Name: 谷歌 DeepMind 首席執行官展示構建世界的人工智能模型精靈 2 (Google DeepMind CEO demonstrates world-building AI model Genie 2)
Uploaded: 2025-04-23T00:12:44.000Z
Duration: 5 min 48 s

程度副詞

This building across the square looks quite new.

廣場對面的這座建築看起來很新。

That is the Google London office in Pancras Square.

這就是位於潘克拉斯廣場的谷歌倫敦辦公室。

This week on 60 Minutes, we reported on DeepMind, Google's artificial intelligence research lab where we explored what's coming next with co-founder and CEO Demis Hassabis, a leader in this new world.

本週，我們在 "60 分鐘 "節目中報道了谷歌的人工智能研究實驗室 DeepMind，並與聯合創始人兼首席執行官德米斯-哈薩比斯（Demis Hassabis）探討了這個新世界的未來。

DeepMind's AI assistant, Astra, can see, hear, and analyze the world around it.

DeepMind 的人工智能助手 Astra 可以看到、聽到並分析周圍的世界。

It can recognize famous paintings like this one.

這就是愛德華-霍珀的《Automat》。

And guess what its subject might be feeling.

猜猜它的主題可能是什麼感覺。

The subject in the painting appears pensive and contemplative.

But with a little push, it can do even more.

但只要稍加推動，它還能做得更多。

Can you create a story around this picture?

你能圍繞這幅圖創作一個故事嗎？

It's a chilly evening in the city, a Tuesday perhaps.

這是個寒冷的夜晚，也許是個星期二。

The woman, perhaps named Eleanor, sits alone in the diner, enjoying a warm cup of coffee.

這位也許名叫埃莉諾的女士獨自坐在餐廳裡，享受著一杯熱咖啡。

Have you found yourself saying, I didn't see that coming?

你有沒有發現自己在說：我沒想到會這樣？

It's happened many times, actually, since the beginning of DeepMind.

實際上，自 DeepMind 成立以來，這種情況已經發生過很多次了。

And Astra, just being able to be that good at understanding the physical world was not something we were expecting it to be that good at that quickly.

而阿斯特拉能夠如此出色地理解物理世界，這並不是我們所期望的。

While reporting this story, we learn more about the strides DeepMind has made with AI that can produce images, videos, and even three-dimensional worlds.

在報道這篇報道時，我們瞭解到 DeepMind 在人工智能方面取得的更多進展，它可以生成影像、視頻，甚至三維世界。

Two years ago, we were shown a technology that could create videos with simple text prompts like Golden Retriever with wings.

兩年前，有人向我們展示了一種技術，它可以通過簡單的文字提示來製作視頻，比如長著翅膀的金毛尋回犬。

Since then, it has taken a massive leap forward.

從那時起，它就開始了巨大的飛躍。

It's just a way clearer, more detailed image.

這只是一個更清晰、更細緻的影像。

And the snowflake's sort of falling on him and collecting as a...

Director of product management, Tom Hume, showed us this model called VO2.

產品管理總監湯姆-休姆（Tom Hume）向我們展示了一款名為 VO2 的模型。

It recreated the fantastical scene of the Golden Retriever, but with astonishing photorealistic detail.

它再現了金毛尋回犬的奇幻場景，但細節逼真得令人驚歎。

It even knows how wings flap, like how it would flap in the air.

它甚至知道翅膀是如何扇動的，比如在空中如何扇動。

Eventually, we've seen this in science fiction as sort of like the holodeck from Star Trek, where you can just imagine any environment.

最終，我們在科幻小說中看到了這一點，就像《星際迷航》中的 "全息甲板 "一樣，你可以想象任何環境。

The system will generate it for you, and then you can sort of immerse yourself in that imagined environment.

系統會為你生成它，然後你就可以沉浸在想象的環境中。

Asabis and DeepMind research scientist Jack Parker Holder wanted to show us Genie2, a world model that uses images to create interactive 3D environments.

阿薩比斯和 DeepMind 研究科學家傑克-帕克-霍爾德希望向我們展示 Genie2，這是一種利用影像創建交互式 3D 環境的世界模型。

This is a photograph taken by someone in our team somewhere in California.

這是我們團隊中的某個人在加利福尼亞某處拍攝的照片。

And what we then do is ask Genie to convert this into an interactive world.

