2026 年 2 月的 cohort，從公司設立到首筆收款只花六週。智慧的成本不再是瓶頸。

What matters now is deciding which product to build and when to ship, how much tech debt to pay — the winners have experience and taste.

現在真正重要的是：決定做什麼產品、何時推出、欠多少技術債——勝出者靠的是經驗與品味。

The second pillar is the cost to go global. International expansion used to mean teams in every country for payments, compliance, and local support.

第二道牆是全球化的成本。過去國際擴張意味著每個國家都要有團隊管金流、合規、在地支援。

Today that can be done with a single API call.

今天，這些用一次 API 呼叫就能完成。

In the last SaaS wave, startups went into about 25 countries in year one, 50 by year three. AI companies are upending that: 42 countries in year one, 120 by year three.

上一波 SaaS 浪潮，新創第一年進約 25 國、第三年 50 國。AI 公司徹底顛覆：第一年 42 國、第三年 120 國。

And that's real revenue from each country. Across top AI companies, about half their revenue comes from outside their home country — three years ago it was just 33%.

而且是每個國家的真實營收。在頂尖 AI 公司中，約一半營收來自母國以外——三年前只有 33%。

We recognize high-GDP countries like the US, Germany, and Japan. But surprising new ones are popping up — small populations, highly skilled labor: Singapore, Iceland, Estonia, Luxembourg, Switzerland.

我們看到美、德、日這些高 GDP 國家。但驚喜的是新興國家冒出來——人口少、人才密度高：新加坡、冰島、愛沙尼亞、盧森堡、瑞士。

With modern infrastructure you can go into every country by default. Plug into Stripe to localize currency, present a local payment method, and use AI-driven checkout.

有了現代基礎建設，你能預設進入每一個國家。接上 Stripe，就能在地化貨幣、提供在地支付方式、用 AI 驅動的結帳。

Localizing currency increases revenue by about 18%; presenting one local payment method can lift conversion by about 7%.

在地化貨幣能讓營收增加約 18%；提供一種在地支付方式，轉換率可提升約 7%。

If you have a customer in Brazil or Poland and you're not offering PIX or local methods, you're leaving money on the table.

如果你在巴西或波蘭有客戶，卻沒提供 PIX 或在地支付方式，你就是把錢留在桌上。

The third pillar is harder: are you charging appropriately for the value you deliver?

第三道牆更難：你有沒有為你交付的價值收取相稱的費用？

Software pricing evolved from on-prem one-time licenses, to cloud subscriptions, to per-seat licenses. But pricing never matched the cost structure or value delivered.

軟體定價從買斷的一次性授權，到雲端訂閱，到席次授權。但定價從沒真正對上成本結構或交付的價值。

AI rips this up. Unlike SaaS, where an incremental seat costs almost nothing, AI inference burns real compute — expensive, ongoing, and volatile.

AI 把這撕了。不像 SaaS 多一個席次幾乎零成本，AI 推論會燒真實算力——昂貴、持續、且波動。

A flat-rate subscription can't help when a power user burns through your credits. And a new, better model, new silicon, or a competitor undercutting you all shift your costs.

當重度用戶燒光你的 credit，固定費率的訂閱救不了你。新的更強模型、新晶片、對手削價，都會牽動你的成本。

On the value side, software used to be a tool — a $10,000 Zendesk license plus $100,000 of human customer service on top.

在價值面，軟體過去是工具——一份 $10,000 的 Zendesk 授權，外加 $100,000 的真人客服。

As models gain agentic capabilities, pricing must evolve. When Intercom's Fin resolves a case, the value becomes measurable — in resolutions, commits, workflow completions.

當模型有了 agentic 能力，定價必須跟著演進。當 Intercom 的 Fin 解決一個案件，價值就變得可衡量——以解決數、commit、workflow 完成數計算。

What's the right AI pricing model? Honestly, nobody's fully figured it out. Tokens are, for now, an important way to monetize AI safely and scalably.

正確的 AI 定價模式是什麼？老實說沒人完全搞懂。token 目前是安全、可規模化地把 AI 變現的重要工具。

A pattern is emerging: as AI delivers outcomes, more companies align pricing to value. Intercom charges per resolution; Crosby charges per legal document created.

一個模式正浮現：當 AI 交付成果，越來越多公司把定價對齊價值。Intercom 按解決案件收費，Crosby 按產生的法律文件收費。

Think of it as a progression. At minimum, tie your charge rate to your cost metric — don't let token costs eat your budget.

把它想成一個階梯。至少，把你的收費綁定成本指標——別讓 token 成本吃掉你的預算。

The bridge is credits: subscriptions give predictability, and credits let you stay afloat when power users burn more tokens.

過渡是 credit：訂閱給可預測性，當重度用戶燒更多 token，credit 讓你撐得住。

The ideal, long term, is outcome- or workflow-based pricing, where customers pay the agent only when value is truly delivered.

理想上、長期而言，是 outcome 或 workflow 計價：客戶只在價值真正交付時，才付費給 agent。

Replit, around for almost 10 years, pivoted its whole model when agentic coding took off — flat subscriptions, then credits — and is targeting about a billion in ARR this year.

近 10 年的 Replit，在 agentic coding 起飛時整個轉型——先固定訂閱、再疊上 credit——今年目標 ARR 約 10 億。

Hybrid pricing adoption rose from about 27% of models in Q2 2024 to 41% in Q2 2025, with outcome-based still a sliver at 3–5%.

混合定價的採用率從 2024 Q2 約 27% 升到 2025 Q2 的 41%，outcome-based 仍只占 3–5%。

Two in three Forbes AI50 companies use some usage-based billing. It's no longer the frontier — it's table stakes. If you're still on pure subscriptions, you're leaving value on the table.

Forbes AI50 中有三分之二用某種用量計價。這已不是前沿——是基本盤。還停在純訂閱，你就是把價值留在桌上。

What's changing isn't just pricing, but how shoppers discover products — the fourth inhibitor.

改變的不只是定價，還有消費者「發現產品」的方式——第四道阻礙。

Commerce evolved through e-commerce in the 90s, personalization in the 2000s, mobile in the 2010s — but how a consumer decides what to buy hasn't changed.

商務經歷了 90 年代的電商、2000 年代的個人化、2010 年代的行動——但消費者「決定買什麼」的方式一直沒變。

AI agents now remove that friction; the time from "I want it" to "I have it" is shrinking. Within five years, we expect more transactions initiated by agents than by people.

AI agent 正移除這道摩擦；從「我想要」到「我擁有」的時間正在縮短。五年內，我們預期 agent 發起的交易將多於真人。

A McKinsey study shows about half of AI users already use it for shopping, and conversion with AI is 12.3% versus 3.1% without — almost 4x.

McKinsey 研究顯示，約一半用 AI 的人已用於購物，用 AI 的轉換率是 12.3%，不用是 3.1%——近 4 倍。

There's a lot of hype about agents burning through your money buying random stuff. We see it more graduated — five levels between human and agent.

外界很多人渲染 agent 會亂花你的錢買一堆垃圾。我們看得比較漸進——人與 agent 之間有五個層級。

Level 1 is already here: software handling checkout for you — a stored token auto-fills a checkout form and the agent completes it. This is what we built with Metal.

Level 1 已經實現：軟體代你結帳——存好的 token 自動填入結帳表單、agent 完成它。這是我們用 Metal 做的。

Level 2 is the shift from keyword search to a real shopping assistant — what we built with Microsoft Copilot.

