- AI is much, much harder than marketers would have you believe
- RAG, often touted as a solution to AI hallucinations, isn't nearly as straightforward as it seems
- The best approach is to understand AI implementation is a process and choose knowledgeable partners
No, the AI bubble hasn’t burst. Despite much hand wringing over a stock swing earlier this week, HPE’s VP and GM of AI Solutions Joey Zwicker and AvidThink founder Roy Chua both told us the market correction was more of a deflation than an outright crash. But, they added, folks are starting to wake up to one key reality: AI is hard.
It’s true, some use cases are easier than others. Turning on Microsoft Copilot or using Google Assistant, for example, are on the easy side. But Chua said some of bigger projects enterprises are reaching for – the ones that involve creating custom workflows or integrating data across silos to implement retrieval augmented generation (RAG) – well, those are a different story.
“That’s not as easy as you think it is and RAG is not as straightforward as you think it is,” Chua said.
Chua said he’s encountered this truth in his own business. AvidThink, he said, fed 10 years’ worth of its reports into ChatGPT to run RAG. But what he found was that answers delivered by the model lacked critical context, making them unsuitable responses to the questions asked. And because there was no metadata included with the reports that would have signified a date, the model would return information that had become outdated.
“So, that’s what I discovered,” Chua said. “It’s not as easy as dumping all the data in there.”
Womp, womp.
Trough of…promise?
The recognition that models aren’t magic and there is no real ‘easy’ button for AI could actually end up being a good thing for the market.
Why? Because knowing that upfront means enterprises will be less likely to give up when they encounter hurdles in the deployment process.
“The way people will get disillusioned is when they all continually bite off more than they can chew,” Zwicker said. “That’s actually the way we as an industry could shoot ourselves in the foot, is by making it seem easier than it is.”
It’s all about finding the balance between the time and money you’re willing to invest and getting the result you want, Zwicker said. For instance, not every use case requires RAG, but it can be worth the headache for some.
“Many customer successes include RAG,” he added. “It’s hard, but the results you get are meaningfully better.”
According to Zwicker and Chua, the best way for enterprises to chart a smooth course through AI’s choppy waters is to choose your use cases and business objectives carefully and partner with someone who knows what they’re doing.
But remember, it won’t just be a case of ‘here’s my money, and problem solved.’
“It’s work,” Zwicker concluded.