This cost six million dollars to develop that to give you some sort of sense around what that means.

為了讓大家瞭解這意味著什麼，我們花費了六百萬美元來開發這個項目。

Some of OpenAI's models have apparently cost in excess of a hundred million dollars.

OpenAI 的一些模型顯然耗資過億。

Is it as cheap as they're claiming?

有他們說的那麼便宜嗎？

We don't know.

我們不知道。

Or why are there still many questions about it?

或者，為什麼還有很多問題？

Yeah, we don't know.

是的，我們不知道。

Six million dollars sounds, a lot of industry experts have said that sounds like it's a made up number, or at the very least it's a very massaged number, because the numbers that we've seen so far, and even when we talk to some of the most innovative AI developers in the world, how quickly and cheaply they might be able to produce something, they've not said they'd ever be able to produce a model at a cost like that.

很多業內專家都說，600 萬美元聽起來像是一個編造出來的數字，或者至少是一個被篡改過的數字，因為我們目前看到的數字，甚至當我們與世界上一些最具創新精神的人工智能開發者交談時，他們都沒有說過他們能夠以多快、多便宜的成本生產出這樣的模型。

So there are some big question marks over those sorts of numbers.

是以，這些數字還需要打上大大的問號。

But we can't forget that they've open sourced this, which is something that OpenAI hasn't done.

但我們不能忘記，他們已經將其開源，這是 OpenAI 沒有做到的。

Theirs is behind a wall.

他們的在牆後面。

You can't do that.

你不能這麼做。

The second big breakthrough is that DeepSeq released a tech report detailing how they trained the model.

第二個重大突破是 DeepSeq 發佈了一份技術報告，詳細介紹了他們是如何訓練模型的。

Part of that recipe had a big innovation of using a method called reinforcement learning to teach the model how to reason without any human supervision.

該配方的一部分有一個重大創新，即使用一種稱為強化學習的方法來教模型如何推理，而不需要任何人的監督。

And this is kind of the holy grail of deep learning, is how can we kind of scale the learning of models, or the improvement of models, without having to have humans in the loop at every step.

這也是深度學習的聖盃，那就是我們如何才能擴大模型學習或模型改進的規模，而不需要在每一步都讓人類參與其中。

I think it's really important to understand that deep learning is not something that you can just download the AI model and run it on your own.

我認為，重要的是要明白，深度學習並不是下載一個人工智能模型就能自己運行的。

So we've seen examples of other people running the DeepSeq R1 model, and then you ask them all kinds of things, like 64Ti doors and so on.

是以，我們已經看到其他人運行 DeepSeq R1 模型的例子，然後你會問他們各種各樣的問題，比如 64Ti 門等等。

The AI model can answer you directly, because it doesn't have the follow-up screening.

人工智能模型可以直接回答你，因為它沒有後續篩選。

I think it's the big moment.

我認為這是一個重要時刻。

It's what you call a black swan.

這就是所謂的黑天鵝。

We should have expected this.

我們早該料到會這樣。

It was always US versus China.

一直以來，美國與中國都是對立的。

If you go back a week ago, when you think about that picture of the tech bros next to the president, and then them announcing half a trillion for Stargate, this big project for AI.

如果你回到一週前，當你想到那張科技兄弟在總統旁邊的照片，然後他們宣佈為星際之門投入 5 萬億美元，這個人工智能的大項目。

And it just looked like American tech supremacy was set for decades.

看起來，美國的科技霸主地位將持續數十年。

But in AI, what matters isn't just the quality of the innovation, but it's also the quality of the data.

但在人工智能領域，重要的不僅是創新的品質，還有數據的品質。

And it turns out that countries like China, for various reasons, have got a lot of data going into a lot more, with less qualms about privacy, perhaps, than the West.

事實證明，像中國這樣的國家，由於種種原因，已經有大量數據進入了更多的領域，對隱私的顧慮可能比西方國家更少。

So could that be a route to better AI models?

那麼，這是否可能成為改進人工智能模型的途徑呢？

Perhaps not in this case.

也許在這種情況下不會。

But it raises all sorts of questions about our assumptions, not just about tech, not just about the economy, but also about power in the world.

但這也引發了我們對各種假設的質疑，不僅是對科技、經濟，還有對世界權力的質疑。

In the Biden administration in the US, the word safety used quite a lot, about AI safety.

在美國拜登政府中，關於人工智能安全的 "安全 "一詞用得很多。

Go slowly.

慢慢來。

Make sure everything's tested.

確保一切都經過測試。

Make sure that you're doing everything in the proper way.

確保你以正確的方式做每一件事。

But now I think DeepSeek has really unleashed the beast, in a sense.

但現在我認為，從某種意義上說，DeepSeek 真正釋放了野獸的力量。

We're going to see companies releasing more products.

我們將看到公司發佈更多產品。

It takes a lot of data.

這需要大量數據。

If you look at the privacy policy of the app, it's similar to a social network.

如果你看一下該應用的隱私政策，就會發現它與社交網絡類似。

For example, it takes things like keystroke rhythms to identify people by how they type, that kind of thing.

例如，它可以利用擊鍵節奏來識別人們的打字方式，諸如此類。

But we're not really hearing about concerns about safety yet, especially from the US.

但我們還沒有真正聽到對安全的擔憂，尤其是來自美國的擔憂。

Donald Trump has only said, it's a wake-up call.

唐納德-特朗普只說了一句話：這是一個警鐘。

And we know, of course, that the Trump administration has basically overridden the safety laws that Biden put in place.

當然，我們也知道，特朗普政府基本上推翻了拜登制定的安全法律。

So I think it's a free-for-all now.

所以我認為現在是自由競爭。

What we will probably see in the coming weeks is people exploring the application of this model in domains like medicine, where it's not strictly the same as mathematics and programming, but being able to reason about difficult case problems is an important skill.

在接下來的幾周裡，我們可能會看到人們在探索這一模型在醫學等領域的應用，在這些領域中，數學和編程嚴格來說並不相同，但能夠推理疑難病例問題卻是一項重要的技能。

I'm fairly excited to see how physicians or people working in medical research use this model to figure out, can you use it as an aid for doctors to get a kind of second opinion on their diagnosis?

我非常期待看到醫生或從事醫學研究的人員如何利用這一模型來研究，能否將其作為一種輔助工具，幫助醫生獲得對其診斷的第二意見？

Well, if they collapse the cost of running, making effective these chips, that shouldn't necessarily be bad for all of these companies.

好吧，如果它們降低了運行和製造這些芯片的成本，對所有這些公司來說未必是壞事。

People thought that some of these tech companies were a little bit bubbly.

人們認為其中一些科技公司有點泡沫化。

They couldn't really sustain these values.

他們無法真正維持這些價值觀。

And some sort of pinpricking moment we were waiting for.

我們一直在等待的某種 "針刺時刻"。

For countries that aren't already the dominant AI superpower and have all of this investment and have the ability to grab hold of this tens of billions of dollars worth of investment, it's inspirational for them because it means that you can start thinking in a more innovative way.

對於那些尚未成為人工智能超級大國的國家來說，擁有這些投資並有能力抓住這些價值數百億美元的投資，對他們來說是鼓舞人心的，因為這意味著你可以開始以一種更具創新性的方式思考問題。