對話OpenAI Jack Clark：人工智慧無需超越人類也能改變世界
導語：Eye on A.I.是由紐約時報資深記者Craig S. Smith主持的一檔雙週部落格節目。每一期節目，Craig都將與這一領域有影響力的人物進行交流，推進廣義環境中的機器智慧新發展，思考技術發展新蘊意。 機器之心為此係列對話的中文合作方。以下為此係列內容的第一篇，Craig Smith與Jack Clark就全球AI發展進行了探討。
Hi, this is Craig Smith with a new podcast about artificial intelligence. I’ll be talking to Jack Clark, author of the popularofollow,noindex" target="_blank"> Import AI newsletter , about what he has learned during the previous week and why it is significant. Jack is a veteran of the British technology journal, The Register, and of Bloomberg News. He now works on policy and communications for OpenAI , the nonprofit artificial intelligence research company founded by Elon Musk . So think of this podcast as a review of what’s happening in the world of AI, curated by one of its keenest observers.
大家好，我是 Craig Smith，這是我建立的一個關於人工智慧的播客。今天與我對話的是 JackClark，他是很受歡迎的 Import AI 新聞訂閱 源 的作者。今天我們要談談他過去一週瞭解到了什麼以及為什麼這很重要。Jack是英國科技雜誌 TheRegister和 Bloomberg News的資深行業老兵。他現在也為 ElonMusk資助的非營利性人工智慧研究公司 OpenAI工作，致力於政策和溝通問題。所以你可以把今天的播客看作是對當前人工智慧世界的概覽，而且是來自一位最敏銳的觀察者。
CRAIG: I wanted to start by talking about the bottom of the newsletter this week in which you talk about what was said at the World Economic Forum in Davos. The thing that made a lot of headlines and certainly jumped out at me, was [01:09] Google CEO Sundar Pichai’s comment that the rise of AI is more important than the discovery of electricity or even fire. And you see a lot of statements like that. It surprised me seeing that come from Pichai although I don’t follow him closely. Is that overstating it in your view?
Craig：首先，我想談談本週訂閱的最後部分的內容，其中你談到了有關達沃斯世界經濟論壇上所涉及的內容。有一件事上了很多頭條，當然也引起了我的關注，那就是谷歌的 CEO Sundar Pichai 對 AI 崛起的 評論 ，他說這比發現電甚至火還要重要。而且你也能看到很多類似的評論。但是，這些話從 Pichai口中說出，還是挺讓我驚訝的，儘管我沒有密切關注過他。在你看來，這樣的表述是否誇張？
JACK: It is not overstating it. And I think it’s pretty commendable that he’s talking about this head on, because a lot of executives have been trying to say that this technology has the same attributes as previous disruptive technologies like the advent of the smartphone. This is of a fundamentally different nature. And though when we hear people talk about electricity and fire, we think that’s incredibly overblown, I believe that if you think of how AI is going to play out over the next, not just few years, but few decades, we’re going to look back on it with something equivalent to that. From a historical perspective, it will suddenly look like you built computers and very soon after, the computers gained human-like capabilities. And soon after that, as we’ve seen with things likeAlphaGo, that the computers quickly surpassed human experience. And that’s very unusual and I think significant.
Jack：並不誇張。而且我覺得他能正面談論這個問題是很值得稱道的，因為有很多公司高管都一直在講，這項技術和之前的顛覆性技術有一樣的屬性，比如智慧手機的出現。實際上人工智慧技術的本質存在根本性的不同。而且儘管當我們聽到人們談論電和火時，我們會認為他們說得實在太過誇張，但我相信如果你思考一下 AI會在未來的幾十年，而不是幾年，會如何發展，我們回頭看時也會有一樣的感覺。從歷史角度看，很可能看起來就像是計算機突然就發明出來了，之後很快計算機就具有了接近人類的能力。在那之後不久，我們看到AlphaGo這樣的計算機很快超越人類的事情出現了。這是非常不同尋常的，我認為這意義重大。
CRAIG: Yes, I agree. I guess I'm a bit of a skeptic in the timeline. I can see super intelligence and the intelligence explosion happening at some point in the human history. I just am not optimistic it’s this century. But we will…
Craig：是的，我同意。我想我只是對時間軸有些疑慮。我能想見超級智慧和智慧爆炸會在人類歷史的某個節點發生。只是我並不樂觀地認為會在這個世紀發生。但未來會發生……
Jack: Well, it does not need to be super intelligent. I'm going to push back on that point with you now which is that, sure super intelligence, yes, that’s going to be obviously a big deal if it happens. But we don’t need for that to happen for this to be huge.
Jack：嗯，其實並不需要是超級智慧。我要反駁一下你的看法；當然，如果確實出現了，那超級智慧顯然是個大事件。但人工智慧成就重大的變革並不需要超級智慧的出現。
Think about exactly how much the world was changed by the advent of the database or the spreadsheet. The ability to just have computers organize rote textual information is one of the forces that led to globalization and the information revolution. Now we have capabilities that let computers access basic human senses, like an approximation of vision, an approximation of hearing and they’re going to live in a world which is actually built for humans to see and hear things.
