Even Djokovic Loses Most Points... But Ultimately Wins [AI Wrong Answer Notebook]

Point Win Rates Hover Around 50% Even for Lower-Ranked Players
But Match Win Rates Show Overwhelming 90% Gap
A Small 2-3% Advantage Accumulates into a Major Difference
In AI, Small Successes Amid Failures Spark Revolutionary Change
Beyond the Success-Failure Dichotomy: Embracing a Probabilistic Mindset

Editor's NoteExamining failures is the fastest route to success. 'AI Wrong Answer Notebook' explores cases of failure in AI-related products, services, companies, and individuals.

The History and Reality of AI: A Graveyard of Failures

Novak Djokovic. AP Yonhap News

The history of artificial intelligence (AI) covered so far (from episode 1 to 39) has been a series of failures. The dazzling successes of AI-related products, services, and companies certainly deserve applause. However, these are only a small part of the overall history of AI. The AI innovations we enjoy today are, in fact, the result of countless trials, errors, and setbacks.

One of the ironies in AI history is that even geniuses could not predict the future of AI. John McCarthy, who coined the term 'AI,' and Marvin Minsky, were so entrenched in symbolism that they effectively stifled connectionism, one of the two main pillars of AI research at the time. This led to what is known as the 'AI winter.' Yann LeCun, who chose a different path from the mainstream research direction, had to endure decades of neglect during the 'AI winter.'

Failures have been just as frequent in the realm of AI products and services. Microsoft's chatbot 'Tay' was shut down after only 16 hours due to controversies over racist remarks. Amazon's AI recruitment system sparked debates over gender and racial discrimination. Self-driving cars caused fatal accidents, and medical AI systems made incorrect diagnoses.

Google's 'Flu Trends' project, which ambitiously set out to predict influenza outbreaks using big data, was eventually scrapped after confusing correlation with causation. In fields such as real estate, finance, and securities, there have been countless cases where expectations for an AI revolution ended in disappointment.

Looking back on the history of AI, errors and failures appear to be inevitable and unavoidable. Since technology is created by humans, it inevitably reflects human limitations and biases. Perfectly predicting or controlling the complexities of the real world is almost impossible from the outset.

Errors and Failures as the Soil of Success: The Power of Diversity

Errors and failures are daunting. However, they are absolutely necessary for success.

This is also true from a biological perspective. If no DNA replication errors ever occurred during cell division, that very perfection would actually lead to extinction.

Perfect accuracy in gene replication blocks mutations. Without mutations, genetic diversity disappears, and without diversity, adaptation to environmental changes becomes impossible. When an ice age arrives, a pandemic breaks out, or the climate changes rapidly, a population with identical genes would be wiped out in an instant. In other words, perfect replication leads to perfect extinction.

Errors create mutations, and a tiny fraction of these mutations provide advantageous traits that survive natural selection. That is why we exist today. In this sense, the very existence of life is predicated on error and failure.

The same principle can be applied to the development of AI technologies. Attempts to pursue perfection often impede innovation. Only in environments that allow for trial, error, and failure do real breakthroughs emerge. Errors are the driving force of innovation.

The Power of Chance and Social Environment

If 'effort' were the only factor determining success and failure, it would be difficult to be tolerant of mistakes and errors. Effort is important, but it is not everything.

Behind every AI success story, there are countless elements of chance and social factors. Kunihiko Fukushima proposed the core structure of the convolutional neural network (CNN) as early as 1980. CNNs are the forerunners of machine learning and deep learning.

However, at that time, there was no computing power available to implement such models. Thirty years later, scholars like Geoffrey Hinton were able to succeed with similar ideas because the timing was right. It was a period when it was possible to experiment with and implement CNN architectures. In other words, no matter how brilliant an idea is, without the support of social infrastructure, it is bound to be buried.

