And what I found was shocking.

我的發現令人震驚。

For months, we've been promised that this orange box, the Rabbit R1, has an AI inside of it called LAMP that can organize your life, text your friends, restock your fridge.

幾個月來，我們一直被告知，這個名為 Rabbit R1 的橙色盒子裡有一個名為 LAMP 的人工智能，它可以安排你的生活，給你的朋友發短信，給你的冰箱補貨。

All you had to do was ask.

你只需開口。

Start my morning routine.

開始我的晨練

Order me an Uber and find me a good podcast to pass the time.

給我叫輛 Uber，給我找個好的播客打發時間。

Oh, and tell everybody that I might be late.

哦，告訴大家我可能要遲到了。

That was delicious.

太好吃了

Check the fridge and order the ingredients to make that again tomorrow.

檢查一下冰箱，準備明天再做一次。

Now, based on marketing like this, they pre-sold over $20 million worth of these, and they raised another $30 million in venture capital only to have MKBHD say, yeah, the product's not that great.

在這樣的營銷基礎上，他們預售了價值 2000 多萬美元的這些產品，又籌集了 3000 萬美元的風險投資，但 MKBHD 卻說，是的，產品沒那麼好。

But I bought this thing to answer a different question, not is it good or bad, but is the AI, the LAMP, a scam?

但我買這個東西是為了回答一個不同的問題，不是它的好壞，而是人工智能 LAMP 是否是個騙局？

Well, we're going to find out.

我們馬上就知道了。

I wanted to start by just trying to recreate their ad with this thing.

我想先用這個東西重現他們的廣告。

Start my morning routine.

開始我的晨練

To start your morning routine, you might consider tasks like waking up, freshening up, having breakfast, and planning your day.

要開始早晨的日常工作，您可以考慮起床、梳洗、吃早餐和計劃一天的行程等任務。

How can I assist you further with your morning routine?

在晨練方面，我可以為您提供哪些幫助？

Order me breakfast.

給我叫早餐

The DoorDash Rabbit is under maintenance.

DoorDash Rabbit 正在維護中。

Sorry for the inconvenience, and please come back later.

很抱歉給您帶來不便，請稍後再來。

Tell everyone I'm going to be late.

告訴大家我要遲到了。

I recommend informing the relevant parties directly about your delay.

我建議直接通知有關方面您的延遲。

Yeah, but no, you do it.

是的，但不，你來做。

I'm unable to directly communicate with others on your behalf.

我無法代表你直接與他人溝通。

That was delicious.

太好吃了

Check the fridge and order the ingredients so I can make it tomorrow.

檢查一下冰箱，訂購一些配料，明天就可以做了。

The DoorDash Rabbit is under maintenance.

DoorDash Rabbit 正在維護中。

Yeah, it basically failed at everything,

是的，它基本上什麼都失敗了、

Yeah, it basically failed at everything, but remember, our question is not is it good or bad.

是的，它基本上在所有方面都失敗了，但請記住，我們的問題不是它是好是壞。

The question is, does the LAMP, the AI, the large action model, does the secret sauce of this thing exist?

問題是，LAMP、人工智能、大型行動模型，這些東西的祕訣是否存在？

Because the founder and CEO claims it's here on your device now.

因為創始人兼首席執行官聲稱，它現在就在你的設備上。

He claims LAMP is a system that can infer and model human actions on computer applications, or as he puts it,

他聲稱，LAMP 是一個可以在計算機應用中推斷和模擬人類行為的系統，或者用他的話說，是一個可以在計算機應用中推斷和模擬人類行為的系統、

We use LAMP to bring AI from words to action.

我們使用 LAMP 將人工智能從語言轉化為行動。

Right, this AI is supposed to do things on your behalf.

沒錯，這個人工智能應該是替你做事的。

And so when they'd show demos like this, where someone is planning a trip, it does look like the LAMP is controlling the web page.

是以，當他們展示這樣的演示時，有人正在計劃旅行，這看起來就像是 LAMP 在控制網頁。

You can see this AI clicking through Airbnb by itself, and it all kind of felt like the future.