想一想資料庫或電子表格的出現究竟給世界帶來了多麼大的變化。讓計算機管理複雜繁瑣的文字資訊是推動實現全球化的一大力量。現在，我們有能力讓計算機具備基本的人類感知能力，比如對視覺的近似、對聽覺的近似，而且它們將生活在一個實際上為人類的聽和看而構建的世界裡。
So, just try to imagine about exactly how much information is becoming available right now to computers. And then think about what happened over the last 30 years with spreadsheets or databases, and I’d say to you that we don’t need this stuff to be super intelligent for it to change the world far more than the digital revolution so far has to date. And that to me is kind of mindboggling.
所以，只要嘗試想象一下現在計算機能夠使用多少資訊，然後再思考一下過去 30年電子表格或資料庫所造成的影響，那我就能讓你明白我們並不需要超級智慧就能改變世界，而且改變的程度將遠遠超過數字革命迄今所造就的一切。對我來說，這非常讓人振奮。
CRAIG: Yes, that much I certainly agree with. He talked about the need for global multi-lateral frameworks and then Theresa May put her country’s flag in the ground as a leader in developing the ethics of AI.
Craig：是，我當然認同這一點。他談到了對全球多邊框架的需求，而且 Theresa May（英國首相）也樹立了國家的旗幟，要成為 AI道德倫理方面的領導者。
Is that overoptimistic to think that [04:33] we’ll be able to come up with a multi-lateral framework that countries will adhere to given how unsuccessful we’ve been with nuclear energy technology or nuclear weapons technology?
考慮到我們在核能技術或核武器技術方面如此地不成功，我們真能構建一個各個國家都會遵守的多邊框架嗎？這是不是過於樂觀？
JACK: You and I are having this conversation and neither of us can point to a place on the earth which has had a nuclear weapon use in anger since the World War II. Again, I'm going to say to you, we built institutions that worked here because we aren’t talking about a tragedy after Hiroshima and Nagasaki. There is huge proof that the international governance system that we built around nuclear weapons works. It has its problems. You and I are speaking in 2018 and it’s a time when, I think, the clock to midnight on how far away we are from nuclear annihilation has ticked up to two minutes to midnight which is the closest it’s been in a while. But nonetheless, we persist.
Jack：在座的我們兩人都不能說出在第二次世界大戰之後地球上有誰因為憤怒而使用了核武器。另外，我也要對你說，我們建立了一些有效的機構，因此我們現在才沒有談論廣島和長崎事件之後的悲劇。有大量的證據表明我們圍繞核武器所構建的國際治理體系是有效的。當然也存在問題。現在是 2018年，現在我們離核毀滅結局的末日時鐘僅有兩分鐘了，這是很長時間以來最接近毀滅的時間。但不管怎樣，我們仍在堅持。
So, no, I don’t think it’s over optimistic. I do think it is optimistic and it’s optimistic especially, as someone coming from the UK, which is unmooring itself from Europe and therefore from power blocks. I do think it’s realistic to think that we’re going to see countries collaborate on various AI standards. And I think it’s more than likely that you’re going to see norms develop in the international community around AI and how it intersects with the military.
所以，不，我並不認為這過於樂觀。我確實認為這有些樂觀，尤其是某個來自英國的人想成為領導者，而這個國家正在脫離歐洲，因此也將脫離一個有力量的區域。我也確實相信我們能看到各個國家合作構建各種 AI標準。而且我還相信你肯定會看到圍繞 AI的國際社群開發出相應的規範，約束其在軍事上的應用。
Other than a few bad actors, we rarely use things like chemical weapons and cluster bombs these days thanks to international norms. And yes, there are exceptions, but we know that we can do this stuff, it’s just immensely difficult. But that doesn’t mean we should shy away from the problem.
除了少量惡人之外，現在我們已很少看到人們使用化學武器和集束炸彈這樣的武器，這都多虧了國際規範的建立。是的，確實存在例外，但我們知道我們能做成這件事，只是極其地困難。但困難並不意味著我們應該回避這個問題。
CRAIG: Yes. I guess the thing that’s been said about AI is unlike a cluster bomb, it’s much more difficult to detect when it’s being used. As far as Theresa May’s claim that the UK will be a leader in AI, from my point of view, I spent much of my career in China and I see China as the clear leader going forward simply because they benefit from a centrally commanded economy and they have a huge state sector that doesn’t worry about cost. They’ve got the largest population and are spinning out the most engineers.
Craig：是的。但我想人工智慧終究不同於集束炸彈，在使用時被檢測到的難度要大得多。對於 Theresa May 宣稱英國將成為 AI領域領導者的言論，就我看來，因為我的工作很多都在中國，我覺得中國才是顯而易見的領導者。原因很簡單，因為他們有一個由中央掌控的經濟模式，得益於此，他們有大量不在乎成本的國有企業和機構。他們擁有最多的人口，並且正在培養最多的工程師。
Do you think that countries like the UK have a chance to play a leadership role? And do you think that countries like China will engage in the kind of frameworks that the West is going to want to control this stuff?
