Name: 想要一個私人助理嗎？人工智能可能就在不遠處 (Ever wanted a personal assistant? An AI one could be around the corner)
Uploaded: 2025-10-27T18:27:59.000Z
Duration: 34 min 56 s

Welcome to beyond the Valley, a CNBC tech podcast with Arjun Karpal and Tom Chitty.

歡迎來到 beyond the Valley，這是 CNBC 的一個 tech podcast，由 Arjun Karpal 和 Tom Chitty 主持。

Have you ever wanted a personal assistant?

你有沒有想過要一個個人助理？

情態助動詞 A2

Someone that can manage your diary, arrange travel, make notes in meetings, but also knows you personally, what your interests are, your habits, your likes and dislikes.

他能幫你管理行程、安排旅行、在會議中做筆記，還能瞭解你個人的興趣、習慣、喜好和不喜歡。

Now, what if that assistant was an AI large language model trained on your individual data?

現在，如果這個助理是一個根據你的個人資料訓練出來的 AI large language model 呢？

This is just one of the many ways that future technology could impact our lives in 2025 and beyond, according to a list of predictions that have often become a reality beyond the Valley.

根據一份常在 beyond the Valley 成真的預測清單，這只是未來科技在 2025 年及以後可能影響我們生活的方式之一。

He is chief analyst at tech research firm CCS Insight.

他是 tech research firm CCS Insight 的 chief analyst。

In 2021, he was ranked as one of the top five tech analysts in the US and Europe by Apollo Research.

2021 年，他被 Apollo Research 評為 US 和 Europe 前五名的 tech analysts 之一。

Ben is also the founder and co curator of the Mobile Phone Museum.

Ben 也是 Mobile Phone Museum 的 founder 和 co curator。

But this week we're going to discuss his work leading CCS Insight's annual predictions event.

但這週我們要討論的是他領導 CCS Insight 年度預測活動的工作。

Ben, thanks for joining us on beyond the Valley again.

Ben，謝謝你再次來到 beyond the Valley。

You are the first repeat guest we've had on the podcast.

你也是我們 podcast 第一次邀請到的回鍋 guest。

Well, I'm flattered to enter that hall of fame.

哇，能進入這個名人堂，我真是受寵若驚。

We're going to get a plaque of some sort, I imagine.

我想我們應該會得到某種獎牌吧？

We did promise maybe to give out mugs, didn't we, to some of our guests?

我們之前是不是有答應要送一些杯子給我們的 guest？

That hasn't materialized yet, but we'll work on that.

這個還沒實現，但我們會努力的。

Talk us through this event and how you come up with these predictions.

跟我們聊聊這次的活動，以及你們是怎麼做出這些預測的。

So I've been in the analyst game for quite a while and I don't know, about 15 years ago or so, there were a few companies who were doing the odd tech prediction, but I felt there was an opportunity to do it in a more structured manner.

我做分析師這行已經很久了，大概 15 年前吧，有些公司會做一些零星的科技預測，但我感覺有機會可以更有系統地來做。

We had a team of analysts and quite frankly, our job is to provide insight into where technology is going to.

我們有一個分析師團隊，說實在的，我們的職責就是提供科技發展趨勢的洞見。

And if a team of analysts can't come up with a few interesting predictions, then we're not doing our job properly.

如果一個分析師團隊都無法提出一些有趣的預測，那我們就沒有盡到我們的職責。

So we started putting together a short list of predictions about how we saw things happening in the next year in the kind of five years out.

所以我們開始整理一份預測清單，關於我們預見未來一年，以及未來五年的發展。

So this year, for example, would be 2025 and beyond.

所以今年的話，例如是 2025 年及以後。

But the charter for the predictions was they have to be definitive.

但預測的宗旨是，它們必須是明確的。

It can't be kind of like might, could, would.

不能是那種「或許」、「可能」、「會」。

I like people to actually put it on the line and actually commit to something.

我希望大家能真正放手一搏，確實地做出承諾。

We don't do stuff that's deliberately provocative.

I think that's a bit of a fool's errand, but we will put things in there which might raise people's eyebrows because if you just go for low hanging fruit and say, you know, the sun will come up and grass will be green if it rains and those sorts of things, it's a waste of time.

我認為那有點像徒勞無功，但我們會放進一些可能會讓大家跌破眼鏡的內容，因為如果你只挑那些顯而易見的，說「你知道，太陽會升起，如果下雨草就會變綠」之類的事情，那就太浪費時間了。

You have to have stuff which is potentially a little bit more edgy and that means you will get things wrong.

