Less than two years ago, the release of ChatGPT kicked off the generative AI boom, a technology that some said would ignite the Fourth Industrial Revolution and completely reshape the world as we know it.
In March 2023, Goldman Sachs predicted that 300 million jobs could be lost or reduced due to AI. Big changes seem to be on the way.
Eighteen months on, generative AI has yet to transform business. McDonald's attempt to automate drive-thru ordering was a hilarious fail that ended up on TikTok, but many projects using the technology have been scrapped. Government efforts to create a system to collect public opinion and calculate welfare payments have met the same fate.
So what happened?
The AI hype cycle
Like many new technologies, generative AI follows a path called the Gartner Hype Cycle, first described by American technology research firm Gartner.
This widely used model describes a recursive process in which the initial success of a technology leads to public inflated expectations that are ultimately unrealized. An initial “peak of inflated expectations” is followed by a “trough of disillusionment,” then a “slope of enlightenment,” and finally a “plateau of productivity.”
The Conversation, CC BY
A June report from Gartner listed most generative AI technologies as either at the peak of hype or still rising, arguing that most of these technologies are two to five years away from being fully operationalized.
Many exciting prototypes of generative AI products have been developed, but their real-world deployments have not met with much success: A study published last week by the US think tank RAND found that 80% of AI projects fail, more than double the rate of non-AI projects.
The shortcomings of current generative AI techniques
The RAND report lists many difficulties with generative AI, from the requirement for significant investment in data and AI infrastructure to a lack of necessary talent. But the unusual nature of GenAI’s limitations represents a significant challenge.
For example, generative AI systems can solve highly complex college entrance exams but fail very simple tasks, making it very difficult to judge the potential of these technologies and leading to a false sense of confidence.
After all, if you can solve complex differential equations or write an essay, you should be able to take on a simple drive-thru order, right?
Recent studies have shown that the capabilities of large-scale language models such as GPT-4 have not always lived up to people's expectations: in particular, more capable models performed significantly worse in critical cases where incorrect responses could have devastating consequences.
These results suggest that these models can give users a false sense of confidence: our fluency in answering questions can lead humans to reach overly optimistic conclusions about the model's capabilities and to deploy the model in situations where it is not well suited.
Experience from successful projects has shown that getting a generative model to follow instructions can be difficult: Khan Academy's Khanmigo tutoring system, for example, would often give the correct answer to a question even when it wasn't prompted to do so.
So why is the generative AI hype not over yet?
There are several reasons for this:
First, generative AI technologies are advancing rapidly, despite challenges, driven primarily by scale and size.
Research has shown that the size of the language model (the number of parameters), as well as the amount of data and computational power used to train it, all contribute to improving the model's performance. In contrast, the architecture of the neural network that powers the model seems to have little impact.
Large-scale language models also exhibit so-called “emergent capabilities” — unexpected abilities in tasks for which they have not been trained. Researchers report that new capabilities “emerge” once the models reach a certain critical “breakthrough” size.
Research has shown that sufficiently complex large-scale language models can develop analogical reasoning capabilities and even replicate illusions experienced by humans. While the exact causes of these observations are disputed, there is no doubt that large-scale language models are becoming increasingly sophisticated.
So AI companies are working to develop bigger, more expensive models, and tech companies like Microsoft and Apple are betting on profits from their existing investments in generative AI. According to one recent estimate, generative AI would need to generate $600 billion in revenue per year to justify current investments, and that figure could grow to $1 trillion over the next few years.
For now, the biggest winner from the generative AI boom is Nvidia, the largest manufacturer of the chips that are fueling the generative AI arms race. As the proverbial gold rush goes, Nvidia recently became the most valuable public company in history, with its stock price tripling in a year and its market capitalization hitting $3 trillion in June.
What happens next?
As the AI hype begins to die down and we move through a period of disillusionment, we are beginning to see more realistic AI adoption strategies.
First, AI is being used to support humans, not replace them: A recent survey of US companies found that they are primarily using AI to improve efficiency (49%), reduce labor costs (47%), and improve product quality (58%).
Second, we are also seeing a rise in smaller, cheaper generative AI models that are trained on specific data and deployed locally to reduce costs and optimize efficiency. Even OpenAI, which has been leading the race to develop larger models, has released its GPT-4o Mini model to reduce costs and improve performance.
Third, there is a strong focus on providing AI literacy training and educating employees on how AI works, its potential and limitations, and best practices for ethical AI use. Over the next few years, we will likely need to learn (and relearn) how to use a variety of AI technologies.
Ultimately, the AI revolution will be like an evolution: the use of AI will gradually increase over time, bit by bit changing and transforming human activities, which will be much better than replacing them.
