Businesses are jumping on a bandwagon of making something, anything that they will launch as a “Generative AI” feature or product. What’s driving this, and why is it an issue?

I used to be recently catching up on back problems with Money Stuff, Matt Levine’s indispensable newsletter/blog at Bloomberg, and there was an interesting piece about how AI stock picking algorithms don’t actually favor AI stocks (and likewise they don’t perform all that well on the picks they do make). Go read Money Stuff to learn more about that.

But certainly one of the points made alongside that evaluation was that companies throughout the economic landscape are gripped with FOMO around AI. This results in a semi-comical assortment of applications of what we’re told is AI.

“Some firms are saying they’re doing AI once they’re really just attempting to determine the fundamentals of automation. The pretenders will likely be shown up for that sooner or later,” he said. …

The style and apparel company Ralph Lauren earlier this month described AI as “really a very important a part of our . . . revenue growth journey”. Restaurant chains resembling KFC owner Yum Brands and Chipotle have touted AI-powered technology to enhance the efficiency of ingredient orders or help making tortilla chips.

Several tourism-related businesses resembling Marriott and Norwegian Cruise Line said they’re working on AI-powered systems to make processes like reservation booking more efficient and personalised.

Not one of the examples above referenced AI of their most up-to-date quarterly filings, though Ralph Lauren did note some initiatives in broad terms in its annual report in May.

(from Money Stuff, but he was quoting the Financial Times)

To me, this hits right, though I’m not so sure they’re going to get caught out. I’ll admit, also, that many firms are in actual fact employing generative AI (often tuned off the shelf things from big development firms), but they’re rarely what anyone actually needs or wants — as an alternative, they’re just attempting to get in on the brand new hot moment.

I assumed it could be useful to speak about how this all happens, nevertheless. When someone decides that their company needs “AI innovation”, whether it’s actually generative AI or not, what’s really occurring?

Let’s revisit what AI really is, before we proceed. As regular readers will know, I really hate when people throw across the term “AI” carelessly, because more often than not they don’t know in any respect what they mean by this. I prefer to be more specific, or at the very least to elucidate what I mean.

To me, AI is after we use machine learning strategies, often but not at all times deep learning, to construct models or combos of models that may complete complex tasks that normally would require human capabilities. When does a machine learning model get complex enough that we must always call it AI? That’s a really hard query, and there’s much disagreement about it. But that is my framing: machine learning is the technique we use to create AI, and machine learning is a giant umbrella including deep learning and numerous other stuff. The sector of knowledge science is kind of an excellent larger umbrella that may include some or all of machine learning but in addition includes many other things.

AI is after we use machine learning strategies, often deep learning, to construct models or combos of models that may complete complex tasks that normally would require human capabilities.

There’s one other subcategory, generative AI, and I feel when most laypeople today speak about AI, that’s actually what they mean. That’s your LLMs, image generation, and so forth (see my previous posts for more discussion of all that). If, say, a search engine is technically AI, which one could argue, it’s definitely not generative AI, and should you ask someone on the road today if an easy search engine is AI, they probably wouldn’t think so.

Let’s discuss an example, to assist perhaps make clear a bit about automations, and what makes them not necessarily AI. A matter-answering chatbot is a great example.

In a single hand, we’ve a reasonably basic automation, something we’ve had around for ages.

• Customer puts a matter or a search word right into a popup box in your website
• An application that appears at this query or set of words, and strips out the stopwords (a, and, the, and so forth — an easy search and replace type function)
• Application then puts the remaining terms in a search box, returning the outcomes of that search of your database/FAQ/wiki to the chat popup box

This can be a very rough approximation of the old way of doing these items. People don’t find it irresistible, and should you ask for the fallacious thing, you’re stuck. It’s mainly a LMGTFY*. This tool doesn’t even imitate the form of problem solving or response strategy a human might use.

However, we could have a ChatGPT equivalent now:

• Customer puts a matter or a search word right into a popup box in your website
• The LLM behind the scenes ingests the shopper’s input as a prompt, interprets this, and returns a text response based on the words, their syntactic embeddings, and the model’s training to provide exceedingly “human-like” responses.

This may mean a couple of big positives. First, the LLM knows not only the words you sent it, but in addition what OTHER words have similar meanings and associations based on its learned token embeddings, so it should give you the option to stretch past the precise words used when responding. In case you asked about “buying a house” it might probably link that to “real estate” or “mortgage” or “home prices”, roughly because texts it has been trained on showed those words in similar contexts and near one another.

As well as, the response will likely be rather more nice to read and devour for the shopper on the web site, making their experience together with your company aesthetically higher. The outcomes of this model are more nuanced and far, rather more complicated than those of your old-school chatbot.

Nevertheless, we want to do not forget that the LLM is just not aware of or concerned with accuracy or currency of knowledge. Remember my comments in previous posts about what an LLM is and the way it’s trained—it’s not learning about factual accuracy, but only about producing text that may be very very similar to what humans write and what humans wish to receive. The facts could be accurate, but there’s every likelihood they may not be. In the primary example, however, you’ve got complete control over all the things that may perhaps be returned out of your database.

