How Not to Build an AI Startup
Building an AI using off-the-shelf technology should be a walk in the park. Programming languages like python have dozens of libraries to make your AI project a breeze. All you need is to:
- Ask a simple question.
- Have a mountain of clean, viable data that backs up a course of action.
- Get the AI optimize the course of action from great to perfect.
Except it is never super easy. Outside the park gates, things that are often missing include:
- The right question;
- The viable data;
- And the meaningful answer.
Having just one of these can be a miracle and on the rare occasion data scientists have one of the above, they will usually have to (1) guess at one of the other two and then (2) extrapolate the third. Guessing can lead to self-selecting a question, an answer, or the data that validates the guess even when the hypothesis itself remains wrong.
At Spotfit, a fitness startup I ran with some MBA buddies for 24 months, we had the question: what workout routines are best for strength, stamina, and flexibility? We assumed our data points: mood, mindset, weight, and physical capability. We then began trying to build data to validate the answer. And it all fell apart rather quickly. AI is super easy when you have all the pieces or all the time. We had neither.
We didn’t have the resources to get the time or the time to get the resources. Our proof of concept in the AI space got us to a question and a hypothesis, but never to data and an answer. Answers take time. Times takes money. What money takes, I am still trying to figure out. Our effort over 24 months building a framework to help people live healthier lives was worthless without something to show at the end. If we wanted it to work, we had to want to finish it even knowing it might still fail.
How to get Strong
We solved our AI dilemma by using a rule engine to generate data based on our assumptions. We guessed that tried and tested strength training techniques (slow, thoughtful repetitions, goal setting, etc.) would eventually get us a dataset that looked like it could prove or disprove our assumptions. We — and the overall the fitness space — believed that repetition and mindfulness can the move the metrics of weight and physical strength. As a startup, we competed against organizations that had datasets through industry connections. We had to build our dataset from the ground up, using assumptions we made as amateurs in the field. Getting strong is hard.
How to get Stamina
A rule engine is not an AI solution, but it is an effective means of creating a dataset. AI needs a question, an answer, and the data. Rule engines use the question and assume an answer to create the dataset. Often times, AI solutions start with datasets and look for questions and answers that can be combined within the data. Once a good question or a good answer is identified then its reciprocal can hopefully be found. Data is the key. Data needs to show a clear trajectory from A to B. It shouldn’t be a coin toss where the data provides the right answer 50% of the time. The question or the answer usually gets refined, but data is the secret sauce of most AI startups. One of the reasons why legacy corporations have such a hard time with AI is because they start with a question, no clear idea of the answer and very dirty data. Without data to back up an answer, there is no intelligence, artificial or otherwise. Data cannot draw the conclusions, it can only validate them.
How to get Flexible
AI systems with data can start with the answer and work backwards by finding trends in the data groupings then formulate a question that describes the results. A simple example is divorce rates. Data scientists know people get divorced. Data scientists know some facts about people in the people dataset. Data scientists can then work backwards to determine how like a person is to divorced people or how like a person is to married people. Finally, data scientists can then ask a question like how likely is this person to get divorced? For Spotfit and our fitness data requirement, we had some options.
1. We could buy data, but that would mean we would have to adjust our data model or massage the purchase (possibly into something it wasn’t).
2. We could create the data, but that would mean getting buy-in from a body of consumers without a product.
3. We could make the data up (synthetic data), but that had high probability of skewing the results to what we wanted the data to say versus what real-world results would be.
AI Did What?
The same month I started MBA school, IBM sent me on training in Philadelphia for a week. I had been avidly reading all about the latest tech trends daydreaming I would be working very soon in a executive role in blockchain or AI or quantum computing. I was about to get my MBA after all. That meant something.
We worked in small groups on projects like better presentation skills and design thinking. During one of these sessions, we had been discussing ways to build eminence online as part of extending our personal and company profile and writing blockchain articles came up.
“80% of the articles about blockchain are about how blockchain is like something totally unrelated to blockchain, then bookended with a sentence or two so that the reader knows the article is still about blockchain.” I said off the cuff.
The following day, we had a presentation on emerging technologies including AI and blockchain. For the AI component we signed into Watson and started feeding it rules using a web interface. For blockchain we played with three columns and private keys linking them together. But what is blockchain? the very young and synchronized presenters asked rhetorically. They then described an ancient Polynesian civilization whose entire economy was based on exchanging these sunken stones. Blockchain was exactly likely kai stones in the ocean.
Ted, turned to me after the presentation and said, “you’re right.”
I had no idea what I was right about. I was still trying to figure out where the Polynesians got the Kai stones from, where they were taking them to and what got them sunk in the first place.
When writing about AI, I’m by no means an expert, but I am right about data being key. Entire companies can be built upon a dataset asset through exploration (questions) and pivots (answers) into a meaningful business. AI is plainly using math to identify patterns in large data sets quickly to find statistical probabilities of an event occurring. Making AI work is having synergy between the question, the data, and the answer. The best way to accomplish that synergy is starting with the data.
When trying to make a go of it with Spotfit, I read a lot on how to build a data set, which statistical model to apply, how best to train an AI engine, what insights it can reveal, and how those insights can be applied. What I didn’t read very often is how hard it can be to get the data. AI research often assumes data is readily available for analysis. The research is about AI not data collection after all. We have been lulled into a belief that AI is just around the corner for society through whizzbang self-driving cars and pattern recognition in photos. Having actually set out in my spare time to build a rudimentary AI using off-the-shelf components, the ugly work of fulfilling the data component is often overlooked and underplayed.
If I had to do it again and try to start an AI business, I would start with data in hand, and I wouldn’t assume the actual business before getting an answer from my data. I would then build the business on selling to everyone struggling with the question. What we were doing with Spotfit amounted to climbing a mountain with no shoes and no mountain. It can be done, but it isn’t easy.
