Unpacking the overloaded term of "artificial intelligence"

After attending IAC2024 earlier this month, I wrote a reflection that outlined "Just One CTA" for each takeaway. I gave myself the action to:

Write an article showcasing different ways "AI" is implemented on popular consumer platforms.

This is that article. And it's mostly for me. As a generalist, I have to really be adept at understanding and gaining alignment on domain areas. This means unpacking organizational structures, processes, rules, acronyms, and of course, words.

While I have been working with clients to craft and align on product and UX strategy, I am not technical expert at artificial intelligence. However, I do believe that we need to establish and maintain a common vocabulary to make conscious AI-related design decisions.

So I'll start by surfacing some definitions that some might believe are basic and boring. Getting alignment on boring and basic is usually a great start to a successful project.

"AI" is all over the place.

You can't miss it. You hear it on the mainstream news. Tons of companies across all industries are talking about how they've integrated "AI" or will be doing it soon. It's all over LinkedIn taglines. We've abandoned "VR", "NFT", and "Web3" for "AI".

We're even using "AI" to help us write Valentine's Day cards. No? Just me? Whoops.

For the most part though, the term "AI" is used today to describe Larger Language Models (LLMs) and Generative AI popularized by OpenAI's ChatGPT in November 2022. I say popularized because LLMs have been around for a long time. One of my favorite LinkedIn taglines on this one real expert's profile is: "Building LLMs since before they were cool." I love that.

In this article, I want to unpack the concept of artificial intelligence and explore what specific kinds of artificial intelligence are used in some popular consumer platforms.

History of the term

The beginning of artificial intelligence can be traced back to the infamous Alan Turing, who asks the philosophical question, "Can machines think?" in 1950.

6 years later in 1956, John McCarthy (then a mathematics professor at Dartmouth College) coined the term "artificial intelligence" at a conference and later describes it in an article in 2007:

It is the science and engineering of making intelligent machines, especially intelligent computer programs. It is related to the similar task of using computers to understand human intelligence, but AI does not have to confine itself to methods that are biologically observable.

where "intelligence" is described as:

...the computational part of the ability to achieve goals in the world. Varying kinds and degrees of intelligence occur in people, many animals, and some machines.

I think this broad description of artificial intelligence is still relevant today.

The range of "AI"

For the most part, academic and technical experts have agreed to categorize AI in three large buckets.

• "Narrow AI". Also known as "Weak AI", this is the only type of AI that exists to date. It enables the machine to perform very specific tasks with human direction.
• "General AI" or "Strong AI". This type of AI enables a machine to learn, assess, and take on new tasks without specific human direction. This type of AI would mimic a human brain and entirely theoretical at the moment.
• "Super AI". This type of AI is sometimes nested under "Strong AI" and would enable machines to surpass the cognitive ability of human beings in all facets: more creativity, better learners, more curious, more empathy, etc.

Fantastic!

Now we can be a little bit more specific about the majority of "AI" related items on LinkedIn. Unless you're doing fundamental research or holding philosophical discussions about machines surpassing humans in cognitive abilities, we're talking about "Narrow AI" or "Weak AI".

For the purposes of this article, we'll refine our scope to Narrow AI and continue unpacking.

Unpacking "Narrow AI"

We can describe "Narrow AI" by capability (Reactive Machines versus Limited Memory) as articulated by Arend Hintze, a Professor at Michigan State University. from.

We can also describe it from a technology perspective. This is the perspective that has helped me to draw better distinctions in consumer-facing applications.

Modern day applications of narrow artificial intelligence contain both "machine learning" (ML) and "deep learning" (DL).

Machine Learning versus Deep Learning

We should recognize the difference between machine learning and deep learning. This means being more curious and specific when we see the term "AI / ML" in the wild.

Broadly speaking, the difference between machine learning and deep learning is the type of algorithms used and the amount of human help with classifying the training data. Also, it's probably safe to say that deep learning is a subset of machine learning.

Traditional Machine Learning algorithms leverage structured and annotated data to make predictions. That means that we have to rely on a smart human to use their experiences, theories, judgements, and beliefs to impose their "obvious" classifications of data before sending it through the ML algorithm. This activity is sometimes referred to as "pre-processing." Some ML algorithms include linear regression, decision trees, random forests, and small neural networks.

Fun side note: The idea of "obvious" is shamelessly borrowed from Liminal Thinking by Dave Gray. This framework has given me words and constructs to collaborate openly with folks from drastically different upbringings, backgrounds, and cultures. Thanks Dave Gray! If you haven't read it yet, it's an amazing read and let's talk about it after you finish.

Deep Learning is a type of machine learning that can use structured and annotated data, but this is not a requirement. Deep learning algorithms can accept unstructured data and still yield predictions that "make sense". These algorithms rely on "deep" neural networks, which have more (100s) of hidden layers (depth) versus the 1-3 in small neural networks in traditional machine learning models.

As you might expect, we can go further to classify the different types of deep neural networks that combine different mathematical techniques and are sometimes optimized for specific use cases. Here are few, but there are so many more.

• Convolutional Neural Network (CNNs). Applications may include image processing, computer vision, speech recognition, and machine translation.
• Generative Adversarial Networks (GANs). Applications may include photo-realistic image generation, photo blending / aging.
• Recurrent Neural Networks (RNNs). Applications may include time-series language translation, national language processing, stock market price predictions.
• Transformer Neural Networks (TNNs). Applications may include computer vision, text generation, text summarization, chatbots. This is also the underlying network for OpenAI ChatGPT, where GPT stands for "Generative Pretrained Transformer."

