Monday, April 4, 2016

Interview with Rachel Law, Founder of Kip


Rachel:
I'm Rachel Law, I'm one of the founders at Kip, I'm the CEO, and I'm very happy to be a part of your interview today.
Michael:
Great. Can you tell me what your company does?
Rachel:
Kip is an artificial intelligence chat-bot that helps you and your team with group shopping. We focus specifically on group shopping, both on enterprise and consumer side. We define group shopping as shopping with 3 or more people.

For instance, when an office needs to order lunch delivery, there’s normally one person, the office manager who does all the ordering. They have to go around the office asking people what they want for lunch, whether they want brown rice, or white rice or whether they want soup or salad. It’s tedious, time-consuming and additional responsibility for the office assistant.

Kip is assists these organizers to make their lives easier. They can just ask Kip, and Kip will ping everyone in the team, collect and coordinate the orders and send the office manager a single checkout link.

At home, there's normally one main caregiver. They manage the spouse's needs and the children, and maybe they have elderly parents as well. All of the responsibilities are directed to this one person. We came up with Kip to assist these people who have too many demands.

Michael:
What's your background?
Rachel:
I first started working in art. I was one of the early artists who started working with digital art back in 2011, and physical computing, so I'm quite familiar with robotics. Afterwards, I did my masters in design and technology. For my thesis, I built a program called Vortex that lets users swap their meta-data across networks so they could pretend to be other people or invent new identities to get better shopping deals.

This project went nation-wide. It was featured in Ad Age and KPCC California radio, as well as some other places. I worked a bit in advertising and then my co-founder Alex and I started Kip together.

My background is mostly e-commerce. I was running an e-commerce site since I was a teenager. I did it for 10 years. We had 12 million daily visitors and $6000 in transactions a week. When we formed Kip, the data corpus came with me so we could train our AI, and some of the profits helped start the company.
Michael:
So what caused you to start your company?
Rachel:
A lot of factors. Firstly, we felt that we had something very interesting and a different approach to artificial intelligence. We saw that a lot of AI companies were very very focused on what we call, “concierge services” and these AI were basically made for a small subset of wealthy people.

Second was that some of these AI startups were focused on starting with humans first.They had call centers of people running their artificial intelligence. So it wasn't really artificial intelligence as much as human automation. In future they would use it for machine learning training, but at present it was all human powered.

We were like, "Well maybe we don't have to do that. What's the lowest modicum of intelligence that people were willing to use for group shopping?”  Shopping is a great way to validate an idea, because when someone is willing to put their money, it means that it works.

So we started with a very stupid kind of chatbot. Barely any intelligence, no reinforcement training, no training, no nothing. Just simple simple rules. Then we put it out there, and people started using it. It was like, "Okay, so people don't actually care that much about intelligence, as long as it works." That's when we started developing and refining. How do we optimize the situation so that it can get usefully smarter and smarter? That's when the artificial intelligence started kicking in.

This was driven by people who would give us all kinds of queries and shopping requests. If you're just using a very standard library, you're going to run into trouble because conversations use short and truncated sentences compared to paragraphs in doc2vec. That's where we started doing what we call intent optimization.

Shopping is an interesting thing. Most of the time, the purchase decision comes after much evaluation. The point of sale comes after you browse the product, you've gone through the evaluation process, and then at the a sale. It's the opposite of a conversation. Most of the time a conversation is when you ask a question, the first line of answer is the most important piece of information, then you get more and more detailed as you go along.

That's what they teach in news journalism, it's an inverted pyramid for information.

How do you reconcile the fact that people normally shop after much evaluation? At the same time, people expect the first answer by Kip to be the perfect one. What we did was start feeding data for reinforcement training. Let's say you want to buy chocolate:
 
You: "Hey, I'm looking for chocolate."
Kip: "Here try these three chocolates, which do you like best?"
You: “I like the second one, but do you have it in dark chocolate?”

Kip will show you more choices, and you'll go: "I like the third one. Do you have it under ten dollars? I like the 1st one, do you have it gift wrapped for me?"