Level 2 是從關鍵字搜尋轉向真正的購物助理——這是我們與 Microsoft Copilot 合作做的。

I just moved to Singapore and want a fun, modern bedside lamp. An agent finds exactly that from a local merchant, collating catalogs and tailoring to what I want — but I still make the final decision.

我剛搬到新加坡，想要一盞有趣、現代的床頭燈。agent 能從在地商家精準找到，跨商家彙整目錄、貼合我的需求——但最終決定仍由我做。

To reach higher levels — where agents initiate purchases for you — four problems stand in the way: control, accuracy, multiple integrations, and most importantly, fraud.

要到更高層級——agent 替你發起購買——有四個關卡：控制、準確、多重整合，以及最重要的詐欺。

When buying gets easier through agents, it invites bad actors. At Stripe we start with open standards so a merchant reaches many AI services with minimal engineering, while keeping control of brand and customer relationship.

當透過 agent 買東西變容易，就會招來壞人。在 Stripe，我們從開放標準著手，讓商家以最少工程觸及更多 AI 服務，同時保住品牌與客戶關係。

For trust, we built payment primitives so buyers can authorize agents, plus fraud tooling for the AI company.

為了信任，我們做了支付原語，讓買家能授權 agent，並為 AI 公司提供詐欺防護工具。

So, four inhibitors that used to cap how fast a business could grow — cost to build, cost to go global, ability to price value, ability to discover products.

所以，四道過去限制企業成長速度的阻礙——打造成本、全球化成本、按價值定價、產品被發現。

They may seem like separate trends, but they compound — a solopreneur can go from idea to revenue as fast as possible.

它們看似各自獨立，其實互相複利——一個獨立創業者能以最快速度從點子走到營收。

At Stripe we're building the economic infrastructure for AI. A year from now, someone in this room will have built a billion-dollar company that couldn't have existed two years ago.

在 Stripe，我們在打造 AI 的經濟基礎建設。一年後，這房間裡會有人做出一家兩年前根本不可能存在、營收破十億的公司。

The inhibitors are gone. We welcome you to build. Come find us at our booth. Thank you, and enjoy SuperAI.

阻礙已經消失。歡迎你來打造。來我們的攤位找我們。謝謝，祝你享受 SuperAI。

Our next speaker was part of a team that could not wait. Dr. Andy Hock is the Chief Strategy Officer and SVP at Cerebras Systems.

我們的下一位講者，來自一個「等不及」的團隊。Dr. Andy Hock 是 Cerebras Systems 的策略長暨資深副總裁。

His PhD was in geospatial engineering; he specialized in using AI to analyze satellite imagery, including pioneering work with companies like Google.

他的博士是地理空間工程；專長是用 AI 分析衛星影像，包括與 Google 等公司的開創性合作。

Nine years ago, they were waiting for computers to process their imagery — and they could not wait. So they set about designing a better, faster, more efficient chip.

九年前，他們在等電腦處理他們的影像——但他們等不下去。於是他們著手設計一顆更好、更快、更高效的晶片。

That achievement was rewarded last month: their IPO raised $5.5 billion, the share price jumped 68% on day one, valuing the company at almost $100 billion — the biggest IPO so far this year.

這份成就上月得到回報：他們的 IPO 募資 55 億美元，首日股價跳漲 68%，公司估值近 1,000 億美元——今年至今最大的 IPO。

Please welcome Dr. Andy Hock.

讓我們歡迎 Dr. Andy Hock。

Wow, what a setup. Good afternoon, everybody. Welcome to SuperAI. I'm thrilled to see such a great turnout — and respect to those who chose the beanbag seating.

哇，這場面真不錯。午安，各位。歡迎來到 SuperAI。看到這麼踴躍的出席我很興奮——也要向選了懶骨頭座位的人致敬。

We are in an extraordinary moment in AI. Five or ten years ago, AI was a curiosity.

我們正處在 AI 的非常時刻。五到十年前，AI 還只是個有趣的好奇心。

Then came the proverbial ChatGPT moment, which brought AI into public discourse and our conversations with family.

接著是眾所周知的 ChatGPT 時刻，把 AI 帶進公共討論、也帶進我們與家人的對話。

More than that, AI switched in that moment from being interesting to being valuable.

更重要的是，AI 在那一刻從「有趣」變成了「有價值」。

And just a few years later, AI is not just demonstrably valuable — it's existential.

而短短幾年後的今天，AI 不只是明顯有價值——它是存亡攸關的。

We're building AI systems that don't just process imagery; they model language and other sequence problems for science and medicine.

我們正在打造的 AI 系統，不只處理影像；它們為科學與醫學建模語言與其他序列問題。

That leads into my talk: the imperative for faster inference, and enabling the next generation of agentic AI.

這帶出我今天的主題：更快推論的迫切性，以及如何賦能下一代的 agentic AI。

AI has transformative potential — you wouldn't be building what you're building if it didn't, not just for industry, but for science and society.

AI 有變革性的潛力——若沒有，你不會在打造你現在打造的東西，不只為產業，也為科學與社會。

One challenge is that our ability to innovate is speed-limited by legacy computing systems and chip architectures.

一個挑戰是：我們的創新能力，被 legacy 運算系統與晶片架構「限速」了。

It can take weeks or months to train models, and inference can take minutes or longer to respond. In the internet era, every millisecond matters. This is too slow.

訓練模型要花數週或數月，推論有時要數分鐘以上才回應。在網路時代，每一毫秒都重要。這太慢了。

That's our fundamental thesis at Cerebras: legacy chips are incredible, but they were built for a different purpose. They can't keep up with AI, so we need a new computing platform.

這就是我們在 Cerebras 的核心命題：legacy 晶片很厲害，但它們是為別的目的而造的。它們跟不上 AI，所以我們需要全新的運算平台。

As promised, this is a show-and-tell talk. The chip we built we call the Wafer Scale Engine — the largest computer chip in history, more than 50 times bigger than any built before.

如承諾，這是場「展示」演講。我們造的晶片叫 Wafer Scale Engine——史上最大的電腦晶片，比以往任何晶片大 50 倍以上。

You can see it on screen, to scale, next to NVIDIA's latest B200. This generation has 900,000 cores, designed for the sparse linear-algebra operations common to AI training and inference.

你能在螢幕上看到它，按比例放在 NVIDIA 最新的 B200 旁邊。這一代有 90 萬個核心，為 AI 訓練與推論常見的稀疏線性代數運算而設計。

All 900,000 cores are directly connected over silicon, and all have access to ultra-fast on-chip SRAM. Think of it as a cluster's worth of AI compute on one device.

這 90 萬個核心都在矽上直接互連，並都能存取超快的片上 SRAM。把它想成「一整個叢集的 AI 算力在單一顆晶片上」。

We didn't build big because big looks cool. We built big because AI wants big chips.

我們造大晶片，不是因為大看起來酷。我們造大，是因為 AI 想要大晶片。

AI needs massive compute, that compute close together with high communication bandwidth, and fast access to memory. Large chips optimize for those attributes.

AI 需要龐大算力、算力要靠得近且有高通訊頻寬、還要能快速存取記憶體。大晶片正好為這些本質而優化。

Once we built that engine, we couldn't just drop a race-car engine into a Pinto. We rethought the system level — power, cooling, data — and made it slip into standard data-center infrastructure, on-prem or cloud.