The process of building ImageNet by Fei-Fei Li, known as the 'godmother of AI,' is also fascinating. She hit a wall in image classification tasks but found a breakthrough through Amazon Mechanical Turk, a crowdsourcing platform. Thanks to millions of images being classified by low-wage workers around the world, the foundation for AI innovation called ImageNet was completed. It was only possible because of the social condition of low-wage labor on the other side of the globe.

The 'data' that fuels AI is no different. The basic blueprints for deep learning have existed for decades. But it was only after the spread of the internet, the proliferation of smartphones as hyper-personalized data collection devices, and the emergence of social media?when the data pipeline as social infrastructure was established?that these blueprints could shine. Only when our society as a whole began supplying AI with data as fuel did true innovation become possible.

Innovation does not arise in a vacuum. Ideas gain the power to change the world only when they meet the right social conditions. The history of AI shows that innovation is not just the achievement of brilliant individuals, but a complex product of social environment and the conditions of the times.

However, we should not try to explain everything as 'luck.' If everything depended solely on social conditions and background, then effort, perseverance, and creativity would be meaningless.

That is not true either. Innovation and success are best seen as complex products of the interaction between individual creativity and excellence, and social conditions.

What Matters Is Completion, Not Perfection

People having a meeting in an office. Photo by Getty Images Bank

The digital world is a binary world of 0s and 1s?a world of clear and simple dichotomies. However, the world of AI we have explored so far shows that this is not the case. Success and failure, perfection and imperfection, are two sides of the same coin. AI is a technology that has developed gradually amid ambiguity and uncertainty.

There are two main approaches to product development and project management. The first is the 'waterfall model,' which pursues perfection at every stage. Each step?from planning to design, development, and testing?is completed perfectly before moving on to the next.

The other is the agile model, which increases completeness through repeated revisions. Even if something is imperfect, it is built, tested, and improved continuously based on feedback.

Today, the advancement of AI follows the typical agile model. Early AI systems are inevitably simple and limited. However, leading players in the AI market do not wait for perfect systems; instead, they keep releasing imperfect versions to the market. Through trial, error, and iterative improvement, they gradually complete sophisticated AI. OpenAI's 'ChatGPT' is a prime example.

GPT-1, released in 2018, started with 117 million parameters. GPT-2 increased this to 1.5 billion, and GPT-3 to 175 billion. For GPT-4, the exact number has not been disclosed, but there are reports that it exceeds 1 trillion.

Each version was imperfect, but continued to improve through user feedback. With every surge in attention, more people flocked to the service, providing more data and better feedback. If they had tried to create a perfect AI from the start, ChatGPT might still be locked away in a laboratory.

"In mathematics, failure is a partial success. By modifying a method that almost succeeded, it often becomes a successful method."

This is a quote from Terence Tao, considered one of the greatest mathematicians of our time.

In mathematics, which forms the academic foundation of AI, even failure is a form of success. Mathematicians rarely solve problems perfectly on the first try. Instead, they form hypotheses, experiment, receive feedback, and repeat the process of revision. As Tao says, mathematical thinking is closer to a process of gradual evolution than to perfection.

The lessons from the history of AI discussed earlier shine here as well. If failure is universal and forms the basis of success, and if chance and environment determine outcomes, then there is no reason to wait for perfection.

The important thing is to complete something and put it out into the world, even if it is imperfect. The initial services launched by Google, IBM, Amazon, and others were far from perfect. But because they completed and released them, these became stepping stones for subsequent research. Perfectionism is an obstacle to progress. The history and world of AI emphasize that what matters is completion, not perfection.

Small Differences Decide Victory and Defeat: Djokovic and the Power of Probabilistic Advantage

Does 'completion' have to be something grand? No. Even a very small completion, a very small difference, is enough. It is small differences that create great achievements. The same is true in tennis.

We often expect something overwhelming from the word 'the best.' When we hear names like Novak Djokovic, Rafael Nadal, or Roger Federer, we imagine them dominating their opponents and ruling the court. We are more familiar with their triumphant roars after victory than with images of them bowing their heads in tears. But the match data tells a very different story.