For in-depth reading on this topic, here are a handful of resources:

• 4 Types of AI: Getting to Know Artificial Intelligence | Coursera
• Machine learning models | Databricks
• What is artificial intelligence? | John McCarthy
• What is artificial intelligence (AI)? | IBM

We can keep going down the rabbit hole, but I'll leave it here and switch gears to exploring how these definitions could be used to describe AI in some popular consumer platforms.

It's Spring time here in the DC Metro Area and traveling has picked up for the season. For this article, we'll attempt to better classify the type of "AI" in travel-related platforms such as:

• Hopper. Finding the best flights to desirable destinations
• Airbnb. Finding a local place to stay while on travel
• Yelp. Finding the best places to eat

"AI" at Hopper

Winter is finally over and you're tired of being inside. You want to fly somewhere fun, but since the economy is in an awkward state, you are very mindful of how much you can save. Enter Hopper.

Hopper is a travel booking app and online marketplace for booking flights, rental cars, and accommodations. In 2010, Hopper released their price prediction feature aiming to answer the question:

Should I buy plane tickets now or wait until later?

Since then, Hopper has been iterating on their offering by ingesting more data and using more parameters to tune the predictions. As of April 2018, their flight database was ingesting about 300 billion prices per month, roughly 10 to 15 million per day. Hopper also brings in demand data. That's the data aggregated from the searches you and I make on websites such as Expedia or Travelocity. In an interview with PhocusWire, the CEO of Hopper Fred Lalonde notes:

It turns out at a very high level that the demand - what people are asking for - precedes the airlines adjusting the rates. In mathematics it's called a leading indicator.

Based on this information, what kind of "AI" do we think this is?

If we use our definitions above, I think the artificial intelligence used in Hoppers price prediction capability started off as Narrow AI: Machine Learning (ML) Algorithms with Linear Regression or Decision Trees.

Over the years, I believe it eventually evolved to Narrow AI: Deep Learning (DL) Algorithms with Recurrent Neural Networks (RNNs). RNNs are type of deep neural network specifically designed for sequence-based data. In the case of pricing analysis, Hopper looks back in time for historical data (demand and previous pricing) to predict the future.

"AI" at Airbnb

Thanks to Hopper, you've got an amazing deal on round-trip tickets from NYC to Rome.

Let's say you want to blend in like a local. You want to immerse yourself in Italian culture. You want to shop at local markets and make fresh food in a kitchen. Sounds like you'll need to find yourself a home rental.

Airbnb enables travelers to browse, request, and book homes in residential neighborhoods to live in for short or long periods of time. In fact, your neighbors may be Airbnb hosts!

In November 2023, Airbnb released an "AI-powered photo tour". It allows for hosts to:

...instantly create a photo tour, which organizes photos by rooms to help guests understand the layout of their home. The photo tour is powered by an Airbnb AI engine that recognizes photos and assigns them to 19 rooms.

Okay. Got it. I upload photos, it auto-magically categorizes each into one of 19 rooms in the home. All made possible by an AI engine.

Based on the information provided, I think the artificial intelligence being offered is: Narrow AI: Deep Learning Algorithm with Convolutional Neural Networks (CNNs) because CNNs are basically built for use cases like image recognition.

"AI" at Yelp

You use your Hopper-purchased flights and get settled in your Airbnb amongst locals. Of course, now you have to find the local grub.

In January 2024, Yelp released a feature that summarized community reviews:

New AI-powered business summaries leverage LLMs to parse recent reviews and describe what you can expect at the business, helping you quickly find a business that fits your needs at a glance.

Generative AI using LLMs are good at summarizing the information they receive. In this case, Yelp is using the constantly growing data source of user reviews to generate a dynamic piece of text to describe the business. I would guess that these summaries are updated on a regular basis, as reviews come in.

Based on the information provided, I think that the artificial intelligence used in Yelp's new feature is Narrow AI: Deep Learning Algorithm with Transformer Neural Networks and Retrieval-Augmented Generation (RAG). In the case of Yelp, I would expect the addition of RAG to help contextualize the summary texts.

One AI box unpacked, which one next?

As we know, artificial intelligence is not new. The concepts, theories, and methods have been around for as long as we've had computers. However, thanks to recent research and development combined with the advancement of graphics processing units (GPUs), there is growing desire to apply "AI" everywhere and anywhere.

I get it. Generative AI is impressive and flashy. If companies aren't putting "powered by AI" in front of their marketing copy, then there is a risk that they will be left behind. It's a business move.

And since it's a business move, my hypothesis is that companies who make conscious and deliberate decisions about artificial intelligence will be the ones who rise to the top. The consciousness and deliberateness requires deep understanding of customers needs, workflows, and pain points. Classic UX work.

But it also requires team / company alignment on definitions, assumptions, and aspirations about "AI". And most importantly, it's requires the company to evaluate the impacts of a single "AI" feature to humans and their environment.

At a minimum, get your team to unpack the first (boring and basics of "AI") box before looking for problems to solve with "AI".

At Devellie, we help teams and organizations unpack complex problems to reveal inspiring solutions that are sustainable and worth-having. Let's be clear about your organizations' challenges and be mindful of whether or not "AI" is the right solution at the moment.