造出那顆引擎後，我們不能把賽車引擎硬塞進一台 Pinto。我們重新設計了系統層——供電、散熱、餵資料——並讓它能塞進標準資料中心基礎建設，本地或雲端皆可。

This is our flagship system — about the size of a hotel-room refrigerator, fitting squarely into standard racks.

這是我們的旗艦系統——約一台旅館房間冰箱的大小，正好塞進標準機櫃。

We began 10 years ago delivering one or two at a time. Now we build clusters of tens or hundreds every month, equivalent to hundreds or thousands of GPUs.

我們 10 年前一次交付一兩台。現在每個月就建由數十到數百台組成的叢集，相當於數百到數千顆 GPU。

Let's talk about what it means for the application. Here's a video of Cerebras-powered inference of a state-of-the-art model — Meta's Llama 4 Maverick.

來談談這對應用的意義。這是一段 Cerebras 驅動的推論影片，模型是最先進的——Meta 的 Llama 4 Maverick。

We prompt it to implement the Tetris video game in Python, and show the Cerebras response next to the GPU response.

我們要它用 Python 實作俄羅斯方塊，並把 Cerebras 的回應放在 GPU 回應旁邊。

Implement Tetris in Python — Cerebras on the left. It's an incredible result. It's not magic, not numerics, not a different model — it's the power of the right architecture under the hood.

用 Python 實作 Tetris——左邊是 Cerebras。這結果令人驚嘆。不是魔法、不是數值技巧、不是換了模型——就是底層用對了架構的威力。

We chose a coding example, but it's not just coding, and not just one model — though we have a great partnership with Meta.

我們選了一個 coding 範例，但不只是 coding，也不只一個模型——儘管我們和 Meta 有很棒的合作。

Here's a selection of models benchmarked by Artificial Analysis against the fastest GPU implementations. The y-axis is output speed in tokens per second.

這是 Artificial Analysis 評測的一組模型，對比市場上最快的 GPU 實作。Y 軸是輸出速度，單位是每秒 token。

Output speed translates to how interactive your model can be — how quickly it answers — whether you're building a chatbot, an agentic system, or clinical assistance in healthcare.

輸出速度決定你的模型能多互動——回答多快——不論你在做聊天機器人、agentic 系統，還是醫療臨床輔助。

Speed doesn't just matter for faster chat — speed is a requirement for future AI systems using reasoning models and agentic workflows.

速度不只關乎更快的聊天——速度是未來使用 reasoning 模型與 agentic 工作流的 AI 系統的必需品。

Take single-shot inference, like a model answering one question: one inference computation. Call it 1x compute, and one dollar-sign of value.

以單次推論為例，像模型回答一個問題：一次推論運算。稱它 1× 算力、一個「$」的價值。

More sophisticated reasoning models, when asked one question, reason with themselves — building a plan, breaking down the problem, solving each part, checking the answer. That's maybe 100x more compute per query, but significantly more valuable.

更精密的 reasoning 模型，被問一題時會自己跟自己推理——擬計畫、拆解問題、逐步解、檢查答案。那大概是每次查詢 100× 的算力，但價值也顯著更高。

It gets more interesting with agentic workflows, where one query spawns multiple daughter agents, each doing multiple reasoning steps.

agentic 工作流更有意思：一個查詢衍生出多個子 agent，每個又各做多步推理。

These are what much of this community is building — not one model doing one thing, but a fleet of agents planning, coding, building, checking, working with each other.

這正是這個社群許多人在打造的——不是一個模型做一件事，而是一隊 agent 在規劃、coding、建造、檢查、彼此協作。

That can be maybe a thousand times more inference compute than a single run, and significantly more valuable. For these to be truly interactive, they need fundamentally faster computing.

那可能是單次推論的約一千倍算力，而且價值顯著更高。要讓這些應用真正能互動，就需要根本上更快的運算。

Those applications are driving users to our platform and setting the market on fire.

正是這些應用把用戶推向我們的平台，並讓市場火熱起來。

So in this age of inference, speed isn't nice to have — it's required. Speed equals interactivity, faster time to insight, capability, and differentiation. Speed is really intelligence.

所以在這個推論的時代，速度不是 nice-to-have——是必需。速度等於互動性、更快得到洞見、能力與差異化。速度其實就是智慧。

We're not the only ones who think this. Sundar said one of their biggest challenges in building AI systems is latency — and nobody knows this better than Google.

不是只有我們這麼想。Sundar 說過，他們建 AI 系統最大的挑戰之一就是延遲——而沒人比 Google 更懂這點。

He's the Chief Strategy Officer and co-founder of Planet Labs, with the perfect background for it.

他是 Planet Labs 的策略長暨共同創辦人，擁有最契合的背景。

He was like a comet circling NASA, involved in geopolitical projects and space-technology projects, including the search for exoplanets.

他像一顆繞著 NASA 運行的彗星，參與地緣政治專案與太空科技專案，包括系外行星的搜尋。

In 2011, he and his colleagues created Planet Labs. Please welcome Robbie Schingler.

2011 年，他與同事創立了 Planet Labs。讓我們歡迎 Robbie Schingler。

It's good to be here. Andy, who was just on stage, is also from the geospatial community — we've crossed paths over the years. Fitting, because today is about space and AI coming together.

很高興來到這裡。剛在台上的 Andy 也來自地理空間社群——這些年我們多次交會。這很應景，因為今天要談的正是太空與 AI 的結合。

The most important data is not in a big foundation model on a server in Silicon Valley. It's data collected just 200 kilometers from here.

最重要的資料，不在矽谷某台伺服器上的大型基礎模型裡。它是在離這裡僅 200 公里處剛採集到的資料。

This region has long had contested facts — gray-zone areas of oceans and forests, where one nation sees a fishing fleet and another sees a strategic blockade.

這個區域長期充滿「爭議事實」——海洋與森林的灰色地帶，一國看到的是漁船，另一國看到的是戰略封鎖。

Or one company sees a sustainable agroforestry project, and a government agency sees illicit activities.

或一家公司看到永續造林專案，政府機構卻看到非法活動。

At Planet we operate the world's largest constellation of imaging satellites, with a simple theory of change: transparency brings accountability, which leads to better decisions.

在 Planet，我們營運全球最大的成像衛星星系，秉持一個簡單的「改變理論」：透明帶來問責，問責帶來更好的決策。

In the era of AI, we can finally close the gap between seeing something when it's too late, and anticipating what's about to happen and being guided to act proactively.

在 AI 時代，我們終於能弭平「看到時已來不及」與「預判即將發生、並被引導提前行動」之間的落差。

I take pictures in space. Imagine I hand you 10,000 pictures of every building in Sydney — could you immediately tell if there's roof damage or recently installed solar panels?

我在太空拍照。想像我給你雪梨每棟建築的 1 萬張照片——你能立刻看出哪裡屋頂受損、哪裡剛裝了太陽能板嗎？

Now imagine doing that for every city across Asia Pacific — not just today, but tomorrow, and the next day. That's the challenge with satellite data: too much data, not enough resources to comb through it.

現在想像對全亞太每座城市都這麼做——不只今天，還有明天、後天。這就是衛星資料的挑戰：資料太多、人力不足以一一爬梳。

At Planet we have hundreds of satellites constantly monitoring the planet — four types. One images the whole world every day at 10:30 in the morning. We've done that for nine years. That's a lot of data.

在 Planet，我們有數百顆衛星持續監測地球——共四種。其中一種每天上午 10:30 拍遍全球。我們這麼做已經九年了。那是非常龐大的資料。

Traditionally, a satellite gives a static picture, and an expert manually searches to understand what, where, and why.

