The first step in AI might surprise you

Cassie Kozyrkov
Oct 15, 2018 · 7 min read
Here’s the audio version of the article, read for you by the author.

Welcome to AI! Welcome to machine learning! Does it matter if you don’t know the difference? Nope, because you’ll start applied projects in them the same way.

What way is that? Perhaps surprisingly, not any of these:

  • Get an AI degree!
  • Hire an AI wizard!
  • Pick an awesome algorithm!
  • Dive into the data!

It’s a trap!

Image for post
Image for post
Image: source.

But first, why are these the favorites? It’s a story of comfort zones.

Your comfort zone can lead you astray

If you’re a business leader, your instinct is to hire someone who sounds qualified. Great instinct! Except the person best qualified to start an applied AI project is not your garden variety AI PhD. It’s… you! Whoops. Hire yourself first and read on to find out what you’re supposed to do before you bring your champion nerd on board.

Don’t ask a team of PhDs to, “Go sprinkle machine learning over the top of the business so… good things happen.”

Eerily familiar?

Or maybe you’re a data scientist. (Also a classic first hire, since today’s market thinks data scientists walk on water.) Perhaps you also have a PhD, but your Great Love isn’t methods. It’s data. Data data data! What data do we have? Let’s figure out what beautiful ingredients we can use!

Wait… use for what?

If you’re a data scientist or AI researcher and this sounds familiar, you just got handed a lemon by your leader. They let you down! Go on strike until they’ve done their part.

Image for post
Image for post

Begin with the decision-maker

Start here

The right first step is to focus on outputs and objectives.

Imagine that this ML/AI system is already operating perfectly. Ask yourself what you would like it to produce when it does the next task. Don’t worry how it does it. Imagine that it works already and it is solving some need your business has. (That’s why you needed those qualifications. Someone fresh out of a PhD doesn’t understand your business yet, so they’re not qualified for this task.)

Image for post
Image for post

The problem with the approaches discussed previously is that the order of operations is all messed up. The right way to approach an applied project is to flip the algorithms-inputs-outputs order on its head, like so: think about outputs, then inputs, then algorithms!

Your order of operations might be a mess.

A kitchen analogy comes in handy here. If you’re running a restaurant (as opposed to an appliance factory or food science lab), why would you think about buying — or, worse, inventing — a pizza oven before you’ve even considered whether adding pizza to your menu makes sense? That sounds like the rookie mistake of someone who doesn’t know what business they’re in. Instead, start with what your customers want and what food quality you’re willing to settle for.

Define success!

Image for post
Image for post

“All of them”? But surely you don’t want your police dog chasing sheep! Or vice versa, for that matter. A better answer is that it depends on what the owner wants. That’s you! Diving into algorithms and data before figuring out what outputs would count as good or bad behavior is a bit like putting a puppy in a basement with food and water, then being surprised what comes out isn’t good at being a police dog. You can’t expect to just sprinkle machine learning on your business, leave it brewing, and get something useful.

Image for post
Image for post
It took plenty of planning to get Peach this good at policing. He even writes witness statements! (This fist-crayon masterpiece comes from officers frustrated by a barrage of requests for an account from PC Peach… despite their having explained that Peach is not a person.)

Analytics might a better fit for you

Spend some time figuring out what looks promising enough to pursue, then come back to machine learning when you’re ready.

Besides, analytics uses some of the same math, so you wouldn’t be lying if you told your friends you’re using ML/AI algorithms (though you’re not building ML/AI systems). Many people who think they want ML/AI actually only need analytics. The latter is a great idea for all projects, while the former is good for only certain kinds. If you’re unsure, go for the sure thing.

Image for post
Image for post
Is this good behavior? I have an opinion and I’d hazard a guess that Fido here has one too.

Before you do anything else

The right time to think about your goals is at the very beginning, while your project is still a puppy!

What goes for puppies goes for ML/AI systems. To figure out what success looks like, you don’t need to understand how the puppy’s brain learns from sensory inputs. You don’t need to think about how those sensory signals are stored and processed (yet). What you do need is to figure out that you want a sheep dog (and what that means to you). To do your job thoroughly, you also need enough imagination to picture what behaviors you’re aiming for and what you’re trying to avoid.

Additionally, it helps to do a quick intuitive reality check: verify that relevant data is within your reach and that you have the hardware muscle to process it. If you’re training a sheep dog, are you confident you can get hold of enough actual sheep to show it? Even if you have sheep, your puppy’s brain needs to be able to take in and use information about them. If your “puppy” is actually a fly larva, it’s not going to be able to do good things with sensory data about sheep. (It can’t run in production either.) I don’t need to tell you that you’ll have a problem.

Image for post
Image for post
“What is my purpose?”

What’s obvious with dogs seems to elude many ML/AI teams I’ve seen. Some only ask what the dog is for when they retrieve it from the basement after a few years. Well, now you know.

The right step taken by the wrong people

It takes business savvy to properly think through what an ML/AI system is supposed to do for you and why it’s worth building. Focus on this first, before getting anywhere near the nitty gritty, including figuring out whether or not the algorithm that’ll solve your problem is considered AI or ML (you deal with that much later). If you have no ML/AI training, tackling this first part before you’ve hired a team or bought sci-fi kit might sound daunting, but I’ve got your back… here’s a step-by-step guide just for you!


Sign up for Get Better Tech Emails via


how hackers start their afternoons. the real shit is on Take a look

By signing up, you will create a Medium account if you don’t already have one. Review our Privacy Policy for more information about our privacy practices.

Check your inbox
Medium sent you an email at to complete your subscription.

Cassie Kozyrkov

Written by

Head of Decision Intelligence, Google. ❤️ Stats, ML/AI, data, puns, art, theatre, decision science. All views are my own.

Elijah McClain, George Floyd, Eric Garner, Breonna Taylor, Ahmaud Arbery, Michael Brown, Oscar Grant, Atatiana Jefferson, Tamir Rice, Bettie Jones, Botham Jean

Cassie Kozyrkov

Written by

Head of Decision Intelligence, Google. ❤️ Stats, ML/AI, data, puns, art, theatre, decision science. All views are my own.

Elijah McClain, George Floyd, Eric Garner, Breonna Taylor, Ahmaud Arbery, Michael Brown, Oscar Grant, Atatiana Jefferson, Tamir Rice, Bettie Jones, Botham Jean

Medium is an open platform where 170 million readers come to find insightful and dynamic thinking. Here, expert and undiscovered voices alike dive into the heart of any topic and bring new ideas to the surface. Learn more

Follow the writers, publications, and topics that matter to you, and you’ll see them on your homepage and in your inbox. Explore

If you have a story to tell, knowledge to share, or a perspective to offer — welcome home. It’s easy and free to post your thinking on any topic. Write on Medium

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store