你認為英國這樣的國家有機會成為領導者嗎？另外，對於這種西方國家想要控制的框架，中國這樣的國家會參與進來嗎？
JACK: You’re handing me a very complicated question there. I'm going to split this open into a couple of things. Do I think countries like the UK can have a role here? I'm skeptical. I'm skeptical because we know that AI requires you to be the pinnacle of scientific and technological development from a national perspective. So, though the UK has some AI advantages, I'd argue a huge amount of that is because of investment into the UK from American companies. Facebook and Google and Microsoft. The startup ecosystem in the UK is somewhat contingent on that.
Jack：這個問題真的非常複雜。我會將其分成幾個小問題來說。英國這樣的國家有機會嗎？我表示懷疑。因為我們知道 AI需要從國家層面佔領科學和技術發展的制高點。儘管英國有一些 AI優勢，但我認為很大一部分優勢都源自 Facebook、谷歌和微軟等美國公司對英國的投資。英國的創業生態環境也在某種程度上取決於此。
So, Prime Minister May is giving the impression that this is entirely domestically developed, but if you look at the enabling factors, it actually comes from somewhere else where there’s a larger amount of money swilling around in it and the proportionally larger scientific and technological culture.
May首相讓人覺得這完全是國內發展起來的，但如果你看看是什麼因素帶來了這樣的發展，你會看到實際上來自別處，那裡有多得多的資金，而且有更廣泛的科學和技術文化。
Now, do I think that larger countries like the US or China or India or Russia, even, by virtue of some of its specific investments, can have a role in international AI? Yes, that seems likely. I think that we see the emergence of what one could term power blocks here.
接下來的問題是，我覺得美國、中國、印度或俄羅斯等更大的國家，甚至通過其特定的投資，能夠成為國際 AI發展的領軍者嗎？是的，很有可能。我認為我們將看到力量核心的興起。
Now to your question of norms, will these various power blocks respect the norms which the majority tries to set? Mostly, no. Mostly, people are going to do what they’ve always done which they try to grow their own economies and compete in the international game of geopolitical brinksmanship. But, and here’s the difference and you alluded to this earlier, AI has the potential for great harm and so, there are incentives here for collaboration around certain norms relating to how one wields and uses AI, which even if you have radically different ideological s towards how you see yourself competing on the global stage, you have a pretty big incentive to want to be able to have certain confidences about your competitors or opponents and how they’ll use this powerful technology.
現在回到關於規範的問題，這些不同的力量核心會尊重大多數人試圖設立的這些規範嗎？基本上不會。基本上而言，人們會繼續做他們一直都在做的事情，也就是發展自己的經濟，在地緣政治邊緣政策的國際博弈中互相競爭。但是，仍會有不同之處，你之前也提到過，AI有可能造成很大的傷害，所以這會激勵人們圍繞特定的規範進行合作，這就涉及到了人們開發和使用 AI的方式。即使你對自己如何參與國際舞臺上的競爭有自己的看法，你也會希望能夠對你的競爭者或對手以及他們使用這項強大技術的方式具有一定的信心。
So, I do think that countries like China as you mentioned will have an incentive to conform to some norms. I do think that those norms are going to be relatively few in number though.
所以，我確實認為你提到的中國這樣的國家有動力去遵守一些規範。但我也確實認為他們會遵守的規範的數量會相對少一些。
CRAIG: Yes, we talked about this before, the question of whether or not China is publishing everything that they do in open forums. My guess is that there’s a lot - and I think you agreed - that there’s a lot going on in China that no one sees. That’s certainly of some concern.
Craig：是的，我們之前談過這個，問題是中國是否會公佈他們所做的一切。我想你肯定也同意，中國在做的很多事情沒人能看到。這肯定會引起某些擔憂。
[10:32] Something else you write about in the newsletter is a student in Florida who used computer vision and deep learning to create a live overview of a popular multiplayer online video game, League of Legends. What struck me about that is the relative ease and low expense, both in dollars and computational time, with which this guy managed to come up with a solution.
你在本週的訂閱訊息中還寫到了另一件事，佛羅里達州的一個學生使用計算機視覺和深度學習建立了一個對流行的多玩家網路視訊遊戲《英雄聯盟》的 實時概覽 。讓我驚訝的是這相當簡單而且成本低，不管是資金成本還是計算時間。這個傢伙成功提供了一個解決方案。
It also highlighted to me how much data there is out there to play with if you know how to scrape it because he was using just videos of gameplay. Beyond that I'm not sure I can see the usefulness of what he’s done except for that world of gaming. But that struck me that this student in Florida with very little money and with publicly available datasets come up with a solution. I would guess we’re going to start seeing more and more of that.
這也凸顯出其實存在很多可以使用的資料——如果你知道如何使用的話，因為他的這個專案其實就只使用了遊戲視訊。除此之外，我不確定我明白他所做的工作在遊戲世界之外的應用。讓我驚訝的人這個佛羅里達的學生僅用少量資金和公開可用的資料集就提出了一個解決方案。我想將會有越來越多這樣的事情出現。
JACK: Yes, I think you’ve hit on a really interesting point which is, why is it interesting? Well, it’s not interesting that they were able to build a cool doohickey for an eSport, for a computer game that you and most people will not have heard of.