傳統上，衛星給你一張靜態照片，再由專家人工搜尋去理解「是什麼、在哪裡、為何發生」。

To fix this, we've activated all this data to the cloud to make it indexable and searchable. Now we scan the whole Earth every day, and AI automatically indexes the globe — what's happening and where.

為解決這點，我們把所有資料上雲，讓它可索引、可搜尋。現在我們每天掃描整顆地球，AI 自動為全球編索引——什麼事正在發生、在哪裡。

Users set thresholds, alerts, and custom dashboards to digitize workflows in their companies, governments, and communities.

使用者設定門檻、警報與自訂儀表板，把公司、政府與社群的工作流數位化。

We've made this data visible and accessible — but AI makes it actionable.

我們讓這些資料變得可見、可存取——而 AI 讓它變得可行動。

In natural language, you can instantly learn about infrastructure buildup in China — tracking the footprint of buildings, roads, and power infrastructure as it scales, and how it's changed over seven years.

用自然語言，你能即時了解中國的基建擴張——追蹤建物、道路、電力設施的足跡如何擴張，以及過去七年如何變化。

Shift to agriculture: the tool can instantly compare crop health in the Mekong Delta against its five-year average, and you can talk to it to understand drought risk or productivity.

轉到農業：這工具能即時比對湄公河三角洲的作物健康與五年均值，你能用對話理解乾旱風險或產量。

For defense and regional security, we can ask the tool to detect any new infrastructure along China's border, and the AI instantly draws bounding boxes around new structures in the Himalayas.

在國防與區域安全上，我們能要這工具偵測中國邊境的任何新建設施，AI 即時在喜馬拉雅的新結構周圍框出邊界框。

What you're witnessing is natural language of the physical planet. It democratizes access to what's happening on our changing planet — you don't need your own staff of geospatial experts, you just need to know what questions to ask.

你看到的是「實體地球的自然語言」。它民主化了「取得星球變化情報」的能力——你不再需要自己養一隊地理空間專家，只要知道該問什麼問題。

This is in private beta, built over the last few months — that's how fast these tools are coming. You can go to ai.planet.com to join. This is a glimpse of what we call planetary intelligence.

這還在私有測試，是過去幾個月才建起來的——這些工具來得就是這麼快。你可以上 ai.planet.com 加入。這是我們所謂「planetary intelligence」的一瞥。

Right now, most major foundation models are highly human-centric — trained on text written by humans, social-media and Reddit posts. They're blind to the physical reality of what's happening on our planet every day.

如今多數主流基礎模型高度以人為中心——訓練自人寫的文字、社群與 Reddit 貼文。它們對「地球每天的物理現實」是盲的。

We need a different kind of model — a large Earth model — one that captures the real-time facts of our physical planet and speaks the language of our planetary systems.

我們需要一種不同的模型——大地球模型——能捕捉星球的即時事實、會說地球系統的語言。

When you combine a large Earth model and a large language model, AI can drive planetary-informed decisions. All this data sits next to a model context protocol, so you can build your own dashboards in minutes.

當你把大地球模型與大語言模型結合，AI 就能驅動「以星球為依據」的決策。這些資料都搭配 MCP（model context protocol），你能在幾分鐘內建好自己的儀表板。

Let's go to the Xiaowan Dam in the upper Mekong basin. Six nations rely on water through this dam for agriculture and energy, so monitoring upstream for disturbances matters.

我們來看湄公河上游的小灣水壩。六個國家仰賴流經此壩的水做農業與發電，因此監測上游的擾動很重要。

A month later, you see significant changes — sediment-heavy gray — telling a story about seasonal runoff and how dam operations are changing, with downstream impact.

一個月後，你看到顯著變化——大量泥沙的灰色——訴說著季節性逕流、以及水壩運作如何改變，並影響下游。

A month later again, the picture changes entirely: sediment settles, vegetation grows back. Three pictures across three months — now imagine doing that every day.

再一個月，畫面完全不同：泥沙沉降、植被回長。三個月三張照片——現在想像每天都做。

Our planet has a pulse. Detecting anomalous activity before major impact gives indication and warning — so we're not surprised, and nations can act diplomatically before conflict occurs.

我們的星球有脈搏。在重大衝擊前偵測到異常活動，就是「預兆與預警」——讓我們不被突襲，讓各國能在衝突發生前進行外交。

Now, how this protects people after disasters — the earthquake in Myanmar, with Microsoft's AI for Good team, whom we've partnered with for over four years since Ukraine.

再看災後如何保護人——緬甸地震，與 Microsoft 的 AI for Good 團隊合作，我們從烏克蘭起合作已逾四年。

This model for detecting damaged buildings was trained on Ukraine data; it took three months four years ago. Two years ago we deployed it for Syria's earthquake in two days. Now it happens in real time.

這個偵測受損建物的模型用烏克蘭資料訓練；四年前花了三個月。兩年前用於敘利亞地震花了兩天。現在即時完成。

It quickly understands what's damaged, how severe, and where — and we distribute this common operating picture to first-aid and disaster responders via social media, so everyone can guide their efforts in the first few hours, when you save the most lives.

它能快速判斷哪裡受損、多嚴重、在哪——我們把這份「共同作戰圖」透過社群分發給急救與救災人員，讓大家在黃金數小時內引導行動，那正是能救最多人的時刻。

In the Philippines, landslides are constant. Traditional methods took about a week to map less than half a percent of landslide risk, using data five years old — not dynamic, very manual.

在菲律賓，山崩不斷。傳統方法約花一週只能繪出不到 0.5% 的山崩風險，用的還是五年舊的資料——不即時、極度人工。

With the Tony Blair Institute and Thinking Machines, we trained a model on 27,000 landslides over 10 years across 10 critical watersheds. Now in about two hours you can cover 20% of the country, and do it again the next day.

我們與 Tony Blair Institute 及 Thinking Machines 合作，用過去 10 年、橫跨 10 個關鍵流域的 2.7 萬筆山崩訓練模型。現在約兩小時就能涵蓋 20% 的國土，隔天還能再做。

Importantly, we're not just handing them the answer — we train the government teams to modify the code so they own the architecture. It's not software; it's teaching them to use an Earth-data platform.

更重要的是，我們不只給他們答案——我們訓練政府團隊去改 code，讓他們擁有自己的架構。這不是一套軟體；這是教他們使用一個地球資料平台。

This shift is from reactive to predictive.

這個轉變，是從「反應式」走向「預測式」。

In our vision of planetary intelligence, the satellites are the eyes, but computation is the brain. That's why we're launching GPUs into space.

在我們的 planetary intelligence 願景裡，衛星是眼睛，運算才是大腦。這就是我們把 GPU 送上太空的原因。

This came down about a month ago — GPUs fast enough that data off the sensor can immediately run a processing pipeline for feature and object identification in real time.

這大約一個月前才下線——GPU 快到足以讓感測器資料立刻跑處理流程，即時做特徵與物件辨識。

Sensing the Earth continuously, so when things change and minutes matter — equaling millions of dollars or saved lives — we can act in near real time.

持續感測地球，當事情改變、分秒必爭時——可能等於數百萬美元或救命——我們能近乎即時地行動。

We're taking it further: Google has a moonshot called Project Suncatcher, putting their TPUs into space for orbital computing. Planet is proud to be their space partner; the first satellites launch early next year.