Maybe it takes a total of eight steps for you to complete the purchase, after much evaluation. What we do is we take the customer's steps in the conversation, and then we feed it back to the system for training. So that when someone else searches for the same term, "chocolate" on the same day and they have similar preferences to you, we can optimize the result higher. Instead of taking eight steps, it’ll take only four steps in order for you to find what you're really looking for. It's a kind of best-match system where we are using people's behavior to give the best result. The same way that people use the rating system, except that the ratings system is calculated for a personal match.
Michael:
Can you say something more about your company's AI and how it works?
Rachel:
Yep, I guess it depends what kind of AI you're looking at. We've been working on this for the past two years. It's kind of broad, if you could be a little more specific on the kind of AI you're specifically interested in, I could tell you how it works.
Michael:
Sure, let's talk about NLP first.
Rachel:
Okay. On the NLP side, one of the things we first wanted to see, "Okay, how do we narrow down user language to exactly what we want it to?"

NLP is a boundless problem. It's one of those things that will never be solved because every time you're trying to build something, someone will have come up with a new slang, a new dialect, another idiomatic language factor. They key to solving NLP problems is to make sure that our NLP problem is super super clear, and has very clear problems to solve.

For us, what we wanted to do was fulfill a sales transaction. Sales transactions, is a very structured kind of conversation. The two main driving factors in the sales conversations is price and availability. Most people will ask are questions centered around price and availability. So that's the first thing you want to consider. The second part is how will they ask these questions? Depending on which platform you're launched on, the way a person will phrase the question will affect how the responses that Kip can give. In language, it’s known as a linguistic register.  Your question and answer system needs to be adapted, designed, and centered on the platform you're launching it on.

For instance, if you're launching it on Slack, the kind of questions that come in are normally formally phrased, very technical and relating to product information because the user's more interested in a price comparison or a product review. "Is this the best monitor that we should use for everyone in the company?" The users on Slack are mainly enterprise, so the type of purchase is a corporate purchase.

Whereas if you launch it on something like Kik, you'll see a lot more slang and idiomatic phrasing that are not related to language as we think of it. We started building our universal emoticon translator for this reason. We have a system that translates emoticons or emojis, across the platform. So you can actually use emojis to search for things. Instead of typing a question, you can just send an emoticon or combinations of emoticons, and it would work as well.

We found that to using emoticons is a lot easier for some people because they’re used to thinking of these as language. So instead of text or typing, they would send these to their friends. We adapted that form of communication into Kip. That’s the idea of language, the easiest and most common types of communication.



Then on the machine learning side, you take all these chat conversations, especially  things that we don't understand. These flagged things are tested at the end of the day by me or by my co-founder Alex. Then we feed it back into the system so the system improves.
Michael:
How do you measure the quality of the NLP?
Rachel:
That's a very good question. We don't  measure it like you'd use for Image Net for accuracy. We have a baseline, which is what we originally started with when we had a stupid system, that it just gave generic responses for everything.

For us, the accuracy is how many things don't get sorted into the response buckets. Each bucket is like a micro-service. When a user initiates a conversation with Kip, all of the users' conversations get directed into different buckets.

We measure is how many things don't fall into the bucket, and whether they went into the right bucket. If it doesn't fall into the bucket it means that it's something that Kip does not understand, if it goes into the wrong bucket it means that Kip has a misunderstanding. One is an ignorance, the other is poor or not enough information.

Does that make sense?
Michael:
Yep. Can you share something awesome or surprising that Kip has done that has impressed you or one of your customers?
Rachel:
Probably the emoticon and image search. Most people don’t consider these things as important or part of language, but it’s very interesting because language is constantly evolving. As they say, an image tells a thousand words.

When I say emoticon search, I really mean it as a combination of a text, text as in Roman alphabet, pressing I'm sad, then you give a funny emoticon. Normally that would be considered a conflict of two ideas. Because why is there a funny picture next to an I'm sad image? But Kip can actually parse that. If you actually go on Kip and try to search combinations of emojis, you would get results that are relevant for what you're searching for.

If you think about it, that's where language is moving toward. People are using emoticons as a way to express themselves. You look at Instagram, half the bios are all full of emoticons.
Michael:
When you talk to emoticons, do they care about the AI, or do they care about the fact that you're solving a business problem?
Rachel:
It depends what kind of investor you're talking to. I know it sounds like a terrible answer, but it really does.

Most conservative investors generally does not invest in AI, simply because they know that AI is a black hole, where you can keep throwing money and you might not see anything returned. Facebook is already facing a bit of this, and so is Google, where they've realized that you can be working decades, and pouring all their resources into this and it might not actually turn out to be anything.

In that sense, because we have a very specialized form of artificial intelligence that's when they start caring. The fact on the day that we launched, we were already generating revenue helps. We do a very specialized form of artificial intelligence. So we don't build robots, we don't make driverless cars, we don't care about philosophical questions. Sorry. We don't really care about what's the meaning of life, who should live or who should die. These are the kinds of questions people like to ask general AI's. We're not interested in taking the Turing test either, we don't think it's not that useful. We only do one thing, and we do it well. We help people with group shopping. That's kind of our focus.