Jack：是，我認為你談到了一個真的非常讓人感興趣的觀點。為什麼有趣呢？嗯，並不是因為他們有能力為電子競技建立炫酷的小道具，或者說為你和大部分人都沒聽說過的計算機遊戲。
What is cool is that this doohickey allows them to access information about the game which the company could expose by the application programming interface of the game to developers but have chosen not to. So, in a sense, the significance of this is they were simply able to point this powerful technology at a bunch of data that they were able to scrape from the game. In doing so, the deep learning algorithms essentially improvised the underlying API which the company didn't want the person to have access to.
真正炫酷的是這個小道具讓他們可以獲取有關遊戲的資訊——遊戲公司可以通過提供給開發者的遊戲應用程式介面來提供，但卻沒有選擇這麼做。所以，從某種程度上講，這個專案的重要意義在於他們只是證明可以使用這種強大的技術來處理從遊戲中收集到的資料。這麼做本質上是讓深度學習演算法臨時創造了公司不希望人們訪問的底層 API。
So, when we think about what this means, it really calls into question a lot of stuff about how proprietary software works. Because the assumption with proprietary software is I can sell you some software because you’re going to get some value from it and you’re not going to be able to clone the software I sell you purely from looking at the inputs and outputs of the system.
所以，當我們思考這個專案的意義時，這真的會勾起我們對專有軟體的工作方式的疑問。因為專有軟體的假設是我能向你出售某個軟體，因為你能從這個軟體獲得價值，你不能就通過觀察這個系統的輸入和輸出就克隆出這個軟體。
What this gives us a flavor of is that, with lots of AI technologies, all I need to know is the inputs and outputs and I can improvise the rest. That calls into question a lot of how we think of IP protection and IP based businesses working in modern era.
而我們現在可以感受到，使用很多 AI技術，我只需要輸入和輸出就能構建出其它部分。這不禁會讓我們思考這個時代的 IP保護方式和基於 IP的商業業務。
CRAIG: Yes, you’ve got this student in Florida that’s able to come up with this pretty interesting solution based on open source software and publicly available data. Then you have these powerful tools, like Detectron that people like Facebook are making publicly available. [13:35] Is the day coming when, this sort of democratization of AI, when it escapes the big heavily funded labs and becomes a garage production thing, where people are doing things out of their basements? I mean doing significant things out of their basements.
Craig：確實，佛羅里達的這個學生只用開源軟體和公開可用的資料就創造了這個出色又有趣的方案。這樣的工具現在有很多，比如 Facebook就公開提供了 Detectron。你覺得，從重金支援的實驗室到車庫就能生產的東西，這種 AI的民主化快要來臨了嗎？我的意思是人們能在家中自己動手就做出重大的成果嗎？
JACK: No, sadly. I could explain why, or I don’t know if my response is just going to depress you so much you want to stop this endeavor. Would you like me to explain why I think that’s not the case?
Jack：不行，很遺憾。我可以解釋原因，我不知道我的回答是否會讓你感到失望，以至於你不想繼續談。你想要我解釋我認為這做不到的原因嗎？
CRAIG: I would be fascinated, yes.
Craig：我很樂意聽，請解釋。
JACK: What you see with these kinds of things is what I term a computational dividend from these large companies. They’ve spent a huge amount of money on electricity bills to develop some system that has a capability. In this case, a research platform that can let me draw bounding boxes around the world. So fine, that represents the basic commoditizing effect of technology R&D. It means that people in garages around the world are now going to be able to access this basic sense.
Jack：你看到的那些公開專案都源自大公司的計算紅利。他們投入了大量電費來開發某些效能卓越的系統。這裡說的是一個能讓我為世界繪上邊界框的研究平臺。這代表了技術研發的基本商業化效應。這意味著世界各自在自家車庫中創造的人現在都可以獲得這種基本能力了。
In the same way that it’s a commodity currently to be able to point a webcam at the world and have it offer a label as to what it thinks the most prominent thing in the image is, where stuff like Detectron, or stuff like the League of Legends thing we’re speaking about. You can point your camera at something specific and have it produce a load of information about that. Or you can point it at a specific domain and get it to tell you something useful.
以同樣的方式，可以讓網路相機觀察世界，然後提供其影象中最可能存在的事物的標籤。現在已有這樣的商品了。你還可以將你的相機對準某個特定的物體，就能得到有關它的大量資訊。或者你可以指向特定的區域，就能讓它為你提供一些有用的資訊。
But none of those things really represent the cutting edge. The cutting edge is going to exist on far larger computers than any start-up can possibly hope to wield or any person in the garage. It’s going to require research techniques which have not yet appeared in research papers. The difference in capability is going to be profound.
但這些都代表不了技術前沿。技術前沿會存在於大得多的計算機上，比任何創業公司所希望的或個人在車庫中所用的計算機大很多。這會需要一些尚未出現在研究論文中的研究技術。在能力的差別將具有重大的影響。
In AI now, every six months, the world changes. It used to be every few years, the world changes. And before that it used to be every decade the world changes. The technological epochs are multiplying and the intervals between them are reducing. So, the competitive moat that people like Facebook and others have is getting deeper over time because they’re able to wield larger and larger models. I’ll give you a very, sort of, tangible example of this.
