Interview with Max Sklar from Foursquare: Bots, AI and Future Predictions
Max Sklar is an engineer and data scientist at Foursquare. As part of the engineering team, Max focuses on using machine learning and heuristics to develop new features and products. Most recently he has led the development effort of the Marsbot app, a bot that texts local recommendations.
Max has spoken at a variety of conferences, universities, and meetups, including Talkabot in Austin and ACM Recommender Systems. He holds an M.S. in Information Systems from NYU, and a B.S. in Computer Science from Yale.
You can also follow Max’s blog for learning more about his thoughts on data science, local recommendations, New York City and tech innovations!
Hi Max, thanks for taking part in this interview series. I’d like to talk a bit about you. You are working as a machine learning engineer at Foursquare. How did you decide to pursue your career in machine learning?
“I was always fascinated with the idea that instead of having to mastermind a program, we can just design algorithms that have the computers program themselves.”
After spending a few years as a software developer, I realized that machine learning was going to be the most exciting and engaging field for me personally.
Another interest of mine is building new products while figuring out how to get people to using them. This came through through working on software projects and particularly on Stickymap.com which was an early social-local site. I also have a strong interest in local recommendations, place-based data, and maps.
The common thread among all of these is the open-endedness of it. The most interesting real world problems are ones with endless possibilities where it is not feasible to explore the space of all possible solutions. This can be frustrating when your in problem-solving mode as an engineer, but I’ve grown to embrace it.
Can you tell us a bit about Marsbot project? Let’s walk through the algorithms and math you used.
Marsbot is a character in your pocket that learns about you based on where you go in the real world, and it uses that to text you recommendations about bars, restaurants, cafes, and things to do near your location. You never have to open the app after you set it up — the learning process is completely automated.
Once you use it for a few days, it will start texting you with different types of suggestions. Sometimes it will give you a suggestion on what to order at a restaurant when you walk in. Sometimes it will tell you where to go next when you’re done. You can also text it directly for recommendations and have a (simple) conversation with it.
We didn’t have a grand new algorithm or technique for this product. Instead, we incorporated all of the smart services that we had built at Foursquare to date. The purpose of the product is to both showcase Foursquare’s technology and to show the industry where search and recommendation could be headed in the next decade.
That being said — there are so many cool pieces of software underneath the hood. One is just Foursquare’s recommendation engine in general. Inside that is our ratings and sentiment analysis algorithm, which I worked on and have spoken about a lot. Another is our NLP stack and taste graph to pull out key terms in Foursquare tips for personalization. And finally underneath all of that is Foursquare’s unparalleled global venue database itself!
Key to all of this is Foursquare’s Pilgrim SDK which detects when you stop at a place and where you stopped. This piece of technology is very crucial to all of our products on both consumer and enterprise, so we put a lot of data science and machine learning resources behind it.
As for the chatbots, where do you think they are heading to? What parts of our lives are they likely to dominate?
They are already in our homes and in our mobile devices. At this point, we can take stock of what they have been good at and what they’re limitations are. In the home — hands-free simple questions that revolve around weather, time, and radio stations are great. A few games and novelty features of the echo and Google Home are cool. In my use, I’ve discovered a few limitations and I have put some thought into how this might play out.
Some of the more complex home-control tasks, such as as switching audio feeds to different rooms, finding smart ways to adjust volume, lighting, and purchases are still difficult. I think that instead of having people memorize commands, these systems ought to be smarter in figuring about what someone could possibly want. They can figure out smart ways to ask followup questions and also memorize past preferences. I think these features are going to slowly evolve over many years.
In particular, conversational search is something that I experimented with on Marsbot. If you ask Marsbot for a suggestion, you can tell it that it’s too far or it’s too expensive, and it’ll readjust. However, it doesn’t understand all the complex parameters in natural language. If it gets something wrong and you tell it why and it won’t understand. I think true chatbots should have that conversational and memory aspect to it.
Finally — the hardware is going to expand, and will include cars and eventually smart glasses.
One important thing to remember is that voice isn’t suited for every task. Visual displays with touch are very powerful and will often crowd out voice and text. One day, we might be able to command actions with simple gestures or even (in the far future) by thought.
Do you think that AI is our friend? What are your future predictions for artificial intelligence?
“In general AI is our friend because ultimately it is a tool that we’re going to use to better solve our problems.”
Like any technology, it can be used to harm us when some bad actor directs it to do so. Acts of destruction and theft are nothing new — but the people involved in those activities are going to have access to AI. That’s why I’m more concerned about AI being in the hands of bad actors than some science fiction scenario where robots turn against their owners.
That said, we can use AI to protect ourselves against this kind of stuff. For example, AI is used in detecting identity theft early and detecting malicious account logins. This arms race will improve our security technology over time and I’m cautiously optimistic that we’ll ended up better off than before.
In terms of applications — that’s everywhere in the economy. Just to name a few of the usual suspects: finance, recruitment, law, medicine, education, entertainment, marketing, transportation, research, product development, and military. It’s not going out on a limb to say that we’ll see self driving cars go mainstream in the 2020s, and more well-connected homes. The next step in local recommendations requires a lot of real-time knowledge that there isn’t enough infrastructure for, but I think we’ll see some of that emerge in the next 10 years.
Machine learning and data science are the pioneers of the next industrial revolution. How exciting is this for you?
It’s very exciting, but I think it’s even broader than just AI/Machine Learning. The explosion of decentralized and peer-to-peer applications (think cryptocurrency and Airbnb respectively) are just as exciting. The common thread here is “incentive engineering”.
When I build a machine learning model, I want to pick a loss function or a gradient that gives the model and incentive to make good predictions. If I have any models dependent on each other, I want to make sure that the full result will be good and if not some model will be directed to correct it.
It’s similar in these large networks where the winners are the ones that have figured out how to get all the agents to produce a good outcome for the whole. In general, this comes from many years of tinkering rather than one ingenious algorithm which ensures there will be lots of interesting work to do for the foreseeable future!
Botanalytics is an analytics tool for conversational interfaces. We will be very pleased to have you as our next guest for this enjoyable interview series. If you’d like to participate in our series, shoot an e-mail to hello@botanalytics.co !
And don’t forget to sign up for our newsletter to stay in the loop with the latest news!