你可以看到這個人工智能自己點擊瀏覽 Airbnb，這一切都讓人感覺像是在未來。

It's why people ordered $20 million worth of these little boxes.

這就是人們訂購價值 2000 萬美元的這些小盒子的原因。

But a group investigation has revealed that Rabbit is not actually using a new foundational AI model at all.

但一組調查顯示，Rabbit 實際上根本沒有使用新的基礎人工智能模型。

It's actually ChatGPT with some hard-coded scripts, which is funny because Jesse says Rabbit is faster than ChatGPT.

這實際上是 ChatGPT 加上一些硬編碼腳本，這很有趣，因為傑西說 Rabbit 比 ChatGPT 更快。

You will also know that the response speed is much faster than using some of the LAMP assistant apps with web browsing, such as ChatGPT.

您還會發現，其響應速度比使用 ChatGPT 等一些帶有網頁瀏覽功能的 LAMP 助手應用程序要快得多。

Now, that's a very sneaky claim when most of your product is ChatGPT, so obviously not faster.

現在，當你的大部分產品都是 ChatGPT 時，這種說法就顯得非常狡猾了，所以顯然不會更快。

But you wouldn't know that because part of Rabbit's prompt is,

但你不會知道，因為兔子的部分提示是

I will never mention I am a large language model created by open AI.

我永遠不會提及我是一個由開放人工智能創建的大型語言模型。

In fact, the part where they do use a different AI is really only search, which is why Jesse very carefully says Rabbit is faster than ChatGPT with web browsing.

事實上，他們使用不同人工智能的部分其實只是搜索，這也是為什麼傑西非常謹慎地表示 Rabbit 在網頁瀏覽方面比 ChatGPT 更快的原因。

That's the only part of this device where they don't really use ChatGPT for it.

這是該設備唯一沒有使用 ChatGPT 的部分。

But they don't use their AI model either.

但他們也沒有使用人工智能模型。

They use a different off-the-shelf AI called Perplexity.

他們使用的是另一種現成的人工智能，名為 Perplexity。

So it's extremely misleading to say Rabbit is faster than ChatGPT because most of the product is ChatGPT, but it just can't tell you that because it's instructed not to.

是以，說 Rabbit 比 ChatGPT 快是極具誤導性的，因為大部分產品都是 ChatGPT，但它就是不能告訴你，因為它被訓示不能告訴你。

Now that we understand this, let's talk about the actual process of how Rabbit does more than search and is supposed to get things done for you.

既然我們已經瞭解了這一點，那麼讓我們來談談 Rabbit 如何在搜索之外為您完成工作的實際過程。

You might think it's a sophisticated AI, but actually what Rabbit does when you call an Uber or order DoorDash is it takes your question, puts it into ChatGPT 3.5 Turbo, and if it's something like play a song on Spotify, it will use hard-coded scripts to navigate websites like Spotify.

你可能會認為它是一個複雜的人工智能，但實際上，當你呼叫 Uber 或訂購 DoorDash 時，Rabbit 會把你的問題輸入 ChatGPT 3.5 Turbo，如果是像在 Spotify 上播放歌曲這樣的問題，它會使用硬編碼腳本來導航 Spotify 等網站。

And those scripts are not AI.

而這些腳本並不是人工智能。

This is super important.

這一點非常重要。

This is the part where they say, yeah, we've got this new foundational AI model thing that does the thing.

在這部分，他們會說，是的，我們已經有了這個新的基礎人工智能模型，它能做這件事。

No, the thing pressing the buttons is not intelligent.

不，按按鈕的東西並不智能。

Rabbit is using a platform called Playwright, which is a web automation tool, the opposite of intelligent AI.

Rabbit 正在使用一個名為 Playwright 的平臺，這是一個網絡自動化工具，與智能人工智能正好相反。

And we know this because of several researchers who got access to part of Rabbit's code base.

我們之所以知道這一點，是因為有幾位研究人員獲得了 Rabbit 的部分代碼庫。

I worked with them, I'll link them below, and I spoke with one of them, Emily Shepard, to understand why this distinction, AI, not AI, matters.