In men's major singles tennis, scores are calculated as points, then games, then sets, then matches. To win a match, a player must win 3 out of 5 sets. To win a set, a player must win 6 games. And to win a game, a player must win 4 points.

What is the point win rate for top players like Djokovic, Nadal, and Federer? About 54%. Barely over half. Lower-ranked players average 51%, while mid-ranked players average 52%. The difference is not large. Even the very best players lose countless rallies and points. 'Small defeats' during matches are familiar to them as well.

However, the difference in match win rate is enormous. Lower-ranked players have a match win rate of 55%, while top players reach as high as 90%. At the point level, the difference from mid- and lower-ranked players is only 2-3%, but the final results are overwhelming.

This shows that a small probabilistic advantage, accumulated over the course of an entire match, creates a huge difference. This is explained by the 'law of large numbers.' As countless points and games are played, even a small advantage maintained consistently will ultimately converge to victory.

There is no such thing as a perfect, overwhelming victory from start to finish. What matters is making even a little progress at every moment?small wins and small probabilistic advantages.

From a Deterministic to a Probabilistic Worldview

The AI Wrong Answer Notebook written so far shows that the dichotomous thinking of failure and success, and the deterministic worldview that seeks perfect answers, have clear limitations. The success of deep learning was also the result of decades of accumulated 'failures,' and innovation came not from predictable plans, but from diversity and trial and error.

AI presents us with an entirely new way of thinking?a probabilistic mindset.

The very operating principle of AI is an embodiment of a probabilistic worldview. Even when you ask ChatGPT the same question, you get slightly different answers each time. Rather than predetermined results, it calculates probabilities in real time based on context and interaction.

'Stochastic gradient descent' is one of the core methods used in AI learning. It is like groping your way down a mountain in thick fog, unable to see even a step ahead.

Here, 'stochastic' means making decisions based on only a randomly selected subset of data each time. Instead of looking at all the data, it samples a portion and determines the direction.

This method is meaningful because it does not try to find a perfect answer. Rather than seeking the 'perfect next step,' it probabilistically searches for a 'good enough' point. This 'imperfect' approach can actually produce better results.

The systems of self-driving cars are a striking example of this probabilistic way of thinking.

When a traffic accident occurs, people usually ask the driver, "Did you see the pedestrian?" The answer is either "Yes" or "No"?1 or 0. If someone answered, "I saw them to some extent," "I barely saw them," "I roughly saw them," or "I saw them a little," it would sound strange.

But that is exactly how AI responds. AI answers with probabilities: "30%," "3.5%," "55%," "25%," and so on. Adopting AI means shifting from a deterministic to a probabilistic approach.

An image combining AI and stock market charts. Getty Images Bank

The AI revolution is a journey in search of better probabilistic possibilities, and even a 0.1%, 0.2%, or 3% increase in probability can bring about meaningful change. A success probability of 49.9% and 50.1% differs by just 0.2%, but AI encourages the 50.1% probability. A small advantage of 0.2% accumulates to create a decisive difference. That is why even small challenges, however trivial, are always meaningful. While we cannot determine the final outcome, we can change the probability of that outcome occurring.

Another important point is the role of humans in this probabilistic system. AI does not operate alone. Human input and feedback have a decisive impact on probability calculations. Numerous studies are being published on medical AI. While many results show high diagnostic accuracy for AI, we should pay attention to research showing that combining humans and AI produces better outcomes. This means that in the AI era, humans are not passive beings, but active participants who can change the direction of probability.

The world shown by AI is not one of predetermined fate, but a space of possibilities. Errors, wrong answers, and failures are not the opposite of success, but alternative paths; diversity and uncertainty are not noise to be eliminated, but sources of innovation. We live not in a deterministic world, but in a probabilistic one. That probability is 50.1%.