然後，我們要求精靈將其轉換成一個互動世界。

So we prompt the model with this image, and Genie converts it into a game-like world that you can then interact in.

是以，我們用這張圖片提示模型，然後精靈將其轉換成一個類似遊戲的世界，然後你就可以在其中互動了。

Every further pixel is generated by a generative AI model.

接下來的每一個像素都是由人工智能生成模型生成的。

So the AI is making up this scene as it goes along.

所以說，人工智能是邊走邊編造這個場景的。

Someone from our team is actually playing this.

They're pressing the W key to move forwards.

And then from that point onwards, every subsequent frame is generated by the AI.

從那時起，接下來的每一幀都由人工智能生成。

Parker Holder told us these models could eventually train AI agents to accomplish tasks in these environments.

帕克-霍爾德告訴我們，這些模型最終可以訓練人工智能代理在這些環境中完成任務。

We've created this world with three different arches, where on the left we see some nice plants.

我們用三個不同的拱門創造了這個世界，在左邊我們可以看到一些漂亮的植物。

In the middle, it's an opening, which we don't quite know what's in there.

中間是一個開口，我們不知道里面有什麼。

And rather than a human playing, we take one of our most capable AI agents, and we say, can you actually go up the staircase for us?

與其讓人扮演，不如讓我們最有能力的人工智能代理來扮演，然後我們說，你能幫我們爬上樓梯嗎？

And as it's climbing the stairs, the Genie world model is creating the world around it on the fly instead of imagining what's up there.

在它爬樓梯的過程中，精靈世界模型正在飛快地創造它周圍的世界，而不是想象那裡有什麼。

Effectively, you've got one AI system playing in the mind of another AI system.

實際上，你已經讓一個人工智能系統在另一個人工智能系統的頭腦中進行遊戲。

So you've got one generating the world, and the other one is moving around and accomplishing goals in that world.

是以，你有一個人在生成世界，另一個人則在這個世界裡四處走動並完成目標。

And the practical implication of that would be what?

So, of course, there's lots of implications for entertainment and generating games and videos.

當然，這對娛樂、生成遊戲和視頻也有很多影響。

But actually, the bigger goal is building a world model, a model that can understand our world.

但實際上，更大的目標是建立一個世界模型，一個能夠理解我們這個世界的模型。

You could imagine future versions creating almost an infinite variety of different simulated environments, which the AIs can learn from and interact in, and then translate that to the real world.

你可以想象，未來的版本幾乎可以創造出無限多種不同的模擬環境，人工智能可以在其中學習和互動，然後將其轉化到現實世界中。

So if you had a full-sized robot, for example, you wouldn't train him on how to navigate the world out in the streets in front of the building.

是以，舉例來說，如果你有一個全尺寸的機器人，你不會訓練它如何在大樓前的街道上瀏覽世界。

You would put him in this scenario so that he could figure out how to climb stairs.

你會把他放在這種情景中，讓他學會如何爬樓梯。

It's much harder to collect data in the real world, much more expensive, much slower.

在現實世界中收集數據要困難得多，成本要高得多，速度要慢得多。

But in simulated worlds, you can sort of collect almost an unlimited amount.

但在模擬世界中，你幾乎可以無限量地收集。

We wondered if Google could use the trove of geographic data that's at its fingertips to advance this technology.

我們想知道，谷歌是否可以利用其掌握的地理數據寶庫來推動這項技術的發展。

Google has access to almost infinite data.

谷歌可以獲取幾乎無限的數據。

You've got Google Maps, you've got Street View, you've got Google Earth.

你有谷歌地圖，你有街景，你有谷歌地球。

Could all of those data be collected together in order to train AI agents?

能否將所有這些數據收集在一起，以訓練人工智能代理？

That's what we're exploring at the moment, actually, is both ways.

這正是我們目前正在探索的，實際上是雙管齊下。

So potentially using Street View kind of data to give real-world understanding and geographical understanding to our AI systems.

是以，我們有可能利用街景數據為我們的人工智能系統提供對現實世界和地理環境的理解。

And then on the other hand, you can imagine things like this bringing to life static images of real places, whether it's your own holiday photos or actually Street View views, which are static currently, and actually making them interactive and 3D so you can look around the place itself.

另一方面，你也可以想象這樣的事情：將真實地點的靜態影像（無論是你自己的度假照片還是街景照片，目前都是靜態的）變成互動的 3D 影像，這樣你就可以環顧四周。