Some years ago, when this industry was first figuring out how to run voice (and video) over IP networks, the quality left a lot to be desired. It took extensive engineering to build networks and real-time applications that, together, could deliver a simple telephone call with reasonable voice quality. Eventually, most of the technology problems, like jitter, were effectively addressed, though not eliminated. Only one basic problem remained—what the engineers wryly called the “speed of light problem.” Basically, when moving voice or video traffic from one part of the world to another over an optical network, you would never be able to move it faster than the speed of light. Overcome that minor limitation, and voice quality theoretically could be perfect some day.
It demonstrated the confidence that marked the Internet era: We’ve solved every other problem, maybe we’ll find a way to work out that speed-of-light thing. It was a joke, of course. But in general, there was confidence that any technology problem would eventually meet with a technology solution.
Fast forward—very fast—to the next era, the AI Era that we now inhabit. Will the principle continue to hold true, that technology problems will find their technology solutions?
One reason to be skeptical is that AI is qualitatively different from something like IP networking. As attorney Martha Buyer points out on No Jitter this week, we don’t necessarily know how generative AI systems come up with the answers the give us, which makes troubleshooting not just difficult but fraught with potential risk.
Still, those on the cutting edge of generative AI are attempting to use the technology itself to mitigate, if not overcome, some of the more difficult risks. This MIT article describes one such effort, centered on one of the biggest obstacles to gen AI’s success as a customer-facing self-service technology: The potential for gen AI to go rogue and spit out toxic responses to user queries.
The process of training an AI bot to avoid giving offensive responses has been laborious and human-driven, but the MIT researchers are attempting to use gen AI itself to automate the training process. The hope is that gen AI-based “red teaming” can come up with exponentially more variations in much less time. According to the MIT article, “The technique outperformed human testers and other machine-learning approaches by generating more distinct prompts that elicited increasingly toxic responses,” which the system then could train the model to avoid.
Experts are also expressing hope that gen AI can be used to mitigate another of its current risks: Technical debt. One reason enterprises are moving cautiously on gen AI is the fear of being stuck with the wrong investment, but according to CIO Dive, improvements to software development—driven by gen AI—could reduce this risk: “While early movers are at risk of technical debt, generative AI efforts may help in that area, too. Gartner predicts that by 2027, enterprises will use generative AI tools to create appropriate replacements for legacy apps, shrinking modernization costs by 70%.”
These mitigation efforts certainly won’t be perfect, and we’re just at the beginning of gen AI adoption, let alone an era when gen AI solves all the problems created by gen AI. Caution absolutely is warranted when testing and deploying gen AI-based solutions, but as multiple speakers said at last month’s Enterprise Connect, you’ve still got to be exploring the technology: You can’t do nothing.
