# Kai-Fu Lee on the future of AI | The TED Interview

Segment00:00:00 – 00:10:00

Full episode transcript, trimmed to training segment.

==================================================

[0000:05] Hello, I'm Chris Anderson, welcome to The TED Interview,

[0000:08] the podcast series

[0000:09] where I get to sit down with a TED speaker

[0000:11] and dive much deeper into their ideas

[0000:14] than was possible during their short TED Talk.

[0000:17] Now, today on the show, Kai-Fu Lee

[0000:21] and the epic race to develop artificial intelligence, AI.

[0000:27] Kai-Fu was 11 years old when he moved to America from Taiwan

[0000:31] to be immersed in the US education system.

[0000:33] And from that point, I guess you could say

[0000:36] his life has been positioned to straddle two worlds:

[0000:39] America and Asia.

[0000:42] After graduating in computer science,

[0000:44] Kai-Fu developed a reputation as a star AI innovator,

[0000:49] and spent time at Apple, Microsoft and Google.

[0000:53] But 10 years ago,

[0000:54] he decided to invest all of his time and energy

[0000:56] into China's high-tech companies.

[0000:59] His motivation?

[0001:00] He saw something truly extraordinary

[0001:02] in the new generation of Chinese tech entrepreneurs.

[0001:06] (TED Talk) Kai-Fu Lee: Chinese entrepreneurs,

[0001:08] whom I fund as a venture capitalist,

[0001:10] are incredible workers, amazing work ethic.

[0001:13] As an example, one start-up tried to claim work-life balance,

[0001:17] "Come work for us, because we are 996."

[0001:21] And what does that mean?

[0001:23] It means the work hours of 9 am to 9 pm, six days a week.

[0001:29] That's contrasted with other start-ups that do 997.

[0001:32] (Music)

[0001:34] CA: Kai-Fu and I are going to discuss

[0001:36] how China's gladiatorial fight to the death

[0001:40] has made China better positioned to lead the world in the AI race.

[0001:44] Also, the amazing ways AI is already being incorporated

[0001:48] into China's everyday life,

[0001:50] the massive risk of disruption to our economies,

[0001:54] and a beautiful way of thinking

[0001:55] about how AI could actually enhance the future of human work.

[0002:00] I found this to be such a fascinating conversation,

[0002:03] alarming and thrilling in equal measure.

[0002:06] Let's do it.

[0002:09] Kai-Fu Lee, welcome.

[0002:11] KFL: Thank you, great to be here.

[0002:12] CA: So I'm so excited for this conversation,

[0002:15] we get to dive into what you consider, I think what I consider,

[0002:20] to be really one of the most important topics,

[0002:22] perhaps the most important topic facing humanity's future:

[0002:25] artificial intelligence, AI,

[0002:28] and where it will take us, and how excited we should be about it,

[0002:32] how worried we should be about it,

[0002:33] what we might do with that excitement and those worries.

[0002:36] A good place to start, Kai-Fu, would be:

[0002:38] you have said that artificial intelligence,

[0002:41] that we should think of it as a sort of an epoch-changing invention,

[0002:46] along the lines of the steam engine and electricity.

[0002:50] Why do you make that case?

[0002:52] KFL: I think we've only had a very small number

[0002:55] of truly groundbreaking technologies that impact every part of our lives,

[0003:00] and artificial intelligence is a technology

[0003:04] that can learn to make very accurate decisions,

[0003:08] predictions, classifications

[0003:12] purely based on seeing a lot of data.

[0003:15] So that kind of capability will bring us amazing efficiency,

[0003:21] greater profits, cost savings,

[0003:24] and also liberate us from a lot of routine jobs.

[0003:28] And it impacts every imaginable profession.

[0003:32] So it is as big as electricity, and maybe even bigger.

[0003:37] CA: So, explain that to me,

[0003:39] and perhaps in the context of the recent breakthroughs

[0003:42] that have been made in deep learning.

[0003:44] That seems to be the fundamental shift that has changed AI

[0003:48] from being a sort of curiosity that some people tried to work on

[0003:52] in university departments,

[0003:53] to a business-changing and life-changing innovation.

[0003:59] KFL: The field of AI actually began in the 50s.

[0004:04] And the field really hasn't made much progress

[0004:08] until deep learning came about.

[0004:10] And deep learning is learning purely on correlation,

[0004:15] on huge amount of data,

[0004:17] and then draws the features and conclusions

[0004:22] pretty much by itself.

[0004:24] So if you think about recognizing pictures of cats,

[0004:29] it just sees millions and even billions of pictures,

[0004:33] and learns what it is that makes a cat a cat.

[0004:37] And it doesn't really use human abstractions or features,

[0004:43] and it's just looking at a thousand-dimensional space

[0004:48] and drawing a thousand-dimensional curve

[0004:52] that separates what is a cat from what isn't.

