James Kaplan is the Co-Founder & CEO of MeetKai, an AI assistant that makes life easier through conversation, personalization & curation.
You initially began programming when you were only six years old, what initially got you excited with coding, and what languages did you learn?
The impetus was the Oregon Trail game built for Windows 95. I was hooked and would play it every day after school. I was in 1st grade, so there wasn’t much else to do! I started thinking of all the things I wish I could change about the game. When I was younger, I bought a copy of “Game Programming for Teens,” a basic coding book at a local bookstore. I was instantly drawn to it and quickly forgot my original Oregon Trail motivations. But I always loved games, and they acted as my motivation to learn new programming languages. Over the next few years, I learned Visual Basic when I was attempting to write a bot for NeoPets (it worked) and then PHP when I was 12 to start making dynamic websites. Even then, my motivation for learning PHP was to make money to purchase video games.
Your previous venture was operating an AI-based hedge fund, what did you learn from this experience?
My biggest takeaway was to throw away the notion that you can’t compete with giants in a crowded space. It can be very tempting to think that, in finance, anything profitable has already been tapped and captured by a giant player. However, I quickly realized that you shouldn’t overestimate your competitors. Laziness and organizational inertia prohibit new ideas. Other times, an idea or campaign may be too niche for a bigger company. Surprisingly enough, they might also think something is too risky to develop, so it’s not worth going outside of their comfort zone.
Could you share the genesis story behind launching the MeetKai AI assistant?
I was tired of finance. Everything had to be measured in terms of Sharpe ratio and PnLs. It removed a lot of the fun of technology, and was far removed from what originally brought me onto the path of programming. In late 2018 I spoke to Weili Dai, MeetKai’s Co-Founder, about what I saw in the tech landscape in general. One of my main observations was that the voice assistant space was stagnated. All of the key players held onto old approaches, and users weren’t benefitting from the tech. Nobody was willing to try new approaches, “since that isn’t how X” does it. There was no differentiation. If I started from scratch and threw away all (well, most) of the preconceived notions of “how to build a voice assistant, then I could fundamentally change the user experience. We started building an actual AI assistant as opposed to a voice-powered chatbot. The lessons learned in my previous venture coupled with Weili’s mentoring led us to found MeetKai.
What are some of the challenges behind building an AI assistant?
There are two classes of problems to building an actual AI assistant – user expectations and technical implementation. The first problem is overlooked, but applicable at MeetKai. Users are trained in what is and isn’t out of bounds when it comes to a voice assistant. In particular, they assume they have to search in a command-oriented fashion. We’re working towards training users to search in natural language. That enables much richer types of capabilities, such as using negations “Find me a Dwayne “The Rock” Johnson movie that is not Moana”. We can handle that perfectly well in everyday speech, but current voice assistants fail to answer us.
The second class of challenges is technical. This shows up in two sub-categories – search and understanding. For search, we are different from other virtual assistants in that we maintain our index of the content. While this allows all of the magic that makes us next-gen, it brings with it the challenges of running and maintaining a custom voice-first search engine. This is an area we are continuously innovating around. Language understanding is the second area where challenges are faced when building an AI assistant. For most voice assistants, this would mean being able to understand English text. MeetKai understands and supports 16 languages. This is not 16 times more work, since we use multilingual approaches, but it’s still a substantial amount more than being “English first, English only.” It’s worth the time investment though, as it’s incredibly important to us that MeetKai is truly global.
How does MeetKai use personalized AI to differentiate itself?
We use personalized AI in two differentiated spaces – understanding, and search.
When a user says, “Can you find me something Chinese tonight?”, that might mean they want a Chinese recipe, a Chinese restaurant, or a Chinese show to watch. With our personalized, deep understanding, we can disambiguate and provide a result for the user that they expect. This is all done without their data ever having to leave our platform.
We bake personalization into the search itself. One of the biggest problems other virtual assistants face is that many of the searches are fed to third-party fulfillment providers. When you search for a restaurant with a conventional assistant, it’s passing that search to Yelp. The downside of that is that Yelp doesn’t know the user personally, and if they do, then it’s a privacy issue. Because MeetKai is a first-party app, all the way to the result, we have true personalization.
What are some use cases for MeetKai?
MeetKai’s goal is to be the first AI Concierge. We want to help users in their everyday life. We don’t want to be a command-oriented assistant. We don’t support features like, “volume up”, “volume down”, “30-second timer”,…you get the picture. I feel strongly that those features are not AI, they are just voice-based input. If you think back to the days of AskJeeves, the entire idea was that it was your butler, no one would ever anthropomorphize Google Assistant. As much as Apple Siri’s or Amazon’s Alexa would like to be, no one thinks of them as anything other than an app. We are still in our early stages, but we built out all of the necessary technology to carry out our three to five-year roadmap towards a true AI assistant.
How can the industry prevent an AI assistant from introducing bias or reinforcing existing bias in a user?
There is a delicate balance between providing personalized results and not creating bias. We have seen what happens when researchers optimize for engagement measurements on social media platforms – it creates bubbles of bias. The first step towards this as an industry is to rethink our metrics and KPIs. At MeetKai, we optimize for a balance between click-through and novelty of the click. The AI should be rewarded much more for finding a novel result that is clicked on in 30% of cases, rather than simply the “top” result clicked on in 50% of cases. However, this approach does have a pretty apparent fallback upon deeper consideration. What if the results the AI is coming up with are just biased results for that user’s small bubble? We build our AI to push in both directions of a user’s personalization zone. If a user asks for a list of articles about beef, then instead of presenting them with articles that match their beliefs, we may sprinkle in articles that are at the edge, or slightly outside the edge, of their preferences. This may include an article on animal ethics and climate change, as well as an article about the potential health benefits of animal-based fats. The technical intuition for the approach we have taken is rooted in the acceptance that it is hard to train AI to determine if Y is biased, but it is much easier to train it to know that X and Z fall on the edges of the same zone as Y.
Where do you see the future for AI assistants in 5 or 10 years?
AI assistants will move more and more to the first-party-based approach that MeetKai has. For the entire existence of their lifetime, Alexa and Google have tried to cultivate an ecosystem of third-party extensions and skills. This runs into some serious problems in privacy. Furthermore, this is saying nothing about the upper bounds it places on capabilities. I expect more industry players to adopt the same approach we have taken, and there are signs of this across the board.
Thank you for the great interview, readers who wish to learn more about this personalized AI assistant should visit MeetKai.
- Attention-Based Deep Learning Networks Could Improve Sonar Systems
- Cerebras CS-1 System Integrated Into Lassen Supercomputer
- Deepfaked Voice Enabled $35 Million Bank Heist in 2020
- Facebook: ‘Nanotargeting’ Users Based Solely on Their Perceived Interests
- IBM Announces AI-Driven Software for Environmental Intelligence