In recent years, the generative artificial intelligence (Gen AI) was the target of exaggerated expectations and promises of a world-changing industrial revolution. Since the launch of ChatGPT, the idea that AI would generate massive changes, including the loss of 300 million jobs, has been widely publicized. However, 18 months after the peak of the hype, the reality has proven different.
The AI Hype Cycle
Like many new technologies, generative AI has followed a well-known path called Gartner's hype cycle. This model describes a recurring process in which a technology's initial success leads to inflated public expectations, which are eventually unfulfilled. After the "peak of inflated expectations" comes the "trough of disillusionment," followed by a "slope of enlightenment" and, finally, a "plateau of productivity."
A Gartner report published in June 2024 indicated that most generative AI technologies are still at or near the peak of inflated expectations. Practical adoption of these technologies has been less successful, with 80% of AI projects failing, according to a RAND study, a rate more than double that of non-AI projects.
The Current Limitations of Generative AI
The challenges facing generative AI are numerous, from the high investment required in data and AI infrastructure to the scarcity of qualified human talent. However, the unusual nature of Gen AI's limitations poses a critical challenge.
For example, generative AI systems are capable of solving complex university tests, but fail on simple tasks, as demonstrated by McDonald's failed attempt to automate drive-thru orders. This discrepancy creates false confidence in users, who end up using the models in inappropriate situations.
Experience from successful projects shows that it's difficult to get a generative model to follow instructions accurately. Khanmigo, Khan Academy's tutoring system, is an example of this, revealing correct answers even when instructed not to.
Why Isn't the Hype Over Yet?
Despite the challenges, generative AI technology is improving rapidly, driven primarily by the increasing scale and size of models. Research shows that the number of parameters, the amount of data, and the computing power used in training all contribute to model performance, while the neural network architecture has minimal impact.
Large language models also exhibit unexpected emergent abilities, such as reasoning by analogy and reproducing optical illusions, which emerge when the models reach a critical size. The causes of these advances are disputed, but there is a consensus that the models are becoming more sophisticated.
AI companies continue to work on larger, more expensive models, while companies like Microsoft and Apple bank on the returns on their existing investments. It's estimated that generative AI will need to generate $600 billion in annual revenue to justify current investments, potentially reaching $1 trillion in the coming years.
What's Next?
As the AI hype begins to fade and we enter the period of disillusionment, we're seeing more realistic adoption strategies. Companies are using AI to support humans, rather than replace them. A recent survey showed that US companies are using AI primarily to improve efficiency (49%), reduce labor costs (47%), and increase product quality (58%).
We've also seen a rise in smaller, cheaper generative AI models trained on specific data and deployed locally to reduce costs and optimize efficiency. OpenAI, for example, released the GPT-4o Mini model to reduce costs and improve performance.
Furthermore, there is a growing focus on AI literacy and workforce education about how AI works, its capabilities and limitations, and best practices for ethical use of AI. We will have to learn and relearn how to use different AI technologies in the coming years.
In the end, the AI revolution will likely be more of an evolution, growing gradually over time and gradually altering and transforming human activities. Which, arguably, is far better than replacing them.