[0004:55] And that curve is really hard for humans to comprehend.

[0005:00] It's just like our brainwaves are hard for deep learning to understand.

[0005:05] So this is not a replication of human thinking process,

[0005:09] but an entirely mathematical classification engine

[0005:14] that learns from a lot of data and what humans tell it,

[0005:18] "This is a cat, this is not, this is Chris, that's Kai-Fu,"

[0005:22] and it learns to recognize them better than people

[0005:27] if they could see millions or even billions of pieces of data.

[0005:32] So they need a lot of data to do well,

[0005:35] the more data, the better it does,

[0005:38] but you don't really have to teach it

[0005:41] specific rules and decision-making rules.

[0005:45] CA: Right, it's able to use that data

[0005:47] because someone put a caption of a cat somewhere on a picture at one point,

[0005:51] so that it knows, basically, cat or no cat.

[0005:54] Is part of it then saying

[0005:56] well, any sphere that you apply that power to

[0005:59] can show this incredible pace of outpacing us.

[0006:04] Talk about some of the fields where there is just vast amounts of data

[0006:08] and where AI is already showing remarkable promise.

[0006:11] KFL: Sure, an example is machine translation.

[0006:14] There's a lot of data on the web,

[0006:16] because we've got all these web pages with translated versions:

[0006:21] United Nations proceeds and so on,

[0006:24] and what the machine-learning algorithm can do

[0006:27] is just take these parallel texts

[0006:29] of French going into English, Chinese going into Arabic,

[0006:33] and all these pairs of languages that have previously been translated,

[0006:38] and feed all of this to an engine, and out comes a machine translator.

[0006:44] Now, the current machine translation isn't yet as good as professionals,

[0006:48] but it's very good,

[0006:50] and that's purely based on text that's been translated on the internet.

[0006:56] And similarly, faces and speech,

[0007:01] turning speech into text,

[0007:03] learning to talk like you or me,

[0007:06] even generating natural language, making up a story,

[0007:10] AI is getting better and better at it.

[0007:12] And there are further advanced applications

[0007:16] where it's not yet beating people,

[0007:19] but autonomous vehicles are now driving millions of miles,

[0007:24] robotics are able to manipulate and pick up objects and move around,

[0007:28] so we're going to see many waves and new applications coming up over time.

[0007:35] CA: In his hilarious sci-fi book,

[0007:37] Douglas Adams, who wrote "The Hitchhiker’s Guide to the Galaxy,"

[0007:40] he had in that book a Babel fish that you put in your ear

[0007:44] and it automatically translated any language coming in.

[0007:49] There are already devices on the market

[0007:50] that are staring to do this in still, I guess, fairly primitive form,

[0007:56] but how long do you think it is

[0007:57] before any tourist can go to any major city on the planet

[0008:02] with a little earbud or something in their ear

[0008:05] and just have a conversation in real time with anyone,

[0008:08] in any of the world's major languages?

[0008:10] KFL: Yeah, you picked a good one.

[0008:12] We're almost there with this one, I think in about three years.

[0008:15] We're basically just cleaning up the engineering hard edges

[0008:20] to make the product flow more smoothly.

[0008:23] We already have text-to-text translation,

[0008:26] speech recognition is getting more and more accurate,

[0008:29] and speech synthesis is getting more and more humanlike.

[0008:33] So this little device will have a latency of maybe one to three seconds.

[0008:39] It can't quite work like humans, which are instantaneous,

[0008:42] but other than this one-to-three-second delay,

[0008:46] I think the tourists can already buy devices that are usable.

[0008:51] In three years, these will be very good devices.

[0008:54] Now, one caveat is, when you said everywhere in the world,

[0008:58] so you've got to have enough data

[0009:00] to train the translation pairs of languages.

[0009:03] So if you go to a country

[0009:06] in which the language is spoken only by a few thousand people,

[0009:10] then you probably don't have the data.

[0009:12] So you've still got to go to one of the more popular languages

[0009:16] or more populous countries

[0009:18] for this to work well.

[0009:20] CA: It seems to me that this is just one area

[0009:22] where people haven't even begun to really follow through

[0009:25] on the implications of that.

[0009:26] People hear about this maybe,

[0009:29] but clearly don't believe it yet,

[0009:31] because if you believed that that was coming,

[0009:33] and say, in three-years' time

[0009:34] this device, with a two- or three-second delay,

[0009:37] but that means that in eight-years' time,

[0009:39] there's a device with less than a one-second delay.

[0009:41] The implications of that are so astonishing.

[0009:43] First of all, in terms of lots of jobs that are there right now get taken out,

[0009:48] but perhaps more interestingly,

[0009:50] just huge amounts more of connection, I imagine.

[0009:53] So many people who right now

[0009:55] maybe don't want to spend too much time in a foreign place,

[0009:58] or if they do, they don't really connect with people