For a median user of your website, this may not actually feel drastically different on the front end — the response could be more nice, and might make them feel “understood” but they don’t have any concept that the answers are at higher risk of inaccuracy within the LLM version. Technically, if we get right all the way down to it, each of those are automating the strategy of answering customers’ questions for you, but just one is a generative AI tool for doing so.

Side note: I’m not going to get into the difference between AGI (artificial general intelligence) and specialized AI straight away, aside from to say that AGI doesn’t, as of this moment, exist and anyone who tells you it does is fallacious. I’ll cover this query more in a future post.

So, let’s proceed our original conversation. What results in an organization throwing some basic automation or a wrapper for ChatGPT in a press release and calling it their latest AI product? Who’s driving this, and what are they really pondering? My theory is that there are three essential paths that lead us here.

• I Want PR: An executive sees the hype cycle happening, they usually wish to get their business some media attention, so that they get their teams to construct something that they will sell as AI. (Or, relabel something they have already got as AI.) They might or may not know or care whether the thing is definitely generative AI.
• I Want Magic: An executive hears something in news or media about AI, they usually want their business to have whatever benefits they consider their competition is getting from AI. They are available in to the office and direct their teams to construct something that may provide the advantages of AI. I‘d be surprised if this executive really knows the difference between generative AI and an easier automation.

None of this necessarily precludes the concept that a great generative AI tool could find yourself happening, of course- plenty exist already! But after we start from the presumption of “We’d like to make use of this technology for something” and never from “We’d like to resolve this real problem”, we’re approaching the event process within the entirely fallacious way.

But come on, what’s the harm in any of this? Does it really matter, or is that this just fodder for some good jokes between data scientists in regards to the latest bonkers “AI” feature some company has rolled out? I’d argue that it does matter (although additionally it is often material for good jokes).

As data scientists, I feel we must always actually be a little bit perturbed by this phenomenon, for a few reasons.

First, this culture devalues our actual contributions. Data science was once the sexiest job around, or so many magazine covers told us, but we’re fortunately settling right into a much calmer, more stable, less flashy place. Data science teams and departments provide sturdy, reliable value to their businesses by determining how the business will be run efficiently and effectively. We determine who to market to and when; we tell firms the right way to organize their supply chains and warehouses; we discover productivity opportunities through changes to systems and processes. As a substitute of “that’s how we’ve at all times done it”, we at the moment are empowered to search out the perfect technique to actually do things, using data, across our firms. Sometimes we construct whole latest products or features to make the things our firms sell higher. It’s incredibly precious work! Go have a look at my article on the Archetypes of the Data Science Role should you’re not convinced.

All these items we do is just not less precious or less vital if it’s not generative AI. We do loads of machine learning work that probably doesn’t reach the mysterious complexity boundary into AI itself, but all that stuff continues to be helping people and making an impact.

Second, we’re just feeding into this silly hype cycle by calling all the things AI. In case you call your linear regression AI, you’re also supporting a race to the underside when it comes to what the phrase means. The phrase goes to die a death of a thousand cuts if we use it to consult with absolutely all the things. Possibly that’s tremendous, I do know I’m definitely able to stop hearing in regards to the phrase “artificial intelligence” any time now. But in principle, those of us in the information science field know higher. I feel we’ve at the very least some responsibility to make use of the terms of our trade accurately and to withstand the pressure to let the meanings turn to mush.

Third, and I feel possibly most significantly, the time we spend humoring the demands of individuals outside the sphere by constructing things to provide AI PR takes time away from the true value we may very well be creating. My conviction is that data scientists should construct stuff that makes people’s lives higher and helps people do things they should do. If we’re constructing a tool, whether it uses AI or not, that no one needs and that helps nobody, that’s a waste of time. Don’t do this. Your customers almost definitely really need something (have a look at your ticket backlog!), and try to be doing that as an alternative. Don’t do projects simply because you “need something with AI”, do projects because they meet a necessity and have an actual purpose.

When executives at firms walk in within the morning and choose “we want AI features”, whether it’s for PR or for Magic, it’s not based on strategically understanding what what you are promoting actually needs, or the issues you’re actually there to resolve on your customers. Now, I do know that as data scientists we’re not at all times ready to thrust back against executive mandates, even once they’re a little bit silly, but I’d really wish to see more firms stepping back a moment and occupied with whether a generative AI tool is in actual fact the fitting solution to an actual problem of their business. If not, possibly just sit tight, and wait until that problem actually comes up. Generative AI isn’t going anywhere, it’ll be there whenever you even have an actual need for it. Within the meantime, keep using all of the tried and true real data science and machine learning techniques we have already got — but just don’t pretend they’re now “AI” to get yourself clicks or press.

See more of my work at www.stephaniekirmer.com.