AI世界目前每隔六個月就會大變樣。過去這個時間段是好幾年。更早之前甚至要每隔十年才有大變化。這項技術正在倍速發展，變革週期也在縮短。所以，Facebook等巨頭的競爭優勢將會隨時間越來越大，因為他們有能力研發越來越龐大的模型。我可以舉一個非常形象的例子。
I as a start-up can have Detectron. I as a start-up can have a residual network, or a highway network, or some kind of advanced deep neural network system. I can even have the data. It’s still going to take me a period of time to train a model based off of my own data to do something useful. The training time is going to be conditioned on the amount of computers I can access. I don’t know how large your garage is, but it’s definitely not as big as a football field I'd wager. Whereas, Facebook has a football field worth of computers. What that means is that when Facebook wants to do something involving R&D, it can do it far faster than anyone in the garage.
如果我有一家創業公司，我可以使用 Detectron，我也可以使用殘差網路、highway 網路或其它的先進 深度神經網路 系統。我甚至能獲得資料。但如果要做出一些有用的東西，我仍然需要使用自己的資料來訓練一個模型，這需要一定的時間。訓練時間的長短取決於我能夠使用的計算機。我不知道你的車庫有多大，但我敢說肯定不會有足球場那麼大。然而，Facebook的計算機能擺滿足球場。這意味著當 Facebook想要做點研發時，開發速度會超過任何在車庫中搞開發的人。
So, if Facebook’s ability to empirically experiment and discover new AI techniques is proportional to the amount of computers Facebook has, it’s very hard to see a world in which start-ups can really easily compete with these AI giants because they simply don’t have a large enough computer to be able to experiment as rapidly as them despite having benefits of the computational dividends from the companies.
所以，如果 Facebook的實驗能力和發現新 AI技術的能力與其所能使用的計算機數量成比例，那麼這個世界上就很難有創業公司能輕鬆與這些 AI巨頭競爭，因為他們根本沒有足夠大的計算機來幫助他們足夠快地完成實驗。就算創業公司能從這些公司取得一些計算紅利，也難以與之抗衡。
CRAIG: Well, maybe it’s the computational dividends though that will finance or make possible smaller applications that the big guys aren’t interested in pursuing or aren’t spending time thinking about.
Craig：也許這些計算紅利能夠產生回報，比如做巨頭沒有興趣或不願花時間的更小型的應用。
You wrote about DroNet , a joint project between universities in Switzerland and Spain to train drones to fly along city streets. They used publicly available data fromself-drivingcars and created their own dataset with bicycles. That to me was fascinating and it was, again, something that was created without huge investment and based on open source technology with, I think there is a Parrot drone, which is a cheap consumer drone. Isn’t the computational dividend there in these applications that are going to start happening on a more localized level?
你還寫到了 DroNet ，這是一個瑞士和西班牙的大學之間的一個聯合專案，目的是訓練能在城市街道上飛行的無人機。他們使用了來自自動駕駛汽車的公開可用資料並且使用了自行車建立了自己的資料集。我覺得這也是個不錯的專案，同樣沒有用到鉅額投資，同樣基於開源技術。我想他們用的是 Parrot無人機，一種較便宜的消費級無人機。這些應用中的計算紅利是否會開始在更區域性的層面上發生？
JACK: Yes, in the sense that you’re going to see some innovation at the edges. Sure, absolutely. The dataset which they trained that drone from, there were two parts to it.
Jack：是的，在某種程度上，你會看到在這些邊緣位置的一些創新。這當然是肯定的。他們訓練無人機所用的資料集包含兩部分。
One was a bicycle, so they didn't have a good dataset of collisions. So, they got a bicycle, strapped a GoPro or something to it and just simply pedaled with intent towards obstacles and tried to simulate collisions. I was quite glad to read in the paper that they didn't actually just generate tens of thousands of collisions. I would have felt very sorry for the researcher who had to do that.
一是自行車。他們沒有優良的撞車資料集。所以他們給一輛自行車綁上了一個 GoPro或類似裝置，然後故意讓它撞到障礙物來模擬撞車情形。我很高興能在他們的論文讀到這一點，因為他們並不是簡單地直接生成數以萬計的撞車事件。對於不得不去做這件事的研究者，我表示同情。
So, what they did is they accelerated towards it, deliberately slows down and then labelled that as something that the drone should avoid. So, sure, they generated some of their own data, but you’ll notice that the really strategic data which is the actual car driving data, the bit which tells the drones how to steer, how to follow the road, that comes from Udacity . That comes from a dataset generated by the Udacity online education course forself-drivingcars.
他們實際上做的是先加速衝向障礙物，然後故意減速，並將障礙物標記成無人機應該避開的東西。所以當然他們生成了一些自己的資料，但你也會注意到實際上至關重要的資料是真實的汽車駕駛資料，這些資料能告訴無人機如何轉向、如何按道路行駛。這些資料來自 Udacity 。是 Udacity為自動駕駛汽車線上教育課程所生成的資料集。
So, the Udacity online education course forself-drivingcars is a course run by Sebastian Thrun who helped design Google’sself-drivingcar. Udacity itself is funded by tens of millions of dollars of venture capital. And Sebastian Thrun knew how to create that dataset after spending four or five years working in secret at Google on itsself-drivingcar project. None of that suggest to me that Google is any less ahead as a consequence of this.