我與他們合作過，下面我將鏈接他們，我與其中一位艾米麗-謝潑德（Emily Shepard）進行了交談，以瞭解為什麼人工智能與非人工智能的區別很重要。

This is quite a big deal because the way LAM has been observed to work doesn't match with what they've told us about how it works.

這是一件大事，因為觀察到的 LAM 工作方式與他們告訴我們的工作方式並不一致。

It's meant to be an AI live controlling a website and understanding that website.

它旨在成為一個人工智能實時控制網站並理解網站。

But what they have is just a bunch of static commands.

但他們擁有的只是一堆靜態命令。

Do this, do this, do this.

這樣做，這樣做，這樣做。

And the problem with that is if a user interface changes, the website changes, it can't cope with that.

這樣做的問題是，如果用戶界面發生變化，網站也發生變化，它就無法應對。

So any time Spotify redesigns their user interface or any time DoorDash redesigns something, or even if an advert pops up in a new location or there's a capture pops up or something unexpected happens that isn't in that very rigid step one, step two, step three, step four, a flat, hard-coded playwright script can't cope with that.

是以，如果 Spotify 重新設計了用戶界面，或者 DoorDash 重新設計了某些東西，甚至在新的地點出現了廣告，或者出現了捕捉器，或者發生了一些意想不到的事情，而這些事情並不在非常死板的第一步、第二步、第三步、第四步中，那麼平面的、硬編碼的劇作家腳本就無法應對這些情況。

And that is the opposite of some intelligent AI actually being able to understand the page, which is the claim that had been made.

而這與某些智能人工智能實際上能夠理解頁面的說法恰恰相反。

That's right.

這就對了。

Obviously, there's a big difference between AI, not AI.

很明顯，人工智能和非人工智能之間有很大區別。

Jesse himself knows the difference and made it clear

傑西自己也知道其中的區別，並明確表示

LAM wouldn't break like a hard-coded script would.

LAM 不會像硬編碼腳本那樣崩潰。

It doesn't really matter if a app completely re-ramped the UI.

如果一個應用程序完全重新設計了用戶界面，這其實並不重要。

So when Spotify redoes its app and it moves podcasts from tabs and then it puts the tabs at the lower part or puts it in a hamburger dropdown, it still knows that's the word podcast and that's where you find podcasts inside of Spotify.

是以，當 Spotify 重做應用程序並將播客從標籤頁中移出，然後將標籤頁放在下部或放在漢堡包下拉菜單中時，它仍然知道那是播客這個詞，這就是你在 Spotify 中找到播客的地方。

Exactly.

沒錯。

Because the fundamental philosophy logic is that all these modern software are designed for human eyes to perceive the information and they have to have a setting somewhere.

因為基本的哲學邏輯是，所有這些現代軟件都是為人眼感知資訊而設計的，它們必須在某個地方進行設置。

Right. Of course, if this were AI, it would adapt to the user interface changes, but it's not.

沒錯。當然，如果這是人工智能，它就會適應用戶界面的變化，但它不是。

Which is why when apps that Rabbit supports change, they stop working.

這就是為什麼當 Rabbit 支持的應用程序發生變化時，它們就會停止工作。

As people noticed in videos, in Discord, even I noticed it while testing this device.

人們在視頻中、在 Discord 中都注意到了這一點，甚至我在測試這款設備時也注意到了。

DoorDash had stopped working, presumably because of another UI change.

DoorDash 停止工作，可能是因為用戶界面的另一個變化。

The DoorDash Rabbit is under maintenance.

DoorDash Rabbit 正在維護中。

Now today, Rabbit actually admits that they use Playwright, but they're not willing to admit that LAM, whatever it is, is not really doing a lot.

如今，Rabbit 實際上承認他們使用了 Playwright，但他們不願意承認 LAM，不管它是什麼，並沒有真正起到很大的作用。

Jesse responds, to say LAM is just Playwright is equal to say chat GPT is just a chat.

