As generative A.I. surges in interest, with ever-expanding applications, industry discussions focus on the best ways to take advantage of the technology. While the possible applications present a tantalizing prospect for companies, adopting the technology will require companies to carefully consider existing infrastructure and data management.
OLIVIA KABELL, Associate Editor
For Jason Schern, field CTO at Cognite, a good chunk of A.I. preparation involves solving what he calls the “data problem.” Data is the backbone of any utilization plan for A.I., and without it—especially in a usable form—companies will make very little headway in incorporating A.I. When discussing his own journey into the field of A.I. and the origins of Cognite’s involvement, Schern emphasized that this data criticality has always been a concern. Pre-A.I., he noted, the company and its founders “…started realizing that—at least in the industrial space, like oil and gas—there was a really big data problem that was preventing them from being able to achieve their A.I.-related aspirations.”
Data diversity is one complicating factor, Schern noted. Since collected data can take a variety of forms—from simple to complex, structured to unstructured—data is siloed by necessity, according to Schern. Subsequently, it becomes difficult to access and utilize. Under those circumstances, “[w]e’re never really going to be that scalable, or that pervasive, because you wouldn't have the data that you need to go take advantage of those [A.I.] capabilities.” While there may be a temptation to merely move all the data to a central location and away from individual silos, Schern laid out the true counter to the problem: “finding the meaningful relationships between the data and then saving those relationships.”
For industrial data, the context and relationships for these data are not built-in—something in stark contrast to headline A.I. systems. Unlike these systems, which are trained on language and benefit from built-in context like grammar, industrial data sets must have their relationships established before they can be accessed at scale by a computer system, Schern cautioned. Even so, he pointed out that A.I. can improve efficiency in this area as well; by using the technology to speed up the process of establishing data relationships, operators can reach the point of being able to call up extensive data sets much faster, Fig. 1. “We automate the process of finding those relationships and saving those relationships as an industrial knowledge graph…to provide a frontline worker in the field doing operators rounds…[with] access to that data immediately.”
Current boundaries. Yet, there are certain boundaries to what A.I. can and cannot do at this stage, Schern was quick to emphasize. “[L]arge language models are not predictive models,” he noted, elaborating that predictive maintenance dreams remain out of reach—for the present. Rather, what A.I. systems available now excel at is locating data within massive data sets. For example, he said, finding a specific answer within a 200-page manual is a task that could be reduced from ten minutes to a few seconds, with the help of A.I. Scaled up, the overall time saved could be significant.
Schern gave another example: a Japanese client would start their shift with two to three hours of sifting through data and reports from the previous shift. With A.I., that hours-long task becomes the work of 10 to 15 minutes. This, in itself, is a major benefit, but the real advantage lies in the flexibility it offers to operators, Fig. 2. In Schern’s words, “The real benefit is that now I'm getting wrenches on task two hours sooner, which means I'm resolving the issue two hours sooner, which means I'm getting uptime two hours sooner, which means I'm getting the additional two hours.” In this scenario, the delegation of tasks to A.I. frees up much-needed expertise for other critical areas, improving overall efficiency.
Managing drawbacks. Yet, as much as A.I. excels at tasks involving data organization and retrieval, it presents serious drawbacks to the unprepared user. “Hallucination” was the word Schern used to describe the common fictional outputs that A.I. is known to produce, in lieu of an exact, data-based answer. The solution, he said, is to provide context-rich data that a given A.I. is asked to utilize to produce outputs, which would considerably lessen the instance of these “hallucinations.” The alternate scenario would be an LLM consistently trained and re-trained on datasets from the operations themselves, rather than the broad datasets used in commercially available models, but this, Schern pointed out, is too costly to pursue, at least for now.
Even so, he predicted a shift in the coming years regarding affordability—referencing similar paths for other technologies—and the need for companies now to anticipate that shift by preparing their infrastructure and data accordingly. To use a hocky metaphor, one of the trends around the technology is “organizations that are skating where the puck is going to be…[and] are doing the things now that prepare themselves to take advantage of those things [A.I., future technologies] when they come.”
Overall, Schern’s outlook on the technology’s future was positive. “[T]he capabilities today are pretty amazing and [are] allowing us to do things that you really couldn’t do before—especially [when they allow us] to have real impact on our customers.” Similar to sentiments expressed at CERAWeek earlier this year, the focus of A.I. usage now is preparation: understanding how the technology can be used now, as well as preparing for where it will be in the years to come. WO
JASON SCHERN serves as Cognite’s Field CTO. He has spent the last 25 years working with some of the largest discrete manufacturing companies to dramatically improve their Data Operations and Machine Learning Analytics capabilities. Mr. Schern’s experiences led him to Cognite, passionate about the value and impact of trusted, accessible contextualized industrial data at scale.