There’s still a lot of work to do before artificial intelligence products are truly useful.
At a company like Costco, there are concerns that hallucinations, a common problem in which AI models and tools provide incorrect information, will inadvertently reroute container ships ferrying tens of millions of dollar’s worth of the retailer’s goods. At other companies, cleaning decades of data so that it’s useful enough to dump into a large language model sometimes requires years of work and investment.
These kinds of complications involving AI, and the technology’s huge potential, were the focus of a session at Coins2Day’s Brainstorm AI conference in San Francisco this week. AI vendors may pitch their products as being magic, but those who are trying to make it work often take a more cautious view.
“I feel like a lot of it is still useless,” one attendee told the panel of speakers, which included Najeeb Uddin, CIO of non-profit AARP; Arvind Jain, CEO of workplace AI chatbot Glean; Alexandra Debow, CEO of photo sharing service swsh, and Nicholas Stevens, vice president of AI product at real estate site Zillow.
Stevens said he sees “exciting” things coming in the next year from AI as it increasingly understands unstructured data, like unlabeled images. Yet, no one argued that AI solved every problem.
“We’re still pretty early,” Debow, of swsh, said about the current wave of AI development. “We’re too early to have a white glove solution. We need a second layer to be built.”
As for where to start building that second layer, Uddin of AARP, which recently developed its own online AI chatbot for its members, said even he was “surprised” at how much of AI remains “a manual exercise” to ensure it works and is accurate for users.
“Usually, it wasn’t the LLM that was the problem, it was our data,” said Uddin, referring to making internal data ready for training large language models that underpin much of AI.
Because AI technology is still a work in progress, Jain, from Glean, urged companies considering AI tools to think of the technology “very much as an enabler” that adds 10% to 20% rather than something that can do a lot more. Responding to concerns about AI steering a container ship loaded with goods to the wrong port, Jain argued that humans should always remain in control.
“There’s no need to eliminate a human in these examples,” Jain said. “When you think what AI can do for you, it shouldn’t really be the bottom line of ‘How can I replace a human?’”