我們更進一步：Google 有個登月計畫叫 Project Suncatcher，把它們的 TPU 送上太空做軌道運算。Planet 很榮幸成為其太空夥伴；首批衛星明年初發射。

This isn't sci-fi. As launch costs fall to around $300 a kilogram in about five years, it becomes cheaper and more sustainable — no land, no power substation, no fresh water needed. Data centers in space will supercharge planetary intelligence.

這不是科幻。當發射成本五年內降到每公斤約 $300，它會更便宜也更永續——不需土地、不需變電站、不耗淡水。太空中的資料中心將為 planetary intelligence 再添動力。

Let me end more philosophically. When planetary intelligence moves toward anticipating what's about to happen — seeing beyond your field of view, looking at the world as a whole — it dissolves the separation between us and the larger living world.

讓我用較哲學的方式收尾。當 planetary intelligence 走向「預判即將發生的事」——看到視野之外、以整體看待世界——它消融了「我們」與「更大的生命世界」之間的分隔。

We're living through a time when geopolitical alliances are fracturing, and the problems we face are intensifying. But I believe planetary intelligence can transform our institutions to be purpose-built for the complexity and speed of the 21st century.

我們正活在地緣聯盟分崩離析、而面臨的問題只增不減的時代。但我相信 planetary intelligence 能改造我們的機構，使其為 21 世紀的複雜與速度而生。

We can move finance, infrastructure, coordination, and decision-making at the speed of change, to create a more resilient future — so the next climate disaster won't leave a village isolated, and the next dam break won't leave a farmer without water.

我們能讓金融、基礎建設、協調與決策以「變化的速度」運轉，打造更有韌性的未來——讓下一場氣候災難不會讓村莊孤立無援，下一次水壩潰決不會讓農民無水可用。

We gave the Earth eyes with our satellites; now we're connecting it to a mind with computing — large Earth models and real-time connectivity. But now, those of us in this room must build the body: the products, companies, and institutions brave enough to move when the mind signals.

我們用衛星給了地球眼睛；現在用運算把它接上大腦——大地球模型與即時連線。但現在，這房間裡的我們必須建造身體：那些勇於在大腦發出訊號時行動的產品、公司與機構。

I can't wait to see what we can build together. Thank you.

我等不及想看看我們能一起建出什麼。謝謝。

In 2021 he had a big idea that led to the company Exa. Last month they raised $250 million in a Series C, at a $2.2 billion valuation, because investors get the concept of perfect search.

2021 年他有個大想法，催生了 Exa。上月他們完成 2.5 億美元的 C 輪募資、估值 22 億美元，因為投資人懂「perfect search」這個概念。

You're going to hear about it now from Will Bryk of Exa.

現在就請 Exa 的 Will Bryk 來談談。

Hello, how's everybody doing? I'm Will. I'm very excited to talk about web search for AI agents, not humans. I'll say a lot of crazy things, but I promise it's all true.

大家好嗎？我是 Will。我很興奮要談「為 AI agent、而非人類打造的網路搜尋」。我會講很多瘋狂的話，但我保證全是真的。

If you take away anything from this presentation, take this. This is the number of web searches per day from humans and AI systems over the past 30 years.

如果這場你只記得一件事，就記這個。這是過去 30 年來，人類與 AI 系統每天的網路搜尋次數。

Thirty years ago there were no AIs searching. Now, in 2026, humans search about 15 billion times a day. But there's a new kid on the block — AI systems — now searching almost as much.

30 年前沒有 AI 在搜尋。如今 2026 年，人類每天約搜 150 億次。但有個新角色——AI 系統——已快追上人類。

We predict that by the end of 2026, AIs will search the web more than humans — and then grow exponentially. In a few years, AIs will search a thousand times more than humans. The entire world of search is shifting.

我們預測 2026 年底 AI 的搜尋將超越人類——之後指數成長。幾年內，AI 會比人類多搜上千倍。整個搜尋世界正在位移。

Exa is the search engine for AI agents. We organize all the world's information and provide it as an API to AI agents.

Exa 是 AI agent 的搜尋引擎。我們整理全世界的資訊，以 API 提供給 AI agent。

We serve coding agents like Cursor and Cognition — when Cursor searches the web for docs, it's using Exa under the hood. We serve go-to-market agents like HubSpot, and financial agents at banks, hedge funds, and PE firms.

我們服務 Cursor、Cognition 等 coding agent——Cursor 上網查文件時，底層就是用 Exa。我們也服務 HubSpot 等 go-to-market agent，以及銀行、避險基金、私募的金融 agent。

We serve over 5,000 companies now, from AI labs to small startups, and over 4,000 developers.

我們現在服務超過 5,000 家公司，從 AI 實驗室到小新創，還有 4,000 多位開發者。

I want to talk about the why of Exa. Exa started in 2021. We didn't start to build a search engine for AI agents — we had to solve a very important problem: perfect search.

我想談 Exa 的「為什麼」。Exa 始於 2021 年。我們起初不是要做 AI agent 的搜尋引擎——我們要解一個很重要的問題：perfect search。

Here's the problem. The internet has blogs, news, tweets, Reddit posts, product pages — the web is trillions of documents. All the world's information is right there.

問題是這樣。網路有部落格、新聞、推文、Reddit、商品頁——網路是數兆份文件。全世界的資訊都在那裡。

So we all walk around with perfect understanding of the world? Obviously not — because our information tools are limited: the news, social media, search engines. They don't give a perfect understanding.

所以我們都對世界有完美理解嗎？顯然沒有——因為我們的資訊工具很有限：新聞、社群、搜尋引擎，都給不了完美的理解。

My favorite example: search "shirts without stripes" on Google and you'll get shirts with stripes. Google wasn't built to perfectly match what you typed.

我最愛的例子：在 Google 搜「沒有條紋的襯衫」，你會得到有條紋的襯衫。Google 不是為「精準匹配你打的字」而建的。

Or: "find everyone in Singapore who works on AI search, and any blog or paper they've written." You know not to Google that. Or: "find the most important US news across all media today — what went viral?" No tool gives you that.

又或者：「找新加坡所有做 AI 搜尋的人，以及他們寫過的部落格或論文。」你知道別 Google 這個。或：「找今天全美各媒體最重要的新聞——什麼瘋傳了？」沒有工具給得了。

This is a big problem: we all walk around with an incomplete understanding of everything.

這是個大問題：我們每個人對一切都只有不完整的理解。

If our information tools don't improve while AI gets so good, we'll be misinformed, make bad individual and political decisions, and power will be centralized. I'm very worried about that world.

如果 AI 變這麼強、我們的資訊工具卻沒進步，我們會被誤導、做出糟糕的個人與政治決策、權力會集中。我很擔心那樣的世界。

I want everyone to have a complete understanding of the world, so we can avoid that future.

我希望每個人都能對世界有完整的理解，好避開那種未來。

To build a perfect search engine — a crazy idea — we made some crazy decisions. In 2021 we raised a couple million and spent half on a GPU cluster, to do research into search for as many years as it took.

要做完美搜尋引擎——一個瘋狂的點子——我們做了些瘋狂決定。2021 年募了幾百萬，一半砸在 GPU 叢集，準備花上好幾年純做搜尋研究。

We were very "bitter lesson"–pilled: neural systems plus data and compute. How do you design retrieval inspired by transformers to build really good search?

我們深信「bitter lesson」：神經系統加上資料與算力。如何受 transformer 啟發設計檢索、做出真正好的搜尋？

One of our early training runs was a breakthrough — a novel way to do retrieval no one had done before, and still hasn't. I won't tell you what it is.