Udacity自動駕駛汽車線上教育課程由 Sebastian Thrun 主持，他曾幫助設計了谷歌的自動駕駛汽車。Udacity本身得到了風險資本的數千萬美元投資。而且在谷歌的自動駕駛汽車專案上祕密工作了四五年後，Sebastian Thrun 也知道如何建立這樣的資料集。但這些都不能讓我信服地表明谷歌失去了任何一點領先優勢。
In fact, it suggests the opposite for what we’re seeing here is late stage research projects made possible by investments made many years ago by companies which are somewhat opaque to us and whose capabilities aren’t super obvious. Yes, it creates a capability that’s interesting, like drones that can do some useful stuff. But imagine how much better that drone would be if it was trained off of Google’s giant internalself-drivingcar dataset.
事實上，剛好相反，我們現在看到的都已經是後期的研究專案，是那些公司很多年前的投資所產生的成果。這些公司的專案對我們而言都有一定程度上的不透明，能力如何也並不非常明顯。是的，DroNet 確實有些值得關注的能力，比如能做些有用事情的無人機。但想象一下，如果用谷歌那巨大的內部自動駕駛汽車資料集來訓練無人機，無人機的效能會好多少？
I think that can give you an idea here for how, say, Waymo at a drop of a hat now, Waymo knows it can train a far more capable drone navigation model than that which appeared in this paper because rather than having 70,000 images, it’ll have 700 million.
我覺得現在你已經明白，Waymo知道自己立馬就能訓練出一個比那篇論文好很多的無人機導航模型，因為他們擁有的影象遠不止 70000張，而是 7億張。
I just want to push on this point, I am struggling to see evidence of the possibility of effective competition in the AI ecosystem. No one has shown me evidence to convince me that start-ups have an easy life here. In fact, all of the evidence I see says the opposite.
我還想強調一下這個觀點。AI生態系統中很難再具有有效的競爭。還沒有任何證據能讓我相信創業公司能輕鬆地活下去。事實上，我看到的所有證據都指向了反面。
It says that start-ups are either using the dividends, so the lag investment of big companies, or they are competing in territories where we know the big companies have innate advantages and could simply exercise those advantages and run over the start-ups.
證據表明，創業公司要麼使用大公司的計算紅利，要麼就抓住大公司投資滯後的間隙，否則他們就要在大公司具備天然優勢的領域與之競爭——大公司只要利用一下這個優勢，就能輕鬆碾壓創業公司。
CRAIG: Are cases like DeepLeague and DroNet more examples of bright young engineers who will eventually go to work for Google, Facebook, Uber, Amazon or one of the big guys?
Craig：這是否說明聰明的年輕工程師基本最終都會為谷歌、Facebook、Uber 或亞馬遜等巨頭工作？
JACK: Yes. As far as I can work out, this shows us that it’s easy to contribute to AI, it’s easy to develop AI in the open using tools which are being given away to you by these larger companies. But I haven’t seen an AI start-up emerge which can truly show a capability that exists in excess of anything you see from either top tier academic research institutions, or far more frequently, the large companies.
Jack：是的，在我看來是這樣。這向我們表明使用開放的工具開發 AI並做出貢獻是很容易的，而很多工具都是由這些大公司提供的。我還沒看到任何一家新興的創業公司具備真正超越頂級學術研究機構的能力，更別說超過那些大公司了。
[22:43] To me AI is going to challenge a lot of our notions of anti-trust and a lot of our notions of how competition works in technology. Because the moats seem to grow deeper over time even though you’re able to release a huge amount of it as open source or in the open via research papers. It’s very paradoxical and it’s going to mean that if you’re a regulator, all of the symptoms are of a healthy market. You have some start-ups, you have lots of open innovation, you have lots of sharing.
在我看來，AI將會挑戰我們的很多反壟斷觀念以及很多有關技術競爭方式的觀念。因為就算大公司開源了大量資源或通過研究論文公開了研究成果，技術實力的差距還是會越來越大。這是非常矛盾的，這意味著監管者看到的現象都表明這個市場是健康的，因為有創業公司，也有大量開放的創新，還存在很多共享。
But then when you prod at the fundamentals, look at where the money is coming from, who has the strategic asset needed to do business like data or compute, or who has the talent, you find that it’s actually just the very large companies which have a vice-like grip on the fundamentals of what we think of as the ingredient to competition.
但如果究其根本，看一看資金的來源，誰有能力支撐資料或計算業務的戰略資本，誰擁有人才，你就會發現正是那些大公司像頑疾一樣掌控了競爭的基礎。
CRAIG: Yes. The stress is on companies, not on nation states.
Craig：是的。承受壓力的是公司，不是國家。
JACK: We can't view this in terms of nation states yet because we don’t really have the manifestation of coherent nation states research agendas yet.
Jack：我們不能從國家方面來看待這個問題，因為目前還沒有國家表現出連貫一致的科研議程。
[24:00] When you look at organizations in countries, in China, you see a very tight loop between government, the private sector and the public sector. None of it is what I'd quite call yet Chinese research. It’s more just what happens if you’re trying to catch up with people like Google, Microsoft and Amazon is you play every card you have.