Having a very clear focus means that we can develop a very specific AI, that is intelligent in it's own way. That makes it interesting to investors. There's that very clear benchmark of what we're going to achieve. It is a very clear benchmark of what is considered intelligent. That's something that people rarely think about.

What would you consider as intelligent? How much intelligence is necessary? What kind of intelligence are we creating? You need benchmarks. If not, you're going to get a team of data scientists and they're not going to know what they're measuring for.
Michael:
Why is there so much excitement about AI startups?
Rachel:
I think there's a lot of interest in artificial intelligence startups because they see the future of work being automated. On a very practical, capitalist sense, I think that investors are interested in artificial intelligence because it's going to be the way forward the way Henry Ford revolutionized the assembly line.

Except instead of using human workers on the assembly line, they're using artificial intelligence on the assembly line. Which means that a large portion of work can now be automated, and goods will be much cheaper to produce. Because an AI doesn't need to eat, doesn't need to sleep, doesn't require healthcare, and all that kind of factors.

The other thing about AI, is that artificial intelligence's aren't as artificial as people think they are. At the end of the day, they still trained on human data sets. They're artificial in the sense that, yes they're not nature-born and they're not human, but they're humanly intelligent. They're an aggregate of all of our intelligence, specifically pointed on one field. You can think of them as a kind of super-human collective knowledge point.




Michael:
Where do you go to get information about AI startups?
Rachel:
Right now,  there’s so many AI startups that you don’t have to find it. It’s in the news every other day. There’s probably there's another new AI startup on Angel List as we speak. AI is in the news constantly. Finding an A-I startup is not difficult. It's finding who's doing things that are interesting, and who's doing things that are actually useful, and who is just creating vapor-ware or fear-mongering.

There's so much news about AI,  that it’s overwhelming. That in itself could be one AI, an AI that can curate useful news. Everyone posts mis-information now.

On our side, we have an AI meetup group called Deep Learn NYC. We host networking events and invite industry speakers every few months to talk about practical applications of machine learning and deep learning.
Michael:
Make a prediction of AI and startups.
Rachel:
In the future ... AI startups or AI?
Michael:
AI startups, AI and startups.
Rachel:
Well that's two different questions actually. In the case of artificial intelligence, it’s inevitable. We are going to have a future with A-I whether you are on board or not.

The reasons for this beyond just technology. We as human beings have constantly tried to create things in our image. It's a kind of compulsion almost. We're fascinated with the things that we can do. We are fascinated with whether we can create a version of ourselves, and challenging our own identity is human beings. It's almost like a God-like thing to do. We try to create artificial intelligence.

Whether you are on board or not on board the artificial intelligence train, it will happen. It's just a matter of time.

In the case of AI startups, if you're not already on an AI startup and you're thinking of creating one, unless you have a very very unique angle and approach or proprietary data it’s not going to be easy. It would be very difficult to get in the game. Now that everyone is talking about it, and it's now an open field. If you started a few years ago, now you’re hot and you have a competitive lead. If you start now, you’re not only behind but you’re competing with everyone who read yesterday’s news. The investors are almost over it. It's like the stock market. The moment laypeople know, "Okay it's time to buy Apple shares," it means that large portion of the momentum is already gone. That whoever started it two years ago, they've already scooped up the resources. If there's already investment, it's probably already committed to somebody else.

It's going to be very very competitive, similar to 1950s manufacturing. A lot a lot of people will try, and then a lot of people will fail, and there will be a market consolidation probably 5 - 6 years down the road. You’ll have a handful of companies who are entirely dominant in one field of AI. There will be an AI startup for healthcare that will manage all aspects of health care. There will be an A-I startup in banking that will manage all the predictive trading, and so on.

You will get that kind of market consolidation simply because it doesn't make sense to have multiple A-I's companies. Because of artificial intelligence are best in large quantities of data, and having too many AI companies is a duplicate of resources.

If everybody has a small data set, then it makes sense one competitor, whoever is going to win it, to acquire all these smaller data sets so that you get a very very large user base, and then you get a very very large access to information so that it actually makes artificial intelligence worth it. Then you can really scale and be extremely defensible, because it would be difficult to duplicate.
Michael:
Great. Thank you so much, I really appreciate it.

No comments:

Post a Comment