The technology can be used to automate mundane tasks, but users should beware of its security lapses and capacity for error.
ChatGPT is grabbing headlines in recent days. The large language model (LLM), which launched in November, is already being discussed as a technology that could solve nearly all business woes.
On the more problematic side, ChatGPT can also be used as a portal through which employees might accidentally share trade secrets with the world, as Samsung recently did.
Whatever roles LLMs will play in the manufacturing world, it’s too late to talk about “when” it will happen because it’s already here.
“We’re past that horizon,” said John Lanza, partner and co-chair of the manufacturing team at Foley & Lardner. While Lanza added there’s excitement about the technology, his clients are struggling with “how do we make sure that our employees don’t use it in ways we don’t want them to use it.”
As ChatGPT progresses further into the manufacturing industry, here’s what companies need to know.
LLMs automate repetitive tasks
Right now, LLMs are taking the monotony out of some workloads. In manufacturing, it’s being used for “automating tasks that people in any job may find tedious or boring or repetitive,” said Lorenzo Veronesi, associate research director at IDC Manufacturing Insights EMEA.
That includes tasks like writing emails, processing data, and also suggesting prompts for things like strategy sessions and quality assurance reports. Used in these ways, LLMs become a “helper to humans” and “can free up your hands and your brain for doing something more relevant,” Veronesi said.
LLMs can be used to create summaries of materials, items or processes at a manufacturing company.
“If we have a lot of data integration between machines and between control systems and networks, it’s a useful tool to capture the data and move the data across the factory,” Veronesi noted. In that way, LLMs can simplify integration to “make it a little bit quicker for people who have to deal with this coding and programming.”
It can also assist with customer interactions and customer service, said Matt Tippets, senior vice president of product at Drift, where an LLM is trained on a manufacturer’s own written material and processes. “A representative can choose to edit or send that respond [directly].”
Keeping the human in the LLM process
The key for manufacturers, Tippets said, is that the LLMs should be drawing from a company’s written materials only, such as internal websites, corporate documents and safety manuals, not the entire internet. Otherwise, the LLM can pull together the wrong answer, but with a technical confidence that would never suggest it’s wrong, jeopardizing the company’s relationship with its customers.
“ChatGPT doesn’t really care about any business. It cares about scraping the internet for the response it thinks is best,” Tippets said. “It’s not going to be aware of a manufacturing process or relationship. It’s not going to be aware of the importance of brand.”
People tend to talk about LLMs in human terms, but it’s important to remember it’s not human, which is why it needs to be “wrapped in the context of the business,” he added.
That’s also why it’s key to train the tech on the company’s specific information, and that it’s used to supplement, not replace, a person’s work. This will ensure that it doesn’t go haywire or alienate customers, who can often tell they’re talking to an LLM and not a person, Tippets said.
LLMs can also “hallucinate,” Veronesi said, a term used when the tech completely makes up information and acts like it hasn’t, such as concocting a fake Washington Post article.
“It’s really tricky to rely on this tool to get an answer on something you don’t know anything about,” he said. “If you already know the answer, you can get it as a tool to do something for you, then review what it’s done and validate it and bring it forward.”
The current limitations of LLMs in manufacturing
There’s another reason to limit what materials a manufacturer’s LLM is trained on in a closed environment, or even whether it should be used: possible exposure of trade secrets.
That’s especially true with ChatGPT, which is an open source application. While the tech does work for its users, they must provide their own information in return.
“Anything you say to it can and will be used against you,” Lanza said. “If you’re using a public model, you have to treat it like any other public person or resource.”
This is why Lanza noted many of his clients who have embraced LLMs remain concerned about how they can ensure employees don’t misuse the technology.
Earlier this year, Samsung employees accidentally shared confidential information when its semiconductor division used ChatGPT to check source code. The company has since limited ChatGPT upload capability to 1024 bytes per person.
LLMs also have their limits in terms of creative capacity. In a February blog post, Veronesi and Mark Casidsid, a senior research analyst at IDC Manufacturing Insights, asked ChatGPT how it could be useful in manufacturing. The results were mixed.
“It came out with some interesting suggestions, but some of them were a little generic,” he said. ChatGPT offered six suggestions for how to use it, including quality control, predictive maintenance and document summarization.
However, when Veronesi and Casidsid tried to get ChatGPT to be more specific, it “goes around in a circle and tells you the same thing it told you before with different wording,” Veronesi added.
While he noted the technology will improve, it’s impossible to know what it will do next and what kinds of advances, particularly in manufacturing, it will enable.
“If you asked people in the early 1900s what a washing machine would look like, they would probably see a robot that takes the wash to the pond and washes it and scrubs it with stones and soap,” Veronesi said. “It’s very difficult to see where the technology is going and how it’s going to work out.”
Written by: Jen A. Miller, for Manufacturing Dive.