Tech experts weigh in on the AI hype cycle
AI expectations couldn't be any higher. Read why leading industry experts believe the hype is deserved and what developers can do to deliver on the technology's weighty promise.
Expectations for AI are exploding as the hype around the technology continues to build. But don't expect another AI winter like we've seen in each previous AI hype cycle, say leading AI experts.
"Our assessment is that this time is different," said Marc Andreessen, founder of the investment firm Andreessen Horowitz. "The technology has started to work in a real and deep way. We think prime time has arrived."
Andreessen, who spoke at the Spark + AI Summit in San Francisco, counted at least five previous AI hype cycles. The development of AI has followed a familiar and repetitive script in which minor technical advancements give rise to overinflated expectations that the existing technology inevitably falls short of, discrediting the entire idea of intelligent machines.
No AI winter this time
Things look different this time, Andreessen said. He pointed to the 2012 ImageNet Large Scale Visual Recognition Challenge, in which an algorithm outperformed humans for the first time at an image recognition and classification task, as a turning point.
Marc AndreessenAndreessen Horowitz
Since then, tools have only gotten better, and engineers have been able to build AI systems that exceed human capability in a number of realms, proving that the technology is not mired in an AI hype cycle. The technology has arrived, and it is ready to make serious contributions in industrial settings.
Andreessen said AI is likely to represent a foundational technical change that will force all sectors of the economy to rethink the way they do things.
"Most advances in technology are incremental, but our sense is that AI has the potential to be one of these architecture changes that leads to almost complete turnover of products and companies," Andreessen said.
Moving away from a human-centric view of AI
Meanwhile, but for the promise of today's AI to pay off and move beyond the AI hype cycle, people need to start thinking more systematically about how and to what ends it's implemented, said Michael I. Jordan, the University of California, Berkeley professor who pioneered early research into neural networks.
In a keynote presentation at the conference, Jordan said that most applications of machine learning involve learning a single user's preferences and then tailoring experiences or product recommendations around those preferences.
But what happens when more services are built around this idea? Jordan said it's not a problem if Netflix delivers similar recommendations to millions of users, but if a financial services firm recommends the same stock to a large number of clients, it can artificially alter markets.
"It's not about the individualized decisions," he said. "It's about the systems of agents that we need to think about."
Jordan also took issue with some of the use cases to which AI is being applied. He said many leading developers are trying to replicate some version of human intelligence, which he described as a dead end. Instead of building algorithms that can perform a given task as well as or slightly better than a human, developers should be developing tools that do things humans can't do.
For example, Jordan said a good application of AI would be a machine learning system that analyzes traffic data and changes traffic lights to move cars through congested roadways faster. It's unlikely a human could analyze all the data necessary for this kind of operation using their own cognitive capabilities, which makes it a promising use case.
"People think this AI revolution is all about machine learning for computer vision, but that's the human-imitative perspective," Jordan said. "We're trying to build systems that humans didn't evolve to do. Human-imitative AI is a mistake if you're looking at it that way."