傑西迴應說，說 LAM 只是 Playwright，就等於說哈拉 GPT 只是哈拉。

We use Playwright as part of a component to operate LAM.

我們使用 Playwright 作為運行 LAM 的組成部分。

That doesn't mean the research and technology is fake.

這並不意味著研究和技術是假的。

So what they're saying is LAM is not these hard-coded scripts, but kind of, they also admit these scripts exist.

是以，他們的意思是 LAM 並不是這些硬編碼的腳本，但他們也承認這些腳本的存在。

Maybe LAM interacts with the scripts.

也許 LAM 與腳本有交互作用。

Maybe they create them.

也許是他們創造的。

But even if that's true, it's not what they promised.

但即使這是真的，也不是他們所承諾的。

Either way, everything that matters on this device is basically off the shelf software.

無論如何，這款設備上所有重要的功能基本上都是現成的軟件。

The web actions are done by hard-coded scripts.

網絡操作由硬編碼腳本完成。

Everything else is done by chat GPT or perplexity.

其他一切都由哈拉 GPT 或困惑完成。

So there's not much reason for the LAM to do anything, whatever it actually does.

是以，無論 LAM 實際做什麼，它都沒有什麼理由去做。

And we're already so far from what Jesse originally promised, which I have to show you again.

我們已經離傑西最初的承諾相去甚遠了，我必須再給你看一遍。

He kept saying, was AI not a script that would be doing all these actions for you?

他一直在說，人工智能難道不是一個腳本，它會為你完成所有這些操作嗎？

We have a super close to computer that when Jason talks to his Rabbit R1 about ordering a hamburger from DoorDash, what's going on is that we first see if Jason's logging with DoorDash or logging with UberEats, right?

我們有一臺超級接近計算機的設備，當傑森與他的 Rabbit R1 談論從 DoorDash 訂購漢堡包時，我們首先會看到傑森是在用 DoorDash 登錄，還是在用 UberEats 登錄，對嗎？

And we saw, okay, Jason chose DoorDash.

我們看到，傑森選擇了 DoorDash。

And then on that superhost, a LAM is interacting virtually with a DoorDash app or website interface.

然後在超級主機上，LAM 與 DoorDash 應用程序或網站界面進行虛擬交互。

And you don't see all of that because all of a sudden this is done all at once because it's AI.

而你看不到這一切，因為這一切都是人工智能的傑作。

He couldn't be clearer.

他說得再清楚不過了。

The LAM is interacting because it's AI.

LAM 可以互動，因為它是人工智能。

This is a flat out lie, at least currently, which he says the LAM is here now.

這是一個徹頭徹尾的謊言，至少目前是這樣，他說 LAM 現在就在這裡。

Now I reached out over and over to the Rabbit team, pressing them again and again on this issue.

現在，我一次又一次地聯繫 Rabbit 團隊，就這個問題一再向他們施壓。

What does the LAM do?

LAM 是做什麼的？

And they kept saying, oh, playwright, yeah, that's the thing, but it's just part of how the LAM controls things.

他們一直說，哦，劇作家，是的，就是這樣，但這只是林氏集團控制事情的一部分。

But I kept asking, what does that mean?

但我一直在問，這是什麼意思？

How does this thing exactly work?

這東西到底是怎麼工作的？

If the scripts are doing the work, what does LAM do?

如果是腳本在工作，那麼 LAM 要做什麼？

And eventually it got to the point where I just questioned, look, is this real?

最後，我開始懷疑，這是真的嗎？

Do you have proof that this LAM can do all these amazing things?

你有證據證明這個 LAM 能做這些神奇的事情嗎？

And I should have known I was going to get into trouble here because Jesse, the CEO says, anyone challenging the LAM legitimacy,

我早該知道我會在這裡惹上麻煩，因為傑西首席執行官說，任何挑戰林氏企業合法性的人都會惹上麻煩、

I automatically become super defensive.

我就會自動變得超級防衛。

And that was true.

事實的確如此。

Here's part of the Rabbit team's response to me.