對於中國這樣的國家中的組織機構而言，你會看到在政府、私營企業和公共部門之間存在非常緊密的聯絡。但其中沒有任何一個能說是中國的研究。更像是在追趕谷歌、微軟和亞馬遜等公司時所不得不做事——必須打出手中的每張牌。
In the case of China that’s using huge amounts of data, huge amounts of built-in commercial aspects like your companies and also the ability of government and funding to make things move a bit quicker and accelerate stuff. But there is no research paper that I'd say has a particular Chinese flavor yet. Nor would I say that there’s research I read where I say, “Oh well that feels like a very Belgium-like AI paper or German AI paper.”
對於中國的情況，他們有大量資料，有大量內部商業，就像你的公司一樣，另外政府還有能力和資金讓事情發展更快並且實現加速。但我要說沒有任何論文是中國特色的，我也不會說“哦，這篇論文是比利時風味的，或者說那是一篇德國的論文。”
I don’t think we’re quite there yet. These power blocks are going to manifest, but they’ll manifest under the purview of the implementation of national AI research agendas. Really the only country even doing national AI research is China and it’s only just initiated that. So, we need to wait about three or four years before we can see the fruits of those initial investments and intents on its path.
我認為我們還沒到這些力量核心能夠體現出來的時候，但從國家 AI研究計劃的實現角度看，它們將會得到體現。實際上中國是唯一一個在國家層面上推動 AI研究的國家，而且它才剛剛開始。所以，我們還需要等上三四年時間才能看到這些初期投資的成果以及他們所選道路的目的。
CRAIG: [25:29] That brings us to PsychLab which I really did get excited about. To me that was the beginning of testing an AI against human intelligence in psychometrics. Can you talk a little bit about that?
Craig：這讓我想起了 PsychLab ，我覺得這個專案很激動人心。我想這應該是心理計量學（psychometrics）方面對比人類智慧測試 AI的開始。你能談談這方面嗎？
JACK: Yes. PsychLab is a deep learning and reinforcement learning testing suite from DeepMind. What PschyLab does is it makes it possible for us to test and evaluate reinforcement learning algorithms for not just their performance but how well they do on human psychological tests. What this means is you can now test an AI algorithm on a cognitive science test and you can also have a human go and run the same test and you can compare their performance.
Jack：PsychLab 是一個來自 DeepMind的深度學習和強化學習測試套件。PsychLab不僅能讓我們測試和評估強化學習演算法的效能，還能讓我們瞭解它們在人類心理學測試上的表現。這意味著現在你可以基於認知科學測試 AI演算法，你還可以讓人類來進行同樣的測試，然後比較他們的結果。
The most interesting thing to me is that PsychLab let the scientists at DeepMind surface some pretty meaningful data points regarding the relative performance and drawbacks of their algorithms. In one case, they performed a test where they ask the reinforcement learning algorithm they are testing which was from a system that they call Unreal, which is one of their best performing and most highly tuned systems.
在我看來，最有意思的是 PsychLab 能讓 DeepMind的科學家知曉某些與他們的演算法的相對錶現和缺陷相關的非常有價值的資料點。在一個案例中，他們執行了一項測試——詢問他們所測試的來自一個名叫 Unreal的系統的強化演算法，哪個系統是表現最好的以及得到了最好調節的。
They asked it to perform a very basic test in cognitive science which is to do with looking at different sets of concentric circles. It’s called, the shattered glass problem. And working out which one has the largest amount of concentricity in it. What they discovered is that the Unreal agents perform very, very badly relative to humans. They were able to hypothesize that perhaps it performed badly because the way that the gaze or the vision system of these Unreal agents works is rather too different to how human vision works.
他們讓它執行了一個認知科學中非常基礎的測試，即檢視不同集合的同心圓。這被稱為碎玻璃問題（shattered glass problem）。目標是確定哪一組具有最大的同心度。他們發現，相比於人類，Unreal智慧體的表現非常差。他們給出了假設，也許其表現差的原因是這些 Unreal智慧體的注視和視覺系統的工作方式與人類視覺的工作方式存在非常大的差異。
So, what they ended up doing is implementing a vision system which was heavily based on what we call foveal vision, which is how in the center if your eyeball, Craig, you have more receptors relative to the periphery. This is what allows you to focus on objects that are close and to have higher resolution on things at the center of your vision.
所以，最後他們實現了一個嚴重依賴於所謂的中央凹視覺（foveal vision）的視覺系統，即你的眼球中央所具有的感受器多於周邊時的視覺。這能讓你重點關注近處的物體，並且你視覺中心的事物在你眼中的解析度也更高。
With a traditional convolutional neural network, the receptor field is going to be uniformed across all of it. What they did is that they bunched up the receptor field of a neural network in the same way you do it with a human eyeball, foveal vision. What they found is that not only were they able to create an agent that started to pass these tests with closer to human-like performance, but they’re able to take that very same agent and place it in a reinforcement learning environment called laser tag, which was not tested anywhere else in this testing suite and is actually much more like you having to play a virtual game of paintball with your friends, or laser tag. They found that suddenly, they had created unreal agents which were able to excel in this environment at a level of performance they really hadn’t been able to achieve before.
