The generative AI tree stands out
You may have noticed that generative AI is dominating the business press in general, not to mention the AI literature. That’s all anyone seems to want to talk about. Unlike Yann Lecun of Meta, who said in a talk I heard recently at Northeastern University that “autoregressive LLMs [generative AI models] are doomed,” and “autoregressive LLMs suck,” I think that much of the hype is warranted. These systems are clearly quite revolutionary for individuals and organizations whose primary job is generating content. That applies, for example, to lawyers, artists, movie and TV production studios, programmers, musicians, content marketers, and writers. Those jobs alone are a pretty good chunk of the economy, and those who inhabit it should be spending a lot of time exploring generative AI tools.
However, there are plenty of businesses for which it will not be revolutionary, and they should probably be more focused on other forms of AI. In fact, they may be missing the AI forest for the generative AI tree. They get excited about ChatGPT, Bard, and PaLM, but they don’t know or care much about the forms of “traditional” (sorry, there doesn’t seem to be a better term) machine learning that have been enabling more accurate predictions for a several decades now.
My Part in Leading a Company Astray
As an example of this phenomenon, I was working with the senior management team of a telecom equipment firm this week. They asked me to come in and speak about generative AI and its potential applications in their industry. It’s good, of course, to educate business leaders on new developments in technology. But I ultimately concluded that these executives needed to know more about conventional machine learning before getting a heavy dose of generative AI instruction.
I do fault myself with this audience for not beginning with a discussion of how generative AI differs from other forms of deep and machine learning. I launched right into the promise and peril of generative AI, which I had been asked to do. It became clear that at least several members of the audience didn’t understand the differences between generative machine learning tools and those that simply predict future quantitative numeric values based on past ones. The focus of discussion, however, was on generating text: legal contracts, product descriptions, and translations of new product announcements.
That would be all fine—generative AI can certainly generate text—except for the fact that most of this company’s important data is quantitative, locational, and sensor-based. It could create plenty of AI use cases involving that sort of data, including predictive maintenance, location optimization, pricing, power management, and the like. The company already has some of those applications underway, but it could use some acceleration in its progress. And of course, this type of data is not generally useful in generative AI use cases.
It’s mostly a good thing that generative AI has grabbed the attention of just about everyone. But it could result in some companies taking their eye off the more important ball. If a company wants an AI use case portfolio in which the right tools are used for the right applications, its leaders need to understand the different types of AI and how they are best employed.
Remember Watson?
This all reminds me of the period in the early 2010s after IBM’s Watson soundly defeated its human opponents on the Jeopardy! game show. Company executives and board members began to ask, “What are our plans for using Watson in our own business?” It turned out that Watson—at least in its original incarnation—was best suited to answering questions involving general knowledge. Perhaps egged on by somewhat misleading marketing messages from IBM, companies and organizations tried to use it for a variety of different applications—from curing cancer to developing new drugs to providing investment advice. None of these worked out very well.
I don’t believe that any of the major vendors with generative AI models are making excessive claims about what the technology can do. But there are many startups employing their models for specific purposes that are bordering on marketing excess. In addition, there is a massive amount of media hype about generative AI’s capabilities. So organizations seeking to solve their problems with generative AI need to be particularly careful about how and where to apply it.
In many cases this situation will mean a need for education of senior executives. I promise that the next time a company asks me to talk with them about generative AI, I will plead with them to let me educate their audience on how generative AI compares to other AI approaches, and the types of problems for which generative AI is and is not suited.