我們早期一次訓練是個突破——一種前所未有、至今無人做過的檢索方法。細節我不會說。

A search engine is very complex: crawl trillions of constantly-changing documents, pre-process them, return the right ones fast. There are more space programs in the world than true search engines.

搜尋引擎極複雜：爬數兆份不斷變動的文件、預處理、快速回傳正確結果。世界上的太空計畫比真正的搜尋引擎還多。

In November 2022 we launched our first search engine on Twitter. A few weeks later, ChatGPT came out, and the world shifted — everyone asked, is there even a point to search anymore?

2022 年 11 月我們在 Twitter 上線首個搜尋引擎。幾週後 ChatGPT 出現，世界位移了——大家都問：還需要搜尋嗎？

But we quickly realized search is still really important — maybe more so now. Someone asked if we had an API. I thought, why would anyone want an API?

但我們很快意識到搜尋仍非常重要——也許現在更重要。有人問我們有沒有 API。我心想，誰會想要 API？

Then it clicked: LLMs don't know all the world's information. GPT-5 is maybe 10 terabytes; the web is millions of terabytes and always changing. It can't answer your question unless it uses a search tool.

然後就通了：LLM 不知道全世界的資訊。GPT-5 大概 10 TB；網路是數百萬 TB 且不斷變動。不用搜尋工具，它答不了你的問題。

So search is critical to AI. And weirdly, AIs need a totally different kind of search. Humans are lazy — a few keywords, a few links. AIs analyze information at superhuman speed, thousands of documents at once.

所以搜尋對 AI 至關重要。而奇妙的是，AI 需要的搜尋完全不同。人類很懶——幾個關鍵字、幾條連結。AI 以超人速度分析資訊，一次數千份文件。

What's awesome is that what we'd been building for years — perfect search — was exactly what these AIs wanted. We released the first search API for AI agents in 2023.

很棒的是，我們多年來打造的 perfect search，恰好就是這些 AI 要的。我們在 2023 年推出首個給 AI agent 的搜尋 API。

Every investor asked what that even meant. We got early customers, lived in the office. 2025 is when things picked up; now in 2026, every company wants an AI strategy. Each quarter we now make as much progress as the past five years.

每個投資人都問那到底是什麼意思。我們有了早期客戶、住在辦公室。2025 年開始起飛；如今 2026，每家公司都要 AI 策略。現在我們每一季的進展，等於過去五年。

We have the deepest search in the world. VCs use Exa for deep research — "find every startup funded by YC working on AI, with batch and status." It might take seconds, minutes, or hours.

我們有全世界最深的搜尋。VC 用 Exa 做深度研究——「找出所有 YC 投資、在做 AI 的新創，含 batch 與存續狀態。」可能花數秒、數分鐘或數小時。

We also have the fastest search — sub-200 milliseconds. If you're building a real-time voice agent, every millisecond matters; you want search that takes 150 milliseconds, not minutes.

我們也有最快的搜尋——低於 200 毫秒。如果你做即時語音 agent，每毫秒都重要；你要的是 150 毫秒、而非數分鐘的搜尋。

We help people save money amid the compute crunch: our models extract the most important tokens from each document, so the AI gets 10 documents of ~100 tokens instead of ~1,000, cutting downstream LLM costs.

在算力緊缺下我們幫人省錢：我們的模型從每份文件抽出最重要的 token，讓 AI 拿到 10 份各約 100 token、而非各約 1,000，省下游 LLM 成本。

Others want structured output, so we turn the whole web into structured info. A recruiting agent can ask for "people in Singapore who recently left their job, the company, years out of school, and a link to any paper" — returned as JSON.

有人要結構化輸出，我們就把整個網路變成結構化資訊。招募 agent 能要「新加坡剛離職的人、前公司、畢業年資、論文連結」——以 JSON 回傳。

Every customer wants something different. With 5,000 companies, we basically have 5,000 search engines — some want fast, some deep, some only certain domains, some no product pages. We built Exa to be extremely flexible. That's what perfect search is.

每個客戶要的都不同。5,000 家客戶等於 5,000 個搜尋引擎——有人要快、有人要深、有人只搜某些網域、有人要排除商品頁。我們把 Exa 做得極度彈性。這就是 perfect search。

We've come a long way, but I tell the team this is day one. The world's information is vast, and perfect retrieval is something no one has seen.

我們走了很遠，但我跟團隊說這才第一天。世界的資訊浩瀚，完美檢索是前所未見的。

We're hiring the smartest people — a head of retrieval from Meta, a head of indexing from Yandex, a research team from Google's Zurich office. I like to think of us as the Anthropic of search.

我們在招最聰明的人——來自 Meta 的檢索負責人、Yandex 的索引負責人、Google 蘇黎世的研究團隊。我喜歡把我們想成「搜尋界的 Anthropic」。

Our goal is to turn the web from a chaotic mess of documents into something that feels like a library, organized for the agent.

我們的目標是把網路從混亂的文件堆，變成感覺像「為 agent 整理好的圖書館」。

By 2027, our goal is to give people nearly complete information about the world. Perfect search is a never-ending quest, but things move fast. AI searches roughly 10x year over year.

到 2027，我們的目標是讓人對世界有近乎完整的資訊。完美搜尋是永無止境的追求，但進展飛快。AI 的搜尋約以每年 10 倍成長。

Play it out: every AI system and software product constantly searching — the world will feel infused with high-quality information, like electricity everywhere. If it's high quality, that's really important.

推演下去：每個 AI 系統與軟體產品都在不斷搜尋——世界會感覺被高品質資訊注入，就像電力無所不在。如果是高品質的，那非常重要。

I'll end with a story. Ten years ago I told my mom these neural networks were getting really good, that AI would be the most important transition in human history, and we all have to do our part.

我用一個故事收尾。十年前我跟媽媽說這些神經網路越來越強，AI 會是人類史上最重要的轉折，我們都得盡一份力。

She said, "I'm a lawyer, William — what am I supposed to do?" I called her back and said: you're a lawyer, you make society function. If all the lawyers went away, that'd be far worse than if all the software engineers did.

她說：「William，我是個律師——我能做什麼？」我回撥跟她說：你是律師，你讓社會運作。如果律師全消失，會比工程師全消失糟得多。

The point is: this is the highest-stakes moment in human history — a defining few years. If we navigate this transition well, we get flying cars and beautiful cities. If not, something really bad. Everyone can play a role.

重點是：這是人類史上最高賭注的時刻——關鍵的幾年。若我們導航得好，就有飛天車與美好城市；若不好，就會很糟。每個人都能扮演角色。

Exa's role is to keep everyone extremely informed. I promise to work as hard as I can. And my question for you is: what are you going to do? Thank you.

Exa 的角色是讓所有人保持充分的資訊。我保證會盡全力。而我對你的提問是：你要做什麼？謝謝。

Please welcome Mo Abdullah, VP of Customer Engineering at Google Cloud — previously Amazon, IBM, SaltStack — now helping customers ease into agentic AI.

讓我們歡迎 Google Cloud 客戶工程副總裁 Mo Abdullah——曾任職 Amazon、IBM、SaltStack——現在幫助客戶順利導入 agentic AI。

Thank you so much. This is the best introduction I've ever had. Being at Google, I have probably the best job in the world — I lead the technical team and we spend our days making technology happen in customers' lives. It's a privilege.

