Peter Norvig, the Director of Research at Google, who has a penchant for colorful shirts, is one of the world’s leading authorities on artificial intelligence. He used to think he was working in a dry, technical area – but then it turned out he was in the ‘search’ business.

‘It’s amazing. I came here because I thought it was an interesting place from a technical point of view. I was working in this very narrow, specific domain. Now it’s this thing that everybody in the world knows about.’

For all the fascination with Google’s multitude of other initiatives at its core remains the web search, one of the most popular destinations on the internet. Behind the familiar and seemingly unchanging search box, Norvig and his team are constantly running a series of experiments to improve the quality of the results it throws up.

The company is all too ready to admit that its own search for the Holy Grail, natural language, or semantic, search, which will truly revolutionize the way we find information on the internet may still be some way off.

Norvig joined Google in 2001 first as director of search quality and four years later became its director of research where he now oversees a hundred or so researchers who investigate subjects that range from networking to machine translation – arguably the world’s leading Machine Translation team.

He says that teaching a computer to understand languages isn’t rocket science – it’s not nearly that easy.

‘In physics, we’ve been able to use computers very well for a long time. We can get our spacecraft to the moon or Mars very accurately. But the problem with language is there’s lots and lots of rules and there are lots and lots of exceptions to those rules.’

To illustrate the point around two years ago  Google took a different approach to teaching a computer how to understand languages, which is more like the way humans learn them. It sounds a fairly simple strategy which comes down to programming a computer to learn through examples – by exposing it to an abundance of texts in a specific language, it can learn to pick out patterns.

‘Most of the answer to how you do this is counting – it’s just the fancy phrase for counting is ‘probability theory.’

He predicts that one of the advances in the next decade will be that most internet searches will start with spoken rather than typed commands and  also expects experiments to begin into monitoring individuals’ brain signals to search the web.

“Why is it that
Wikipedia is
flourishing while
the Open Directory
Project is floundering?”

Norvig joined Google from NASA where he had been the head of the Computational Sciences Division at  Ames Research Center, helping to develop the Remote Agent experiment that flew on the Deep Space 1 spacecraft – the first use of autonomous planning, scheduling, and fault identification onboard a spacecraft. It won the 1999 NASA Software of the Year award and was cited in two AAAI Presidential addresses as one of the top achievements in the history of Artificial Intelligence.

Between Google, NASA, Norvig was at Junglee, the comparison ads and shopping site and Sun. He has also served as an assistant professor at the University of Southern California and as a research faculty member at the University of California at Berkeley Computer Science Department, from which he earned a Ph.D. in 1986 and the distinguished alumni award in 2006.

He has over fifty publications in Computer Science, concentrating on Artificial Intelligence, Natural Language Processing and Software Engineering. He also co-wrote with Stuart Russell, Artificial Intelligence: A Modern Approach, which has been the leading textbook in AI since 1995, selling over 200,000 copies.

RM:

Peter, Google recently said it’s working on a news format called Living Story. The plan, if I have this right, is that it will be adopted by major news outlets worldwide in a bid to standardise the way information is used. But if everyone takes up the new product won’t this lead to news looking hackneyed?

PN:

Living Stories is an experiment we are trying with a couple of papers. Journalism has always been a mix of stereotyped formats (headlines, by-line, lead paragraph, etc.) and varied content. We don’t know if the Living Stories format will thrive and be picked up by other papers. But if it is, they will each have their own unique implementation, and I don’t see why it would be any more constraining or hackneyed than other format elements, such as headlines.

RM:

How do you know when a development is right? Does it hit you right the first time, or are you critical as you go along?

PN:

Always critical, always looking for continual improvements. For developments that have online users there is the advantage that you get to keep tracking how the users like the product. You can launch early and often, monitor feedback, and make changes to try to improve the product.

RM:

What happens when you finish one development? Is the next project one that has been waiting in line? Or is the choice more spontaneous?

PN:

There are always more ideas than we can tackle at one time. The choice of what to do next is a combination of the company goals, the personal interests and skills of available engineers, and schedule fit: for example, if a proposed project X relies on components Y and Z, and if we are currently working on a new version of Z, then probably we should wait until that work is settled before working on X.

RM:

As the company gets larger, does it get harder for people to push their ideas through and then get them implemented? How do you prevent inertia taking over and slowing everything down?

RM:

Yes, it is harder to launch a product in a big company than in a start-up.  This is not a Google issue in particular; it strikes all growing companies.  When you’re a start-up, you rush an idea out because what do you have to lose?  Nobody will remember or care if the product fails. You can take a prototype running on one engineer’s desktop, throw it on a server, and see if people like it.  If they do, you can improve the product gradually as your user base grows. With a bigger company, there are many more concerns: scalability (you could get millions of users, not dozens, on the first day), security (you need to worry about cross-site scripting and other attacks across your whole suite of products), reputation (a poor product will be panned, not ignored, thus tainting the whole company), and others. On the other hand, a big company benefits from economy of scale. A big-company product that is a hit might reach 100 million users in the first few months, while a similar product from a start-up might gather only 100,000.  The big-company product is harder to develop, but it isn’t 1,000 times harder.

