Name: 走進 ChatGPT 架構師 OpenAI，特邀 Mira Murati | The Circuit with Emily Chang (Inside OpenAI, the Architect of ChatGPT, featuring Mira Murati | The Circuit with Emily Chang)
Uploaded: 2025-01-02T23:43:24.000Z
Duration: 24 min 2 s

Inside a nondescript building in the heart of San Francisco, one of the world's buzziest startups is making our AI-powered future feel more real than ever before.

在舊金山市中心一棟不起眼的大樓裡，世界上最熱門的初創公司之一正在讓我們感覺人工智能驅動的未來比以往任何時候都更加真實。

They're behind two monster hits, ChatGPT and Ali, and somehow beat the biggest tech giants to market, kicking off a competitive race that's forced them all to show us what they've got.

他們是 ChatGPT 和阿里這兩款熱門遊戲的幕後推手，並以某種方式擊敗了最大的科技巨頭，掀起了一場競爭激烈的競賽，迫使他們都向我們展示自己的實力。

But how did this under-the-radar startup pull it off?

但是，這家名不見經傳的初創公司是如何做到這一點的呢？

We're inside OpenAI, and we're going to get some answers.

我們在 OpenAI 內部，我們將得到一些答案。

這給我《西部世界》水療中心的感覺。

It's almost like suspended in space and time a little bit.

它幾乎就像懸浮在時空中一樣。

Yeah, it is a little bit of futuristic feel.

This is one of the most introspective minds at OpenAI.

這是 OpenAI 最善於反思的人之一。

我們都認識首席執行官山姆-奧特曼。

But Mira Moradi is a chief architect behind OpenAI's strategy.

但米拉-莫拉迪（Mira Moradi）是 OpenAI 戰略背後的首席架構師。

My guess is that it's AI that loves humanity.

我的猜測是，愛人類的是人工智能。

We're very focused on dealing with the challenges of hallucination, truthfulness, reliability, alignment of these models.

我們非常專注於應對幻覺、真實性、可靠性、這些模型的一致性等挑戰。

Has anyone left because they're like, you know what, I disagree?

有人因為 "你知道嗎，我不同意 "而離開嗎？

There have been, over time, people that left to start other organizations because of disagreements on the strategy around deployment.

隨著時間的推移，有些人因為在部署戰略上出現分歧而離開，成立了其他組織。

And how do you find common ground when disagreements do arise?

當出現分歧時，如何找到共同點？

You want to be able to have this constant dialogue and figure out how to systematize these concerns.

你希望能夠不斷進行對話，並找出如何將這些問題系統化的方法。

It's a combination of guiding the teams on the ground, thinking about longer-term strategy, figuring out our gaps, and making sure that the teams are well-supported to succeed.

這是指導實地團隊、思考長期戰略、找出我們的差距，以及確保團隊獲得良好支持以取得成功的綜合工作。

When you were making the decision about releasing Chat GPT into the wild, I'm sure there was like a go or no-go moment.

當你決定將 Chat GPT 發佈到野外時，我相信你一定經歷了 "去 "還是 "不去 "的抉擇。

We had Chat GPT for a while, and we sort of hit a point where we could really benefit from having more feedback from how people are using it, what are the risks, what are the limitations, and learn more about this technology that we have created and start bringing it in the public consciousness.

我們已經使用 Chat GPT 有一段時間了，我們發現，如果能從人們如何使用它、有哪些風險、有哪些限制等方面獲得更多反饋，我們就能真正受益。

It became the fastest-growing tech product in history.

它成為歷史上增長最快的科技產品。

I mean, what was your reaction to the world's reaction?

我是說，你對世界的反應是什麼？

We were surprised by how much it captured the imaginations of the general public and how much people just loved spending time talking to this AI system and interacting with it.

讓我們感到驚訝的是，它如此吸引大眾的想象力，人們如此喜歡花時間與這個人工智能系統交談和互動。

At the most basic level, how does this all happen?

在最基本的層面上，這一切是如何發生的呢？

So, Chat GPT is a neural network that has been trained on a huge amount of data on a massive supercomputer, and the goal during this training process was to predict the next word in a sentence, and it turns out that as you train larger and larger models and more and more data, the capabilities of these models also increase.

是以，Chat GPT 是一個在大型超級計算機上通過海量數據訓練出來的神經網絡，訓練過程中的目標是預測句子中的下一個單詞，事實證明，隨著訓練的模型越來越大，數據越來越多，這些模型的能力也會越來越強。

They become more powerful, more helpful, and as you invest more on alignment and safety, they become more reliable and safe over time.

隨著時間的推移，它們會變得更強大、更有用，而且隨著你在調整和安全方面投入更多，它們也會變得更可靠、更安全。

Open AI has kind of turbocharged this competitive frenzy.

開放式人工智能為這種競爭狂潮注入了新的活力。

Do you think you can beat Google at its own game?

你認為你能在谷歌的遊戲中擊敗它嗎？

Do you think you can take significant market share in search?

您認為您能在搜索領域佔據重要的市場份額嗎？

What Chat GPT offers is a different way to understand information, and you could be, you know, searching, but you're searching in a much more intuitive way versus keyword-based.

Chat GPT 提供的是一種理解資訊的不同方式，你可以進行搜索，但搜索方式要比基於關鍵字的搜索直觀得多。

I think the whole world is sort of now moving in this direction.

我認為現在整個世界都在朝著這個方向發展。

The air of confidence, obviously, that Chat GPT sometimes delivers an answer with.