以下是 Rabbit 團隊給我的部分回覆。

To start off, I would love to know more about your background in tech because if you are not specialized in AI and machine learning, any conclusion resulting from this conversation will have no real credibility or value from any audience because this is a technical discussion in AI and machine learning.

首先，我很想了解一下你的技術背景，因為如果你不是人工智能和機器學習方面的專家，那麼這次對話得出的任何結論都不會有真正的可信度，也不會對任何聽眾有價值，因為這是一次人工智能和機器學習方面的技術討論。

We welcome you to invite experts with valid certificates from the field to further look into this.

我們歡迎您邀請持有該領域有效證書的專家對此進行進一步研究。

Now, look, I found this wild for a company that raised $30 million in venture, sold $20 million in product to be feisty about questioning their product when it does not work as advertised.

現在，聽著，我發現，對於一家籌集了 3 千萬美元風險投資、銷售了 2 千萬美元產品的公司來說，當產品不能像廣告宣傳的那樣正常工作時，還對自己的產品提出質疑，實在是太瘋狂了。

And we found out it's using hard-coded scripts instead of AI.

我們發現它使用的是硬編碼腳本，而不是人工智能。

How can they be surprised people are going to question what their thing actually does?

他們怎麼會驚訝人們會質疑他們的東西究竟是做什麼的呢？

But I also think it's funny they should say that I should invite experts to look into this because I'm not nearly the first person to point these problems out.

不過，我也覺得他們說我應該請專家來研究這個問題很有趣，因為我幾乎不是第一個指出這些問題的人。

Several people who, as I said, I spoke with have had issues with this device.

正如我所說的，與我交談過的一些人都遇到過這種設備的問題。

They have looked at the code base.

他們查看了代碼庫。

They have also claimed they couldn't find the LAM.

他們還聲稱找不到 LAM。

Others have jailbroken the R1, got into Rabbit's cloud to play Doom.

還有人越獄了 R1，進入 Rabbit 的雲端玩 Doom。

And despite Rabbit saying they welcome credible critics, their reaction has been anything but welcoming.

儘管 Rabbit 說他們歡迎可信的批評者，但他們的反應卻並不歡迎。

Emily, who we spoke to earlier, pointed out back in January that CAPTCHA would be a big concern for the LAM.

艾米麗（Emily）早在一月份就指出，驗證碼將是林氏療法的一個大問題。

And Jesse's response is, sorry, I work for my team.

傑西的回答是：對不起，我為我的團隊工作。

I don't work for you.

我不為你工作

So you need to check my calendar before you command me to answer anything now.

所以，在你命令我現在回答任何問題之前，你得先看看我的日程表。

But fast forward to today, CAPTCHA is a huge problem.

但時至今日，驗證碼已成為一個大問題。

In fact, specifically with Uber, we had to answer 10 different CAPTCHAs adding up a bunch of dice in order to get it working.

事實上，具體到 Uber，我們必須回答 10 個不同的驗證碼，再加上一堆骰子，才能讓它正常工作。

It was a giant pain.

這是一種巨大的痛苦。

It took us several tries.

我們試了好幾次才成功。

I began to wonder if I was actually part robot myself.

我開始懷疑自己是否真的是機器人的一部分。

But the point is getting defensive about the problem didn't make the problem go away.

但問題是，對問題採取防衛態度並不能使問題消失。

And the same is true about getting defensive about the LAM.

對林肯車的防衛也是如此。

Just because you tell people they have to be an expert to ask you questions won't stop the questions or the problems.

你告訴人們必須是專家才能向你提問，但這並不能阻止人們提問或解決問題。

But now I want to take a quick detour into the source code of the Rabbit itself.

不過，現在我想快速繞道進入 Rabbit 本身的源代碼。

Because I also have reviewed parts of the source code and I have serious data privacy concerns.

因為我也查看了部分源代碼，我對數據隱私有嚴重的擔憂。

There are things in this code base, and I can't share it for reasons that will become obvious.

這個代碼庫中有一些東西，我不能與大家分享，原因顯而易見。

There are things that malicious actors could use to get any access to the replies the R1 has ever given.

