Google Brain Helps Marginal Revolution University

In early November Google Translate took a Japanese translation of the opening of Hemingway’s “The Snows of Kilimanjaro,” and returned:

Kilimanjaro is 19,710 feet of the mountain covered with snow, and it is said that the highest mountain in Africa. Top of the west, “Ngaje Ngai” in the Maasai language, has been referred to as the house of God. The top close to the west, there is a dry, frozen carcass of a leopard. Whether the leopard had what the demand at that altitude, there is no that nobody explained.

One day later Google Translate took the same passage and returned:

Kilimanjaro is a mountain of 19,710 feet covered with snow and is said to be the highest mountain in Africa. The summit of the west is called “Ngaje Ngai” in Masai, the house of God. Near the top of the west there is a dry and frozen dead body of leopard. No one has ever explained what leopard wanted at that altitude.

What happened on that day is that Google turned its Translate service over to Google Brain, it’s new division that uses “neural networks” to solve AI problems. Google Brain and it’s history is the subject of  an excellent longread, The Great AI Awakening, from Gideon Lewis-Kraus (from which I have drawn the example).

Today, however, I want to make a different point. In my paper, Why Online Education Works, I wrote:

Online education has the potential to break the cost disease by substituting capital for labor and hitching productivity improvements in education to productivity improvements in software, artificial intelligence, and computing.

The improvements to Google Translate provide an example. Our Principles of Microeconomics and Principles of Macroeconomics courses at Marginal Revolution University are captioned in over a hundred languages. Professional human-written captions have been produced for most of our videos in English, Spanish, French, Chinese and Arabic and we are working on more translations. Most of the translations, however, including those for Corsican, Kyrgyz, and Urdu are provided by Google. The earlier machine-translations weren’t great but were still useful to students in Pakistan who might need a bit of extra help to understand a new concept. The translations, however, are getting better.

Indeed, every improvement in Google Translate automatically becomes an improvement to Marginal Revolution University. Amazing.


'Professional human-written captions have been produced for most of our videos in English, Spanish, French, Chinese and Arabic and we are working on more translations.'

Amazing how two GMU econ professors, a $4 app, and youtube are able to produce such impressive results - no cost disease there.

Unless, of course, one can look at the figures that someone like Roman Hardgrave sees. Who, and you can not make this is up, is General Manager and Chief of Product, Marginal Revolution University, at least according the to the page that google saw at on December 13th, 2016 09:24:32 GMT.


(For a second I thought we'd been Rickrolled)

The 80s were a strange place

"All your base are belong to us" - Google circa 2016

I suppose online education might save the universities by sending the dud subjects online and leaving the campuses dedicated to the laboratory subjects.

Does this innovation improve measured productivity? how does this improvement EVER get reflected in GDP???

Maybe not this one.

But if the cost of producing some service goes down, people might produce more of it. That's kind of like increasing production. Whether it's reflected into CPI or GDP deflator stats for a specific micro-item on the market is not at all clear, of course ...

There's also the theory that education today causes some things related to future production potential.

I'm guessing that it lowers measured GDP, as translators are no longer employable and they need to move to other, perhaps lower wages, professions. However, the increase to human welfare of anyone being able to translate almost any language is almost incalculable, much like search itself, which has transformed the knowledge economy.

Hey, econ professors: how about a thoughtful response to anon's question?

Inquiring minds want to know.

Wouldn't good, free translation lower the barrier to doing business in all kinds of ways, which would show up in economic stats?

I agree that I would like to hear the authors thoughtfully respond, though.

I use Google Translate to communicate in writing with my Greek lawyer, who speaks nor reads little English, since my written Greek is primitive. It's quite good. But for famous works of art, I bet humans are 'fine tuning' the program of Translate using Google Brain, using perhaps famous works of literature just as Hemingway's. The same trick is used by certain chess programs (they use actual brilliant chess games, known to be sound, to 'fine tune' things like positional exchange sacrifices, to make chess engines stronger). Now I will translate this passage from English -->Greek -->Japanese -->English and you can judge the mediocre results my friend!

Result (English to Greek to Japanese to English):
I can use Google Translate to communicate in writing in writing to my Greek lawyer, because I write written Greek is primitive, so I can speak a little English or read. That's pretty good. However, for famous works of art, indeed people are «fine tuning», translating program using Google's brain using possibly famous works of literature like Hemingway. The same tricks used in some chess programs (using truly brilliant chess games, to make the strongest chess engine, to be 'perfect' as a place exchange sacrifice, to be healthy It is known). Now translating this passage from English -> Greek -> Japanese -> English and you can judge my friends a small result!

Maybe it's because my English is terrible? LOL 私の日本人の友達を読んでくれてありがとう!

Is that what you do to all your comments?

1996 prediction: I returned from living in Japan a few years in 1995, studied Japanese while there, and after using the internet a few months, wondered how long it would take to translate on the side. I was sitting in front of a computer in a small lab and suddenly it occurred to me what Yahoo and the internet along with a thousand times more powerful computers, using a statistical approach, could do to human translation. I explained to a friend at a computer next to me: "There goes human translation!"

A German translator in the room heard that and said I was a complete idiot because computers could never think. I responded they don't have to think but internet content will balloon in ten years and using a statistical method, machine translation will be excellent, with many errors, at least for French, Spanish, German, English, and you'll start to see translation jobs disappear in the early 2010s. Arabic and especially Korean and Japanese are grammatically much further from English so the same thing will happen but with a five year lag.

In 2002, I was more specific and said Japanese into English translation will be mostly over except for literature sometime between 2008 and 2015. Unlike 1996, I could see in 2002 that Indians were translating at very low rates, often just 15 or 20 percent that of Americans translators I knew, but I underestimated how something like Google Translate would make any lower level/lower wage translator's job easier.