你必須要有那種可能有點比較前衛的內容，這意味著你也會有預測錯的時候。

And then the final thing that we put into it was the fact that we, we're always happy to after we'd done our predictions the year later or a couple of years later, we'd go back and reflect on them and say, here's some we got right, but more importantly, here's the ones we got wrong.

最後我們放進去的一點是，我們總是樂於在做出預測後，隔一年或幾年後，回頭檢視並說，這裡有些我們預測對了，但更重要的是，這裡有些我們預測錯了。

Well, we're going to get into those later in the episode, but first let's talk about the prediction that we're most interested and explore it a little deeper before we do that though, Stat of the Week Ben, you played last time, I can't actually remember the result of I got in the ballpark, but I didn't get it exactly right.

嗯，我們稍後會在節目中討論這些，但首先，讓我們來談談我們最感興趣的預測，並深入探討一下，不過在那之前，本週統計數據，Ben，上次是你講的，我不太記得結果了，我猜得差不多，但沒有完全正確。

Yeah, I think you were really far off as well, Tom, as yeah, as usual.

對啊，湯姆，我想你上次也猜得離譜，一如往常。

And remember a few episodes ago, just want to make this very clear.

還記得幾集前，我想讓大家很清楚一點。

Well, for looking at Google during the salad.

嗯，因為你看了 Google 的沙拉。

I think it was reviewed and inconclusive.

我覺得這是有爭議的，結果也不確定。

There's still an inquiry going on as we speak.

But the start of the week this week, 990 billion.

All right, so the prediction that we're going to discuss in the main for this episode is this.

好的，那我們這集要討論的主要預測是這個。

So the first personal large language model for AI arrives in 2025.

第一款個人 large language model for AI 將在 2025 年問世。

A personal large language model is trained using only data specific to an individual user.

個人 large language model 是使用僅限於個別使用者資料來訓練的。

In contrast to a standard model, it is tuned specifically to a person's communication style, name, needs and preferences.

相較於標準模型，它會特別針對個人的溝通風格、姓名、需求和偏好進行調整。

It is trained using the individual's data based on previous interactions, such as emails, messages, documents and social media interactions.

它是根據過去的互動，像是電子郵件、訊息、文件和社群媒體互動，使用個人資料來訓練的。

The model is continuously updated by drawing on the person's evolving data sources.

這個模型會透過擷取個人不斷變化的資料來源來持續更新。

So the thesis behind this prediction is the fact that we've seen AI becoming more prominent.

所以這個預測背後的論點，就是我們看到 AI 越來越重要。

We're certainly seeing it becoming very much an on device phenomenon centered around the smartphone and we're seeing the addition of things like contextual awareness.

我們當然也看到它正成為一個以智慧型手機為中心的裝置上現象，並且加入了像是情境感知之類的功能。

So if there's something on the screen, the AI model can look at it and it can help you provide information on that.

所以如果畫面上出現了什麼，AI 模型就能看到它，並協助你提供相關資訊。

Something that we're intrigued about and excited about is the idea of taking all of the personal data that people have, and of course with all of the Uzi caveats around privacy, and using that to, to create a model which is more tailored to that specific individual.

讓我們感到好奇且興奮的是，能夠取得人們擁有的所有個人資料，當然也包含所有關於隱私的注意事項，並利用這些資料來建立一個更貼近特定個人的模型。

And if there's one company that's well positioned to do this, it's Apple, because of the unbelievable repository of personal data that people tend to have on an iPhone or in icloud, because Apple have been fantastically good at taking people on a journey of keeping all of their data from one generation of a device to another.

如果有一家公司能做好這件事，那就是 Apple，因為人們傾向於在 iPhone 或 iCloud 上儲存大量的個人資料，而 Apple 在協助人們將資料從一代裝置轉移到下一代裝置方面做得非常出色。

And what this means is, is you would then be able to get some amazing results out of this kind of model.

這意味著，你將能從這種模型中獲得一些驚人的成果。

So just take for example, some of these scenarios we see about asking AI to create a holiday for me or those sorts of things.

舉例來說，我們看到一些關於要求 AI 為我規劃假期的情境。

It's still relying somewhat on some sort of general algorithms around what people like you like.

