Why AI Won't Be Awesome on Day One (But Will Go Beyond Awesome Later)

Posted by Alyssa Verzino on Dec 7, 2018 2:59:28 PM


The phrase "artificial intelligence is like lily pads in a pond" sounds like a really bad Zen koan, but it is an important concept for properly setting expectations with modern AI solutions. If you expect your business AI solution to be awesome on Day 1, you're setting yourself up for disappointment and, possibly, failure. The lily pads explain why.

A classic math brain teaser tells the story of lily pads in a pond that double each day. If it takes 30 days to go from one lily pad to a fully covered pond, on what day was only half the pond covered? Most people answer 15 days, but the correct answer is 29 days, as a doubling rate each day means just a day ago the pond was only half-covered, the day before that only a quarter, and as recently as a week ago only 0.8% of the pond was covered with lily pads.

So, what does this have to do with AI?

Most artificial intelligence solutions have similarly exponential growth curves, and the "just okay" AI system you deploy today will likely stay relatively adequate until your metaphorical "Day 27," and shortly later your business pond will be overflowing with AI awesomeness.

Now, lily pads thrive in ponds because ponds have what lily pads need: nutrient-rich water, sunlight, and air. What AI systems need to grow is well annotated training data.

Training data is information that deep-learning algorithms use to self-generate their programming. For example, when Google's reCAPTCHA security asks you to select all images with a street sign, it is assigning you a task that conventional software can't yet consistently do (proving you're a human) and developing training data for Google's own image recognition system These tests annotate the regions of an image that humans think contain a street sign, so Google can develop a wide range of annotated street sign images to train, say, self-driving car navigation systems.

It should say something that even Google, with its billions in cash and petabytes of information, still needs help annotating data. Even Google's AI lily pads need help to grow.

Most businesses have lots of data about their operations, but many don't have well annotated data. If, for example, you wanted an AI system to scan your email archives to learn best practices for handling customer contacts, have you identified who is the customer and who is the sales rep in each email string? And have you noted which customers did and did not leave the email chain satisfied? Without that annotation, the email conversation is just noise.

Once you identify what your AI system needs to thrive, you can adapt your workflows to annotate go-forward data as a matter of course. In fact, the best AI systems will help you do that after they are installed. But your historical data won't be helpful without a lot of work (just ask Google).

Thus, the time is now to start seeding your pond with training data and growing your AI lily pads. They won't cover much at first, but after a while, the gains will be noticeable -- and improve quickly from there. You aren't buying a fully covered AI pond; you're buying your first lily pad and directions on how to cultivate it. To expect otherwise is to misunderstand AI.

Talla is building an AI solution to improve customer support, and that system will grow and thrive as it is exposed to more and better annotated customer conversations. If you'd like to start growing your own awesome AI solution, and start moving towards your metaphorical "Day 30" of full AI support coverage, contact Talla today.

New call-to-action

Topics: AI, future of ai