I watched the improvements in Google Translate from 2006 and saw clear jumps every two or three years.
By 2010, I realized my "by 2015" prediction would be off but emailed translator friends that "By 2020, a titanium wall goes up that no translator can get past - except the rare literary translation in Japanese to English."

By 2014, it had turned into a powerful grammar tool for the lower wage/less skilled translators. And the jump I saw in November that happened to be a couple of weeks after the new system went up was a clear jump from that. I think it will take just one more jump -not even a very big one - so that most Western translators will not find it worth their time to be MT editors at fairly low wages around 2019.

Thank you Todd for that reply. I find Google Translate translates simple sentences like this one pretty good. Google Translate fails were there's nuance or slang involved, dig it? But you are correct: by the year 2020 we will all be working for the machines, not working with machines, and loving it!

その返信をお寄せいただきありがとうございます。 私はGoogle翻訳がこのような単純な文章をかなり良いものに翻訳することを知っています。 ニュアンスやスラングが関わっているのであれば、Google翻訳が失敗します。 しかし、あなたは正しいです:2020年までに、我々はすべて機械で作業し、機械で作業しないで、それを愛するでしょう!

Thank you for your reply. I know that Google Translate translates such simple sentences into pretty good ones. If nuance and slang are involved, Google translation fails. But you are right: by 2020, we will all work on machines, not working on machines, I will love it! <-- not bad, Japanese back into English, Google Translate dropped the slang 'dig it' and 'Todd' automatically

"Google Translate dropped the slang 'dig it' and 'Todd' automatically"

I lived in Japan 15 years but at the end there was a small mishap - a slight misunderstanding - and... well... the government has erased me from all records after the "June 13 Incident".

I didn't start it.

But Google Translate keeps getting better and better every two to three years.

P.S. Note that GT messed up: "we will all be working for the machines, not working with machines" ==> "we will all work on machines, not working on machines" Also, Troll Me is right that backward translation isn't the best of tests.

It messed up, that's just what the machine wants you to think.

What incident? The only reference to June 13 in that part of the world was this: involving Paiwan aboriginals in Taiwan.

Translation just doesn't really quite work as A--> B and then B--> A as a quality check. Maybe sometimes for engineering documents or something, but not for most stuff.

If you're counting on a computer to figure that out for you, well, there's probably going to be a lot of messed up stuff along the way.

The point is, there will be less and less messed up stuff, and soon none at all.

*fewer and fewer

Communication just isn't that tidy. But definitely less for given situations.

However, the main observation is that as machine translation becomes more accessible, more people are using it and average translation quality plummeted at the same time as machine translation itself was improving.

And, the computer will probably never tell you when the original, not the translation, is the problem.

"However, the main observation is that as machine translation becomes more accessible, more people are using it and average translation quality plummeted at the same time as machine translation itself was improving.

Where are you getting this from? This is clearly not happening in Japanese/English or French/English.

Yet, you have not experienced Marginal Revolution University until you have read the captions in the original Portuguese.

I remember when "out of sight out of mind" when into French as "hors de vue, hors de l'esprit" and a computer translated it back as "invisible maniac".

AI has taken all the fun out of machine translation!

I think they could use autocaptioning for Youtube and ideally convert between formats seemlessly.

I'm actually impressed with the recent improvements in Google Translate. To suddenly hear it producing something that's actually intelligible instead of grammatical nonsense that makes it less useful than a online bilingual dictionary most of the time makes me hopeful that I'll live long enough for technology to grant me the ability to browse the non-English internet with as much ease as I browse the English Internet.

Now, if only using Google Translate without the benefit of a monitor and mouse wasn't so cumbersome(Google Translate's webpage isn't the most screen reader friendly, and precisely selecting text with a keyboard isn't the easiest task when you can see what's being highlighted).

When we were wandering around China, my girlfriend did quite well with her i-phone translation app. Generally she used the voice input; no typing needed.

Unfortunately nothing Google related works in China. Job security for translators there.

Right, until next year or the next.

The ass backward CCP is no match for the mighty Google.

I imagine the homegrown Chinese tech giants will be able to make their own version without too much trouble.

Google Translate's earlier achievement on Latin translation.

Regarding the mentioned very long "AI Awakening" article, this one paragraph caught my attention:

"Once you’ve built a robust pattern-matching apparatus for one purpose, it can be tweaked in the service of others. One Translate engineer took a network he put together to judge artwork and used it to drive an autonomous radio-controlled car. A network built to recognize a cat can be turned around and trained on CT scans — and on infinitely more examples than even the best doctor could ever review. A neural network built to translate could work through millions of pages of documents of legal discovery in the tiniest fraction of the time it would take the most expensively credentialed lawyer. The kinds of jobs taken by automatons will no longer be just repetitive tasks that were once — unfairly, it ought to be emphasized — associated with the supposed lower intelligence of the uneducated classes. We’re not only talking about three and a half million truck drivers who may soon lack careers. We’re talking about inventory managers, economists, financial advisers, real estate agents. What Brain did over nine months is just one example of how quickly a small group at a large company can automate a task nobody ever would have associated with machines."

There will be multiple shocks to the system resulting in long tem structural unemployment. The human suffering will cause political instabiliy.

Tthere will ne blood.

I wonder if Google Brain is doing one-to-one translation, e.g. from English-to-Corsican and from English-to-Kyrgyz, or if it's doing multiple-to-one given that there are several professional human-written translations already available, e.g. English-Spanish-French-Chinese-Arabic-to-Corsican. If it's doing the former well already, how much improvement can it get from the latter?

Comments for this post are closed