使用傳統的 卷積神經網路 ，感受野會在所有位置均勻分佈。他們所做的是將神經網路的感受野進行聚焦，正如你的眼球所實現的中央凹視覺一樣。他們發現，他們不僅能創造出能以接近人類的水平通過這些測試的智慧體，而且還能將完全一樣的智慧體放入被稱為鐳射標籤（laser tag）的強化學習環境中，這沒有在這個測試套件的任何地方測試過，實際上更像是你與朋友玩的虛擬彩彈遊戲。他們發現突然之間他們就創造出了擅長應對這一環境的 Unreal智慧體，並且達到了他們之前從未達到過的水平。
So, what that tells you is that by doing this psychological or cognitiveneurosciencetesting on their agents, they were able to actually find a drawback, fix the drawback and pass the test. Then take that agent out of the standard testing regime of PsychLab and still show good performance, and in some cases, exceptional performance on wholly new tasks that hadn’t been tested on before.
這說明，通過在他們的智慧體上執行這樣的心理學或認知神經科學測試，他們實際上找到了一個缺陷，並且修復了這個缺陷並使之通過了測試。然後他們將這個智慧體用在了 PsychLab 的標準測試方案之外，也仍然表現優良。在某些案例中，其在之前從未見過的全新任務上也取得了突出的表現。
That to me is just so exciting because it gives us a whole new way of thinking about how we can stress test these algorithms and how scientists can have another tool in their experimentation kit to help them diagnose things about it.
我覺得這振奮人心，因為這為我們提供了一種全新的思考方式——我們可以如何對這些演算法進行壓力測試，科學家可以如何將額外的工具用在他們的實驗套件中以幫助他們進行診斷。
CRAIG: I was also fascinated by that foveal vision fix, if you want to call it that. One of the things that I found interesting is where the gaps in the abilities of both the unreal agent and the human matched in that glass pattern test. Where one dot is white and the other is black, neither the human nor the AI agent could perform, could recognize the pattern. To me that was fascinating. So, there’s something in the cognitive processing of that that breaks down in both the human and the artificial agent.
Craig：這種中央凹視覺修復確實讓人著迷。另一件事我覺得也很有意思，即 Unreal智慧體與人類在玻璃圖案測試（glass pattern test）中所表現出的差異。其中一個點是白色，另一個點是黑色，人類和 AI智慧體都不能有效地識別這種圖案。我覺得這個問題很讓人著迷。所以，在認知處理中，存在某些讓人類和人工智慧體都束手無策的東西。
The other thing that really struck me in the analysis, I think in their paper they talked about how the convolutional neural networks can process features in parallel on a GPU. But humans are primarily restricted to serial processing of visual data. When they asked rhetorically why evolution didn't give humans such an ability as you would presume that it’s an advantage.
另一個讓我印象深刻的是其分析思路。在他們的論文中，他們談到了 卷積神經網路 在 GPU上並行處理特徵的方式。但人類對視覺資料的處理方式主要是序列處理。他們提出了一個問題：為什麼進化沒有為人類提供這種看起來更具優勢的能力呢？
But then they notice that the subjective experience of serial processing feels very much like the essence of thought. Because you think in a sequential stream, which, to me that was fascinating that you can start to identify features in human mental processing that are different from an artificial agent’s processing, t hat might explain some of our subjective experience. I don’t know if you picked up on that, but that was fascinating to me.
但然後他們就注意到，序列處理的主觀體驗非常類似於思維的本質。因為我們是按序列串流的方式思考的，我覺得這個問題很迷人，讓人開始明白人類心智過程並不同於人工智慧體的處理過程，這也許能夠為我們的主觀體驗提供某些解釋。我不知道你是否認同這一點，但我覺得這很有吸引力。
JACK: Yes, I think if AI continues to develop, we are going to learn very surprising things about our own biases and proclivities. Because we’re going to be able to hold up a cognitive system to a mirror.
Jack：我認為如果 AI繼續發展，我們將會了解到有關我們自身偏見和偏好的非常讓人驚訝的事情。因為我們能夠創造一個能反映我們自身的認知系統，就像一面鏡子。
When you look in the mirror in the morning or the evening, you see yourself, but you don’t really have a way of seeing the mirror to how your brain thinks you should act in the world. Because there’s no mirror that can simulate another person. We just look at other people and model ourselves on them.
當你在早上或夜晚窺鏡自視時，你能看到自己，但你實際上並不能從鏡子中看到你的大腦認為你應該在這個世界中如何行事。因為並不存在能夠模擬另一個人的鏡子。我們看待他人時，只是將我們的模型套在他們身上。
What these AI systems and PsychLab shows us is, as we get more creative and advance agents, perhaps we can have something that looks like a cognitive mirror where we can look at a different way of thinking and dealing with the same task, and in doing so, learn more about ourselves. That to me is so exciting.
這些 AI系統和 PsychLab 向我們表明，隨著我們的智慧體越來越有創造力和先進，也許我們可以創造出類似認知之鏡的東西，我們可以從其中看到處理同一任務時不同的思維和處理方式，這樣我們也能更理解我們自己。我覺得這很激動人心。
理論 OpenAI Jack Clark 人工智慧