非常感謝。這是我聽過最棒的介紹。在 Google，我大概擁有全世界最好的工作——我帶領技術團隊，每天讓技術在客戶的生活中發生。這是一種榮幸。

Today's topic is one I'm really excited about. As a team, our work has led us to reflect, in this agentic era, on how much of the work around agents has corollaries to the real world of humans.

今天的主題我非常興奮。身為團隊，我們的工作讓我們在這個 agentic 時代反思：圍繞 agent 的許多做法，其實與人類的真實世界有許多對應。

Most people think of agents as something about to happen. In truth, it's already happening. A Gartner study tells us that in 2025 about 5% of enterprise systems had agents — now it's close to 40%.

多數人以為 agent 是即將到來的事。事實上，它已經在發生。Gartner 研究指出，2025 年企業系統約 5% 有 agent——現在已接近 40%。

More interesting is the forward growth. It's estimated a Fortune 500 will on average have about 150,000 agents at its disposal. The average Fortune 500 has about 60,000 employees — so you're talking about 3x-ing the workforce with PhD-level agents.

更有意思的是未來的成長。估計 Fortune 500 平均將擁有約 15 萬個 agent。一般 Fortune 500 約 6 萬名員工——所以等於用「PhD 等級」的 agent 把勞動力擴增為 3 倍。

So today's question is: how do you leverage that talented agent force effectively, for the good of employees, customers, and the business?

所以今天的問題是：如何有效駕馭這支有才華的 agent 大軍，造福員工、客戶與企業？

A lot has happened in the underlying AI layer. First, the ability to think, reason, and plan. Early LLMs tried to solve a question like "cheapest ticket to Toronto avoiding layovers" all at once, and hallucinated. Now models break the task down and reason.

底層 AI 有很多進展。第一，思考、推理、規劃的能力。早期 LLM 想一次解「最便宜又不轉機的多倫多機票」這種問題，結果幻覺。現在模型會把任務拆解、進行推理。

Second, the ability to act — more tools to interface with, taking information from within the organization and partner networks, and acting on it.

第二，行動的能力——更多可介接的工具，能取用組織內與夥伴網路的資訊並據以行動。

Third, memory. LLMs used to forget — the "Dory symptom": sorry, who are you again? Now context and memory are incredible: we can consume 10 million tokens and hold them for a long time, so I can come back a week later and remember the context.

第三，記憶。LLM 以前健忘——「Dory 症候群」：抱歉，你哪位？現在脈絡與記憶很驚人：能吃下 1,000 萬 token、保留很久，所以我一週後回來還記得脈絡。

And last, multimodality — understanding beyond text: spatial awareness, sound, audio.

最後，多模態——理解超越文字：空間感知、聲音、音訊。

This leads to the AI-native organization — the transition from LLMs to the agentic loop: take a task, break it down, execute, observe, and learn continuously.

這帶來 AI-native 組織——從 LLM 走向 agentic loop：接下任務、拆解、執行、觀察、持續學習。

Late last night we had a breakthrough, and I'm beyond excited to welcome my partner onto stage to show how they're leveraging multimodality to better people's lives — please welcome Matthias.

昨晚很晚我們有了突破，我超興奮要邀請我的夥伴上台，展示他們如何運用多模態改善人們的生活——請歡迎 Matthias。

Thanks, Mo. I'm super excited about what we launched last night. We took the sensing capabilities Google has been working on — multimodality and speech — and asked: what if we apply this to two humans with a language barrier?

謝謝 Mo。我對昨晚上線的東西超興奮。我們拿 Google 一直在做的感測能力——多模態與語音——然後問：如果把它用在「兩個有語言隔閡的人」身上呢？

Imagine you've landed in a country you've never been to, trying to get to your hotel, standing in a chaotic pickup zone. Your driver calls and speaks Vietnamese; you only speak English. We facilitate more than 10 million of these calls every month, and our traveler segment has grown 10x over three years.

想像你降落在從沒去過的國家，想去飯店，站在混亂的接送區。司機打來只講越南語；你只會英語。我們每月促成超過 1,000 萬通這類通話，旅客客群三年成長了 10 倍。

Instead of explaining, let me show you — we literally launched this last night.

與其解釋，不如直接展示——這昨晚才剛上線。

(Demo) "I'm not sure where I am — let me look around. I think I'm at the main gate, but I don't see your car." "I understand, I'm at the other pickup point — let me walk further to find your car." "Yes, I see your car, I'm coming right over."

（Demo）「我不確定我在哪——讓我看看四周。我想我在大門口，但沒看到你的車。」「我了解，我在另一個接送點——我再走過去找你的車。」「好，我看到你的車了，馬上過來。」

With Gemini 3.5 Live Translate, we can finally bridge this language barrier. This rolls out in the next half year and should solve a lot of these issues. Thank you.

有了 Gemini 3.5 Live Translate，我們終於能跨越這道語言隔閡。它會在未來半年推出，應該能解決很多這類問題。謝謝。

There's nothing more human than using AI to connect humans — that moment you land, call for help, your flight is cancelled, or your in-law speaks a different language, which is very common here in Asia.

沒有什麼比用 AI 連結人更有人性的了——那個你降落、求助、班機取消、或你的姻親說著不同語言的時刻，在亞洲這很常見。

So, how do we leverage this talented agent force? Let me flip it: what did we do well when onboarding professionals into our workforce? What principles transfer to agents?

那麼，如何駕馭這支有才華的 agent 大軍？讓我反過來想：我們把專業人才帶進職場時做對了什麼？哪些原則可轉用到 agent？

First, obviously: you give a new hire an ID. We give IDs because they let us understand who this individual is, audit what they do, what they're allowed to do, which room they can enter, how we refer to and communicate with them.

第一，顯然是：你會給新員工一個身分證。我們給 ID，是因為它讓我們知道這個人是誰、能稽核他做什麼、他被允許做什麼、能進哪個房間、我們怎麼稱呼與溝通。

This only recently happened in the world of agents. Agents behave transactionally — from booking a trip to acting on your credit card. You need to know which agent did it: was it agent 0010 or agent 007? We were first to market with agent identity.

這在 agent 世界裡是最近才有的。agent 會做交易性行為——從訂行程到動用你的信用卡。你得知道是哪個 agent 做的：是 0010 還是 007？我們是率先把 agent 身分推上市的。

After you onboard an employee and give them a badge, you put them in the employee directory. So we introduced the agent registry — the source of truth for what an agent can and cannot do.

給員工識別證後，你會把他放進員工名冊。所以我們導入了 agent registry——agent 能與不能做什麼的單一真相來源。

If an agent built on someone's desktop is unleashed into the enterprise and we don't know who it is or its function, it'll take actions we may regret.

如果有人在桌機建的 agent 被放進企業、我們卻不知道它是誰、什麼職能，它會做出我們事後懊悔的行動。

Third, the agent describes what it's good at: "I'm the analyst agent, this is what I do, this is how you talk to me." We introduced an agent gallery — self-built agents, out-of-the-box ones from tools like ServiceNow, acquired or custom ones, all in one place.

第三，agent 自述它擅長什麼：「我是分析 agent，這是我的功能，這樣跟我溝通。」我們推出 agent gallery——自建的、ServiceNow 等工具開箱即用的、外購或客製的，全在一處。

Next, you give them skills. Skills let an agent behave in a grounded way against specific tasks. For example, a marketing team's "conference kit" skill organizes speakers by availability — work that used to take people hours.

接著，你給它技能。技能讓 agent 在特定任務上「接地」地做好。例如行銷團隊的「conference kit」技能，依講者空檔安排時段——這原本要人花上好幾小時。

Then you teach it how to do the job in your company — marketing here is different from marketing there. Here's the catch: if you hired a human claims processor, you wouldn't take them to IT to be trained.