RM:

What areas in development give you the most trouble?

PN:

One area that has proven vexing is understanding networks of people: for example why is it that Wikipedia is flourishing while the Open Directory Project is floundering? Why is Orkut the leader in Brazil and India, but trailing in other countries? Why has Facebook overtaken MySpace in the US? Is it something about the look and feel of the sites? Is it the policies they have for interactions? Or just luck in terms of who captures enthusiastic early adopters? These questions are beyond the current state of understanding, which makes this a difficult area.

RM:

What does Google’s extraordinary progression mean for the company’s future? Surely people will expect you to trump yourself each time another app comes out how do you manage that sort of expectancy?

PN:

We know there remains a lot that we can do for our users and we are glad to have the opportunity to provide new capabilities. I don’t think that every new product needs to be better than all the others, but a new product does need to be useful and easy to use.

RM:

Has any Google project been scrapped due to cost?

PN:

Every engineering solution must make trade-offs on cost: storage, computing resources, bandwidth and latency are all important factors, and a project has to meet requirements on all of them before it can be launched. For most projects it is easier to get a working prototype than it is to turn that prototype into a hardened solution that is efficient and cost-effective.

RM:

Google attracted a lot of flak for its decision not to use the GNU General Public License (GPL) for Android and to go with the Apache License instead. Do you read what critics say about the company and has what they said ever influenced your work or change your view of yourself as a developer?

PN:

Parts of Android (such as the kernel) are under GPL while the majority is under Apache, and individual applications made by developers would fall under their own licenses. The Apache license is more permissive, allowing more companies to participate. There would be some criticism of any choice of license, but the choice of Apache seemed to be the best to satisfy the most people, as many outside analysts have confirmed.

Certainly it is worth listening to criticism to consider what we could do better. If we’ve made a mistake, or if there is a good idea we haven’t thought of, we want to hear about it so we can improve. But keep in mind that improvement means doing what is right for users; it doesn’t mean just avoiding criticism.

RM:

In 2003 you had the idea of rendering the Gettysburg Address as a PowerPoint presentation, and the result is one of the sharpest pieces of satire to appear on the web. After that did you think of an app that could do away with PowerPoint once and for all?

PN:

Honestly, I have no strong feelings for or against PowerPoint. What I was objecting to is certain ways of using PowerPoint, which lead to boring presentations with a low content to noise ratio. Maybe there are applications that can help, but I think mostly it comes down to the person creating the presentation, not the tools. On the other hand, it would be useful if the tool would separate out the phases of (a) taking notes of material to possibly include, (b) turning that material into a viable presentation, (c) practicing the presentation, and (d) delivering the presentation. PowerPoint tends to merge these phases together.

RM:

Do you think narcissism is a necessary quality for working at Google?

PN:

I would say it is a disqualifying quality. The necessary (or at least highly desirable) qualities are the ability to work with others, to evaluate evidence to find the best solutions (rather than just push your own ideas), and to be willing to change your mind. It is also good to have the confidence to tackle a big problem by yourself or in a small team.

RM:

You’ve had an extraordinary career and I was reading one luminary who said that you are one of the smartest people he’s ever met. Who would be your choice?

PN:

That’s a tough one.  I could quote a certain ex-governor and say “all of them.” It is hard to compare people across fields, so I’ll stick to my field of AI and go with Judea Pearl. He revolutionized how we think about uncertainty in AI and proved to be very strong at three very different tasks: First, he was able to work out some of the important technical mathematical details. Second, he used this to address underlying philosophical issues, such as the notion of causality (as opposed to just correlation). Third, he was able to popularize the ideas, to teach and convince others to the extent that the basic ideas went from being extreme to mainstream in the course of a decade or so.

If you take out the requirement of “ever met,” and allow someone I’ve just read, I’ll go with Pierre Laplace, the French physicist. He didn’t get everything right, but he was far ahead of his time. 

And as for my own level of smartness, I’m reminded of the comment of Herb Cain, the former San Francisco Chronicle columnist: after The New Yorker did a restaurant review proclaiming a certain San Francisco restaurant the best Chinese restaurant in the world, Cain wrote that it wasn’t even the best on the block.  At Google, I’m feeling pretty good if I’m the second or third smartest in the room.

RM:

To wrap things up, can you give me an example of a preposterous lie that tells a great deal about life?

PN:

How about ‘You have free will.’ Many philosophers and physicists would say that this is either a lie or at least that most lay people have a vastly oversimplified understanding of the concepts “you” and “free will.” And yet it is this illusion that allows us each to go on living our lives every day and doing good.