The Overestimation of AI Technology
Overhyped in the Short Run, Underestimated in the Long Run: The AI Paradox Explained Through Amara’s Law
Recent buzz around artificial intelligence (AI) has led to significantly overestimating the technology’s current capabilities. Companies like Nvidia have seen their stock prices skyrocket due to investor excitement about AI’s potential. The company’s stock has risen approximately 120% this year and over 600% since the beginning of last year. However, hear my warning that many of the supposed applications are not ready for prime time and may never deliver the expected value.
The Reality Check on AI Readiness
While AI shows promise, its readiness for widespread practical applications is still questionable. Many AI solutions touted today face significant hurdles:
- Cost-efficiency: Implementing AI systems can be prohibitively expensive. For instance, the development and maintenance of AI models require substantial financial investment in hardware, software, and expertise. A recent report highlighted that the cost of training a state-of-the-art AI model can exceed millions of dollars.
- Operational Reliability: Some applications may not work as expected, leading to unreliable outcomes. This is particularly true for complex systems that need to operate in dynamic and unpredictable environments.
- Energy Consumption: The energy demands of AI systems are substantial and often unsustainable. Studies have shown that training large AI models can have a significant carbon footprint, comparable to the lifetime emissions of multiple cars.
- Trustworthiness: Ensuring AI systems produce trustworthy results remains a significant challenge. Issues such as bias in training data and lack of transparency in decision-making processes can lead to untrustworthy and potentially harmful outcomes.
For example, the high costs associated with AI development and deployment can be seen in the substantial investments companies like Microsoft and Amazon have made in AI infrastructure, primarily purchasing expensive GPUs from Nvidia. Despite these investments, the practical benefits remain limited, with many applications failing to deliver the expected returns on investment.
The Unknown Future of AI
AI’s future remains an enigma, much like past technological revolutions. Consider the iPhone and the Internet — both had uncertain beginnings but eventually transformed the world in unimaginable ways:
- The iPhone: Initially seen as a niche product, the iPhone redefined mobile communication and computing. When it was first introduced in 2007, few could have predicted the massive impact it would have on various industries, from telecommunications to software development.
- The Internet: Starting as a project for academic and military use, the Internet eventually became integral to everyday life. Its early days were marked by scepticism and uncertainty about its potential, but it has since revolutionized communication, commerce, and information sharing globally.
AI could follow a similar trajectory, but predicting its exact path is challenging. This unpredictability means that while current AI applications may seem overhyped, future developments could still surprise us. Historical precedents suggest that transformative technologies often face initial scepticism and underestimation, only to exceed expectations once they mature and find their niche applications.
“We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run”. Roy Amara
The Confusion Among AI Terms
One major issue is the confusion between different AI-related concepts.
- Probability & Statistics: These are mathematical foundations that underpin data analysis and are essential for making predictions based on data. Probability and statistics are used in various fields to model uncertainty and draw inferences from data.
- Machine Learning (ML): A subset of AI, ML involves algorithms that learn from data to make predictions or decisions without being explicitly programmed for each task. For example, recommendation systems on platforms like Netflix and Amazon use ML to suggest content based on user preferences.
- Artificial Intelligence (AI): A broader field encompassing ML, AI aims to create systems that can perform tasks typically requiring human intelligence. This includes natural language processing, image recognition, and autonomous vehicles.
- Generative AI: This specific type of AI generates new content, such as text, images, or music, based on learning patterns from existing data. Examples include OpenAI’s GPT models, which can produce human-like text based on a given prompt.
- Artificial General Intelligence (AGI): A theoretical form of AI that possesses general cognitive abilities akin to human intelligence, capable of understanding and learning any intellectual task that a human can. AGI remains a distant goal, with current AI systems being far from achieving such capabilities.
Most people, including some experts, often conflate these terms, leading to misunderstandings about AI’s capabilities and limitations. For instance, many believe that current AI systems, which are predominantly ML-based, possess AGI-like capabilities, which is far from reality. Surveys have shown that even professionals in the tech industry can struggle to distinguish between these different concepts, contributing to the overhyped perception of AI.
The Non-Deterministic Nature of AI
AI’s non-deterministic nature — where outcomes are not always predictable — poses significant challenges for many industries.
- Healthcare: AI in medical diagnostics can lead to unpredictable results, making it difficult for healthcare providers to trust AI recommendations without human oversight. Cases of AI misdiagnosing conditions or failing to recognize critical symptoms have raised concerns about its reliability in clinical settings.
- Finance: In trading and risk management, AI’s unpredictability can result in significant financial losses if not carefully managed. Algorithms used in high-frequency trading can behave unpredictably under certain market conditions, leading to unforeseen consequences.
- Autonomous Vehicles: The non-deterministic behaviour of AI in self-driving cars can lead to safety issues. Instances where autonomous vehicles have failed to correctly interpret road signs or react appropriately to unexpected obstacles highlight the challenges of deploying AI in critical real-world applications.
Despite these challenges, non-determinism is more acceptable in certain contexts. For instance, as a personal assistant, AI can afford to be less predictable. This flexibility allows for the development of an Artificial Intelligence Operating System (AI OS), where the AI aids in everyday tasks, learning and adapting to user preferences over time without the stringent requirements of deterministic accuracy.
Blockchain Hype: A Parallel Perspective
The current AI hype bears striking similarities to the earlier hype around blockchain technology. In my article Enterprise Blockchain: A Promising Technology Stalled by Steep Challenges, I discussed how the potential of blockchain technology has been frequently overestimated while its practical challenges are often underestimated. Just like AI, blockchain promised revolutionary changes but faced significant hurdles in cost-efficiency, scalability, and regulatory compliance. Similarly, in Is Enterprise Blockchain Dead? I argued that while blockchain remains a transformative technology, its widespread adoption is hampered by high costs, integration complexities, and the absence of standardized solutions.
Both AI and blockchain share a history of initial overhype followed by a gradual realization of their practical limitations. We need to set realistic expectations and invest in a deeper understanding of these technologies. This will eventually become critical for their development and adoption.
References
- Cost of Training AI Models: AI training costs are growing exponentially — IBM says quantum computing could be a solution
- Operational Reliability in AI Systems: ‘Trustworthy AI’ is a framework to help manage unique risk
- Energy Consumption of AI Models: Sustainable electrification in the era of AI
- Trustworthiness and Bias in AI: AI’s Trust Problem
- Historical Impact of the iPhone: iPhone at 10: how it changed everything
- Evolution of the Internet: The Invention of the Internet
- Differences Between AI Concepts: Artificial Intelligence vs Machine Learning vs Data Science
- Non-Deterministic Nature of AI: AI Prompt Engineering Is Dead