顯然，哈拉 GPT 有時會讓人感到信心十足。

Why not just sometimes say, I don't know?

為什麼有時不說 "我不知道 "呢？

The goal is not to predict the next word reliably or safely.

我們的目標不是可靠或安全地預測下一個單詞。

When you have such general capabilities, it's very difficult to handle some of the limitations, such as what is correct.

當你有了這樣的通用能力，就很難處理一些限制，比如什麼是正確的。

Some of these texts and some of the data is biased.

其中一些文本和一些數據有失偏頗。

Isn't this going to accelerate the misinformation problem?

這不是會加速誤報問題的解決嗎？

I mean, we haven't been able to crack it on social media for like a couple of decades.

我的意思是，我們在社交媒體上已經有幾十年無法破解了。

Misinformation is a really complex, hard problem.

錯誤信息確實是一個複雜而棘手的問題。

Right now, one of the things that I'm most worried about is the ability of models like GPT-4 to make up things.

現在，我最擔心的事情之一就是 GPT-4 等模型的編造能力。

So they will convincingly make up things and it requires being aware and just really knowing that you cannot fully, blindly rely on what the technology is providing as an output.

是以，他們會令人信服地胡編亂造，這就要求我們必須意識到，不能完全、盲目地依賴技術提供的輸出結果。

I want to talk about this term hallucination because it's a very human term.

我想談談 "幻覺 "這個詞，因為這是一個非常人性化的詞。

Why use such a human term for basically an AI that's just making mistakes?

為什麼要用這麼人性化的詞語來形容一個只會犯錯的人工智能？

A lot of these general capabilities are actually quite human-like.

這些一般能力中，有很多其實都很像人類。

Sometimes when we don't know the answer to something, we will just make up an answer.

有時，當我們不知道某件事情的答案時，我們就會編造一個答案。

And so, there is a lot of human hallucination in a conversation and sometimes we don't do it on purpose.

是以，對話中有很多人類的幻覺，有時我們並不是故意的。

Should we be worried about AI though that feels more and more human?

我們是否應該擔心感覺越來越像人類的人工智能？

Like, should AI have to identify itself as artificial when it's interacting with us?

比如，人工智能在與我們互動時，是否應該表明自己是人工智能？

I think it's a different kind of intelligence.

我認為這是一種與眾不同的智慧。

It is important to distinguish output that's been provided by a machine versus another human.

重要的是要區分機器輸出和人工輸出。

But we are moving towards a world where we're collaborating with these machines more and more and so output will be hybrid.

但是，我們正在邁向這樣一個世界：我們越來越多地與這些機器合作，是以產出將是混合型的。

All of the data that you're training this AI on, it's coming from writers, it's coming from artists.

你訓練人工智能的所有數據都來自作家和藝術家。

How do you think about giving value back to those people when these are also people who are worried about their jobs going away?

當這些人也在擔心自己的工作不保時，你如何考慮讓他們重新獲得價值？

I don't know exactly how it would work in practice that you can sort of account for information created by everyone on the internet.

我不知道在實踐中如何才能對互聯網上每個人創造的資訊進行分類。

I think there are definitely going to be jobs that will be lost and jobs that will be changed as AI continues to advance and integrate in the workforce.

我認為，隨著人工智能的不斷進步和融入勞動力大軍，肯定會有工作崗位流失，也會有工作崗位發生變化。

That's not something that we could have predicted.

Think of prompt engineers like AI whisperers.

把提示工程師想象成人工智能耳語者。

They're highly skilled at selecting the right words to coax AI tools into generating the most accurate and illuminating responses.

他們善於選擇合適的詞語，哄騙人工智能工具生成最準確、最有啟發性的回答。

It's a new job born from AI that's fetching hundreds of thousands of dollars a year.

這是人工智能催生的新工作，年薪高達數十萬美元。

What are some tips to being an ace prompt engineer?

成為王牌提示工程師有哪些祕訣？

You know, it's this ability to really develop an intuition for how to get the most out of the model.

你知道，這種能力能夠真正形成一種直覺，知道如何最大限度地利用模型。

How to prompt it in the right ways, give it enough context for what you're looking for.

如何以正確的方式進行提示，提供足夠的背景資訊，以滿足您的需求。

One of the things that we talked about earlier was hallucinations and these large language models not having the ability to always be highly accurate.

我們之前談到的其中一件事就是幻覺，這些大型語言模型並不能始終保持高度準確性。

So I'm asking the model with a browsing plugin to fact check this information.

是以，我要求帶有瀏覽插件的模型對這些資訊進行事實核查。

So there's this report that these workers in Kenya were getting paid $2 an hour to do the work on the backend to make answers less toxic.

是以有報道稱，肯亞的這些工人每小時能拿到 2 美元的報酬，他們在後端進行工作，以減少答案的毒性。

And my understanding is this work is, it can be difficult, right?

我的理解是，這項工作可能很困難，對嗎？

Because you're reading texts that might be disturbing and trying to clean them up.

因為你正在閱讀那些可能令人不安的文本，並試圖將它們清理乾淨。

So we need to use contractors sometimes to scale.

是以，我們有時需要利用承包商來擴大規模。

We chose that particular contractor because of their known safety standards.

我們之所以選擇該承包商，是因為他們的安全標準眾所周知。

And since then we've stopped working with them.

從那以後，我們就不再與他們合作了。

走進 ChatGPT 架構師 OpenAI，特邀 Mira Murati | The Circuit with Emily Chang (Inside OpenAI, the Architect of ChatGPT, featuring Mira Murati | The Circuit with Emily Chang)