The GenAI Disillusion

Have you ever had a dream that you were so sure was real? How would you know the difference between the dream world and the real world?

— Morpheus (The Matrix, 1999)

"The room was large, dark, and cold, with smooth, hard walls. The only light, in the opposite corner to the window, featured the teenager's fixed gaze on the screen. Unaware of the food scraps on the table, all he saw was a ceaseless cascade of glowing letters and numbers. A fluid, frantic, and poorly rehearsed dance.

The mission was humanly impossible but perfect for the AI: learn to recognize thousands of people by reducing millions of pixels to the simplest possible alphanumeric representation. Thus, when seeing one of these faces in the future, the AI would be capable of recognizing it in real time.

The screen represented different memory states created by the AI, all supposed to become more efficient over time. However, as the young programmer's gaze swiftly alternated over different points on the monitor, he noticed less order and more random digits. Agile and tireless, yet still confused, the AI connected previously isolated parts and sought new pathways to solve the problem it had not yet fully understood.

Hours later, as if awakening from a trance, the programmer felt the first rays of the sun touching his face. Rubbing his eyes vigorously, he was sure of one single truth: he would remain trapped in that silent film of quick, disconnected flashes—a thriller with unpredictable ending, projected, watched, and critiqued simultaneously as it was scripted and performed... By a machine."

The Desert of the Real

Do you think the above text was written by a human, or was it copied and pasted directly from Chat GPT? Take all the time you need.

If you paused for a few seconds, needed to review the text, or checked an online AI detector, don't worry: all three reactions are completely acceptable. If you didn't resort to these devices, a quick and confident answer shows that you trust your ability. And if you guessed the proper answer, the question remains: for how long will you still be able to get it right?

I was the one who just wrote the programmer micro tale. The scene reflects a true story from my own sleepless college nights, but it is easily mistaken by many as a typical response from conversational AI.

Thirty years ago, when researching machine learning for face recognition at UFMG's Parallel Computing Laboratory, artificial neural networks took hours or even days to learn small data sets. Today, FaceID is so common that it goes unnoticed. No one needs to stay up nights training their smartphones, laptops, or their doorbell cameras to be promptly recognized. This trend is irreversible, as more and more autonomous intelligent agents will perform specific tasks in our routine at home and work. How many and which ones are you already using?

At this point, it is relevant to understand the difference between Generative AI (or GenAI) and other types of AI. Instead of merely repeating specific behaviors it was trained for, GenAI can generate new content and ideas. From a vast base of examples and billions of parameters, it makes inferences and creates internal representations of content that was previously seen. It draws inspiration from mathematics and statistics to generate texts, images, videos, and music as if by magic, with the apparent quality of an expert.

In case you haven't noticed yet, Generative AI is one of the pillars for the new gold rush in technology: AGI (Artificial General Intelligence), which would exhibit human-like reasoning and have the capability to teach itself. But with everything moving so fast, so many information sources, and so little time available, how can we comprehend these technologies, apply them to our advantage, and ensure our place under the sun in the future? How much can we trust Generative AI, and how will it evolve in the coming years?

Cycles and Expectations

I believe it is very important to train our natural neural networks to detect and analyze short and long-term technological cycles. A useful tool for understanding innovations and predicting their impact on the technology sector is the Hype Cycle.

Created by Gartner Group, a renowned strategic technology consulting firm, the Hype Cycle represents the market's expectations regarding different innovations over time. The cycle below is quite illustrative for tracking the acceptance and evolution of artificial intelligence technologies. On the X-axis, time runs from left to right. On the Y-axis, expectations about AI rise from bottom to top. The point that locates each technology is marked by an icon that indicates the time in years until it becomes productive and widely adopted.

Observing each stage in more detail:

1. The first stage is the Innovation Trigger, when a new technology is introduced without commercially viable products or proven use cases. The estimate is that it will captivate the few innovators interested in adopting it at such an early stage.
2. Over time (usually, a few years), the innovation moves toward the Peak of Inflated Expectations. At this high point, enthusiasm about that innovation reaches its peak with the proliferation of successful news and strong storytelling... but many failures. At this moment, it is common to overestimate the technology's maturity and create unrealistic expectations for strongly positive results.
3. As time passes, the new technology declines into the Trough of Disillusionment. Reality sets in, interest wanes, and many abandon its adoption. Very few companies defy the odds, learning from their mistakes, refining, and improving the technology. This stage tends to last many years, proportional to the innovation's complexity.
4. Competitive businesses with strong differentiators then begin to climb the Slope of Enlightenment, better understanding how the technology can be applied with real benefits in the day-to-day operations of companies and the routine of society. Use cases are more clearly defined and begin to demonstrate the practical value of the technology, leading to a more realistic and structured understanding of its capabilities and limitations.
5. The expectation about the innovation stabilizes in the final stretch of the cycle, called the Plateau of Productivity. It is at the start of this stage that the hype ends and the innovation reaches maturity, being widely adopted, accepted, and consistently adding value.

At this point in the cycle, everyone understands the capabilities of the technology, leading to even broader adoption and the stabilization of the innovation. An entire ecosystem of solutions proliferates around the players in this stretch.

Concluding...

This is a great article to learn more about the Hype Cycle and study about your specific niche or sector. But what does all this say about GenAI?

The video in this post shows how Generative Artificial Intelligence behaved on the AI hype cycle between 2021 and 2023. It was one of the fastest innovations in history to approach the peak of inflated expectations with such a broad market share, and soon after being mapped by Gartner. This is a very interesting metric: more than 200 million current Chat GPT users form the innovators expected during the first cycle of expectations. Notice in the same video how General AI (GAI) regressed in the time prediction to reach its maturity.

If the journey through the peak and valley awakens the first early adopters, there is a very large user base yet to be conquered as GenAI users until it catches up to the Slope of Enlightenment. The innovation would then be characterized as a potential technological commodity, whether in a market concentrated with a few players or fragmented among hundreds of them. As of today, thousands of startups are seeking their spot at the peak of Generative AI, but they focus on very specific small tasks like face swap, video generation, automatic subtitles, and hundreds of niche applications.

Few startups like these will be acquired by larger competitors, very few will reach scale, and the vast majority will die (or walk sideways until death) in the next 12 months. I believe that by 2026, GenAI will have two main business models: aggregators (gathering various specialized GenAI apps paid by the amount of use, like a B2C or B2B Swiss Army knife) and huge platforms, which provide services in hardware and services with AI similar to how AWS works for cloud computing.

It is in this second category that Sam Altman hopes to stake territory by trying to capture trillions of dollars in investment and maintaining the current lead of Open AI. Supported directly by Microsoft, Sam wants to create proprietary AI hardware to compete with Nvidia. Nvidia, in turn, is a market leader and has already begun to offer conversational features to consumers of its RTX cards. It is a war for potential millions of paying users with monstrous revenues... And by the looks of it, the group of trillion-dollar big techs — Google, Apple, Meta, Microsoft, and Amazon, collectively known as GAMMA — will also need to invest hundreds of billions in new fronts or alliances that allow their future perpetuity on the productivity plateau of Generative AI. Because let's face it, Google, Apple, Meta, and Amazon have not shown a recent history of strong success in this front.

Who would have thought, and it's hard to believe: all this hype about the wonders of GenAI is just beginning to slip into the Trough of Disillusionment in 2024. It will likely disappoint many heavy users, who will demand more than the technology delivers today, and the corporations themselves, which lack innovation policies and structured data to take advantage of the wave of efficiency brought by Generative AI technologies.