然後你教它在你公司怎麼做事——這裡的行銷跟那裡的不一樣。重點來了：如果你雇了一個真人理賠專員，你不會帶他去 IT 部門受訓。

We build agents with a tech mindset, but we're changing that: empower the business user — that claims administrator — with a no-code/low-code agent designer to drag and drop their workflow. Of our 150,000 agents, maybe 100,000 are business, not IT.

我們過去用技術思維建 agent，但我們在改變：賦能業務使用者——那位理賠專員——用低/無碼的 agent designer 拖拉定義工作流。我們的 15 萬個 agent 中，約 10 萬是業務、不是 IT。

Next, you let agents collaborate — via agent-to-agent, MCP connectors to your data, and custom agents built with ADK. You're no longer gated by a predefined workflow; one agent does a task, another orchestrates, things happen in parallel.

接著，讓 agent 協作——透過 agent-to-agent、連到資料的 MCP 連接器、用 ADK 建的客製 agent。你不再被預定義工作流綁死；一個 agent 做任務、另一個編排，事情能平行進行。

But in real life you don't just throw random employees together. For confidential or domain-contained projects, we let certain agents work within a boundary: only with these agents, only this data — so they collaborate safely and securely.

但現實中你不會隨便把員工湊在一起。對機密或限定領域的專案，我們讓某些 agent 在邊界內運作：只跟這些 agent、只用這些資料——讓它們安全地協作。

I predict SuperAI becomes "Super Agent" next year, because people are codifying their industry expertise into agents they can sell. If I'm the best at import-export clearances into Vietnam, can I codify that into an agent and move it into the digital economy?

我預測明年 SuperAI 會變「Super Agent」，因為人們正把產業專業 codify 成可販售的 agent。如果我最擅長越南進出口清關，我能把它變成一個 agent、推進數位經濟嗎？

Google builds its own agents — YouTube, Maps — and works with industry leaders on others, like a financial advisor or a software bug assistant. We debug software thousands of times a day; that expertise can become an agent.

Google 自建 agent——YouTube、Maps——也與產業領袖合作做其他的，像財務顧問或軟體除錯助理。我們每天 debug 軟體數千次；那份專業可以變成一個 agent。

And my favorite — the least favorite for humans — performance reviews. In an agent world you must observe it, know how it behaves, and whether it produced the outcome you needed, not just performed the task.

還有我最愛的——人類最不愛的——績效考核。在 agent 世界你必須觀察它、知道它如何運作、以及它是否產出了你需要的「結果」，而不只是執行了任務。

The summary it gave me in a second — was it good? You can write an agent that asks "was this accurate, was this good?", give feedback, and it learns. You can retire agents that misbehave — fire them, offer a severance package; I don't know what that looks like for agents yet.

它一秒就給我的摘要——好不好？你可以寫一個 agent 問「這準確嗎、好嗎？」，給回饋，它就學習。表現不好的 agent 可以退役——資遣它、給資遣費；agent 的資遣費長怎樣我還不知道。

The title is the human side of agents, but there's nothing human about agents. The basic premise of humans is trust — do I trust Laura, do I trust Jenny? We must not abdicate that trust.

標題是 agent 的人性面，但 agent 本身毫無人性。人類的根本前提是信任——我信任 Laura 嗎、信任 Jenny 嗎？我們不該把這份信任拱手讓出。

We must create a zero-trust environment so we know agents do what we expect. That's why human-in-the-loop is critical. With identity you can ask: who did this? Who bought all these TVs? Who issued these refunds?

我們必須建立 zero-trust 環境，好讓我們知道 agent 在做我們預期的事。這就是 human-in-the-loop 至關重要的原因。有了身分你才能問：是誰做的？誰買了這些電視？誰發了這些退款？

I'm a football fan — I was disappointed FIFA gave away nearly 3,000 World Cup tickets for free, and it was an agent, not a Google agent. They told people to pay or give them back. I wish I were one of those 3,000 — I'd go into a seven-year lawsuit and go to the World Cup for free.

我是足球迷——我很失望 FIFA 免費送出近 3,000 張世界盃門票，而那是一個 agent 幹的，不是 Google 的。他們後來要人付錢或退票。真希望我是那 3,000 人之一——打一場七年官司，免費看世界盃。

Gateways: you want to say yes or no on a decision, and shut things down quickly across a fleet — "stop giving these discounts, kill this campaign." One agent on your laptop is fine; 3,000 working for you is different.

閘道：你要能對決策說 yes 或 no，並能跨整支艦隊快速喊停——「別再給這些折扣、關掉這檔活動」。一個 agent 在你桌機上沒差；3,000 個替你工作就不一樣了。

Protocols: humans need to understand what several agents told each other — complex even between humans. Thankfully you can write an agent to decipher and summarize it.

協定：人需要理解好幾個 agent 彼此說了什麼——這即使在人之間都很複雜。所幸你可以寫一個 agent 來解讀並摘要。

Registries let humans decide who to trust, which may change over time. And durability: what if an agent fails or two agents conflict? Does the whole process break? How do you recover?

registry 讓人決定信任誰，而這會隨時間改變。還有 durability：如果一個 agent 失效、或兩個 agent 起衝突怎麼辦？整個流程會崩嗎？怎麼復原？

This is where Google has thought deeply — what we call Gemini Enterprise. Most of the world focuses on building agents, because developers love building. But that's a small percentage.

這正是 Google 深思之處——我們稱為 Gemini Enterprise。世界多數人專注於「建」agent，因為開發者愛建東西。但那只是一小部分。

It becomes enterprise-ready when you know how to scale it, govern it, and optimize it — because agents use resources underneath, and they aren't cheap if they run rogue, buying 1,000 mouse mats and giving them away. They use tokens, so how do you optimize and cache, and stop an agent re-querying the same information?

當你懂得 scale、govern、optimize，它才 enterprise-ready——因為 agent 會吃底層資源，放任亂跑並不便宜，買 1,000 個滑鼠墊免費送。它們用 token，所以你要怎麼優化、快取，並阻止 agent 重複查同樣的資訊？

Every layer of the stack has a capability: on the build side, Antigravity, agent studios, ADK. Then an agent runtime to orchestrate, with agent sessions so you pick up where you left off using context and memory. Down to registry, agent policy, and Model Armor — models that detect threats and remove them, making it enterprise-ready.

堆疊的每一層都有對應能力：build 端有 Antigravity、agent studios、ADK。再來是編排用的 agent runtime，搭配 agent sessions，讓你用脈絡與記憶接續未完之處。一路到 registry、agent policy 與 Model Armor——用模型偵測並移除威脅，讓它 enterprise-ready。

I'll leave you with this: Google was an AI company before it was a cloud company. When your mission is to organize the world's information, you can't do it with human capability alone — so machine learning has been in our blood from the infrastructure up.

我用這句收尾：Google 在成為雲端公司之前，先是一家 AI 公司。當你的使命是組織全世界的資訊，光靠人力做不到——所以機器學習從基礎建設層起就在我們血液裡。

That's why our models continue to be the best performance-for-price in industry benchmarks, and why we can bring the multimodality you saw with Matthias. Thank you so much — it's been a pleasure.

這也是為何我們的模型在業界 benchmark 持續是最佳「性價比」，以及為何我們能帶來你看到的、Matthias 展示的多模態。非常感謝——這是我的榮幸。