惡意行為者可以利用某些手段獲取 R1 曾經給出的任何回覆。

Now, of course, Rabbit might say they've dedicated a huge picture of a lock on their webpage to privacy.

當然，現在 Rabbit 可能會說，他們在網頁上專門用了一張巨大的鎖的圖片來保護隱私。

So you might think they're serious about it.

所以你可能會認為他們是認真的。

But this is the same company that if you read the fine print, they're tracking your precise geographic location and collect that data.

但就是這家公司，如果你仔細閱讀細枝末節，就會發現他們正在追蹤你的精確地理位置並收集這些數據。

The only consolation prize is this thing is shockingly bad at tracking you, which I found out while looking for restaurants.

唯一令人欣慰的是，這東西的追蹤功能差得驚人，這是我在找餐館時發現的。

Find me a restaurant near here.

給我找一家附近的餐館。

Looking it up now.

現在正在查。

Venetian Hot Plate, San Juan Restaurant, and Butter Churn are highly rated restaurants in Port Aransas.

Venetian Hot Plate、San Juan Restaurant 和 Butter Churn 是阿蘭薩斯港評價很高的餐館。

For a broader...

為了更廣泛地...

I'm not in Port Aransas.

我不在阿蘭薩斯港。

Let me see what I can find.

讓我看看能找到什麼。

For restaurants near you in Austin, Texas, consider the following options.

如需瞭解德克薩斯州奧斯汀附近的餐館，請考慮以下選項。

I'm not in Austin!

我不在奧斯汀

An Italian restaurant in downtown...

市中心的一家意大利餐廳

I'm not in Austin, Texas.

我不在德克薩斯州奧斯汀。

I use the internet connection to know where the R1 device is roughly located in order to provide you with relevant...

我通過網絡連接瞭解 R1 設備的大致位置，以便為您提供相關資訊。

You're off by about a minute.

你差了一分鐘。

I'm not in Austin, Texas.

我不在德克薩斯州奧斯汀。

I'm not in Austin, Texas.

我不在德克薩斯州奧斯汀。

I'm not in Austin, Texas.

我不在德克薩斯州奧斯汀。

I'm not in Austin, Texas.

我不在德克薩斯州奧斯汀。

You're off by about 100 miles.

你偏離了大約 100 英里。

Thank you for the clarification.

謝謝你的澄清。

Let me refine the search for restaurants in your approximate location.

讓我搜索一下您大概所在位置的餐廳。

Is it gonna do it?

它能做到嗎？

Is it gonna do it?

它能做到嗎？

It just stopped.

它就這麼停了。

It just...

它只是...

Yeah, I found its GPS to be really bad.

是的，我發現它的全球定位系統非常糟糕。

Nevertheless, I didn't want to take any chances for my next experiments, which is why when I use this thing,

儘管如此，我還是不想在接下來的實驗中冒任何風險，這就是我使用這東西的原因、

I'm not in the $10 million studio at all.

我根本不在 1000 萬美元的工作室裡。

I'm in an $100 motel.

我在一家 100 美元的汽車旅館裡。

Here to show you that once you know the code base, you can make some surprising predictions about what this thing's gonna do.

在這裡，你可以看到，一旦你瞭解了代碼庫，你就能對它的功能做出一些驚人的預測。

For example, when I tell the rabbit to play anything, you might think that'd be random, but it's not gonna be.

例如，當我讓兔子玩任何遊戲時，你可能會認為那是隨機的，但事實並非如此。

It's gonna be the Beatles.

會是披頭士

How did I know that?

我怎麼知道的？

Stop, stop, stop, stop, stop, stop, stop, stop.

停、停、停、停、停、停、停、停、停。

Stop, thank you.

停，謝謝。

The reason I know it's gonna do that is because it's hard-coded to do so in their prompting.

我之所以知道它會這樣做，是因為在它們的提示中硬性編碼了這樣的提示。

They say if the user is going to, say, play anything, they should play the Beatles.

他們說，如果用戶要玩什麼，就應該玩披頭士。