它仍然在某種程度上依賴一些普遍的演算法，來瞭解像你一樣的人喜歡什麼。

But this would take it right down to that specific view in terms of where you've been in the past, what you like, perhaps doing analysis of pictures you've taken so you can tell who you like, going with what things.

但這將會進一步深入到你的個人觀點，例如你過去去過哪裡、你喜歡什麼，甚至分析你拍的照片，以便了解你喜歡和誰一起做什麼。

It could be a remarkable way of delivering context and for other things, just prioritizing important things in your life.

這可能是一種傳達脈絡的絕佳方式，也能幫你釐清生活中重要的事物。

If things are appearing in your email that maybe haven't been there for a while, but they all of a sudden are super hyper relevant and you need to know about them, it could surface those better than a general model.

如果你的電子郵件裡出現一些許久未見但突然變得超級相關、你必須知道的內容，它就能比一般模型更好地將這些資訊呈現出來。

So this would be something you see working on devices at this point, whether it's a smartphone or a tablet, laptop, etc.

所以這會是你在裝置上看到的功能，像是智慧型手機、平板電腦、筆記型電腦等等。

So in the holiday example, that could be something like scraping sort of your, your photo day, saying, well, you've got a lot of pictures of food when you're on holiday.

所以以規劃假期為例，它可能會像這樣，抓取你的照片資料，說「嗯，你假期時有很多食物的照片。」

Specifically you're really into food when you go abroad.

特別是你出國時非常熱衷於美食。

Maybe we should throw up some good restaurant recommendations for you.

也許我們可以推薦一些不錯的餐廳給你。

When you go on holiday, there's a lot of beaches you like.

你度假時，有很多海灘是你喜歡的。

A beach holiday, is it, is it that kind of way that it will work?

海灘假期，它會是這樣運作的嗎？

And then down to all of the individual things.

You know, looking in your email to see what hotel reservations you've made and making sure that it's the kind of profile of hotel that you like, looking at the kind of experiences that you like.

你知道，看看你的電子郵件，找出你預訂了哪些飯店，確保它是你喜歡的那種飯店類型，看看你喜歡的旅遊體驗。

You know, make no mistake, we're seeing with AI already you can get some general information in this context and it will draw on some of your personal information already.

你知道，別搞錯了，我們已經看到 AI 可以提供一些這方面的通用資訊，而且它會利用你的一些個人資訊。

And the reason it needs to be on device is because it means that it's ring fenced.

而它需要在裝置上運作的原因是，這代表它被隔離了。

From a privacy perspective, you don't want that deepest personal data all being relinquished up into the cloud and then to goodness knows who.

從隱私的角度來看，你不會希望你最私密的個人資料全部被釋放到雲端，然後被不知道是誰存取。

So can we just talk about that from a technical perspective on how it will work?

所以，從技術角度來看，我們可以談談它是如何運作的嗎？

Because we did an episode recently on large language models and we're talking about all of these large language models being trained via these huge, you know, GPU chipsets, data servers, etc.

因為我們最近有一集節目講到 large language models，我們談到所有這些 large language models 都是透過這些龐大的 GPU 晶片組、資料伺服器等進行訓練的。

And then obviously in the cloud you can access the services that are a result of them, Whether it be ChatGPT or something else, so that we understand.

然後很明顯地，在雲端你可以存取它們所產生的服務，無論是 ChatGPT 或其他東西，這樣我們就能理解了。

And when we're accessing some of the current smartphone digital assistants, whether it be, you know, Google, Gemini, Samsung's Bixby, or the Apple Siri and Apple intelligence, when it comes again, it's not fully on device yet.

而且當我們使用現在的智慧型手機數位助理時，像是 Google 的 Gemini、三星的 Bixby，或是 Apple 的 Siri 和 Apple intelligence，它們目前還不是完全在裝置上運作。

A lot of it will still be coming from the cloud.

So when it comes to a sort of personalized LLM, how would one train or how would a company train that and then sort of deploy that to the user?

所以，關於個人化的 large language model，要怎麼訓練，或是公司要怎麼訓練然後部署給使用者呢？

You know, we obviously continue to see the processing power on devices increasing.

你知道，我們當然持續看到裝置的處理能力在增加。

We're seeing the size of the models growing.

我們看到模型的規模也在變大。

So the latest Gemini nano model, 3.8 trillion parameters.

所以最新的 Gemini nano 模型，有 3.8 兆個參數。

I mean, it's staggering, staggering numbers.

我是說，這數字真是驚人，驚人的數字。