AI replacing human jobs: How does it happen?

With the rapid advancement of artificial intelligence (AI) technology, the topic of AI replacing human jobs has become a popular and controversial subject. Many people wonder which jobs are most likely to be replaced by AI, and in what order. In this article, we will explore the idea that AI’s ability to replace human jobs is not determined by human abilities, but by the industry’s fault tolerance. We will discuss how the error rate is the key factor that determines which industries will be affected by AI first and in what order.

AI Path is Not Determined by Human Abilities:

Previously, many people have attempted to predict which jobs will be replaced by AI based on human abilities such as education level, intellectual complexity, and skill/creativity. However, the AI path is not determined by human abilities, but by the AI’s application area. The determining factor for AI’s application area is very simple – fault tolerance. In summary, AI will replace human jobs starting from the industry with the highest fault tolerance and gradually move towards industries with lower fault tolerance.

GPT Cannot Fully Replace Humans:

OpenAI often uses the GPT’s ability and professional skills to directly compare and determine which jobs can be replaced by AI. However, the GPT’s ability does not necessarily mean it can perform well. For example, the reason why self-driving cars have yet to be fully developed is that their error rate is still not comparable to that of humans. The responsibility issue is also a serious problem, which means that fully autonomous L4 error rates must be reduced to a sufficient level before they can handle astronomical compensation.

Accounting is another example where the GPT’s result may be well-written, but it still requires an accountant to sign for responsibility. If the GPT’s error rate is at the 1% level, it still needs human inspection and verification. In fact, compared to current accounting tools, the GPT does not show any significant improvement, and it may even be less accurate than existing tools (existing tools can also automatically import files and have many angle checking functions). Similarly, although copilot programming has been developed, the efficiency of programming with GPT4 assistance has not significantly increased.

The Error Rate Determines the Degree of AI Disruption:

Thus, it is easy to see the timeline for AI replacing human jobs from the error rate, starting from high fault tolerance to low. As mentioned in the previous article, the GPT application scenarios are similar, and we can directly use the chart from the previous article.

E1 Level: The industry with the highest fault tolerance will be the first to be affected by GPT/generative AI (or the first to be applied), such as content generation assistance (writing/graphic design), and fortune-telling. These industries require high creativity, intelligence, or professional skills, but their fault tolerance is high because there is no standard answer. On the contrary, these industries are also the most amazing and valuable places for people to create more value (so those who can use tools “replace” those who cannot use tools).

When the GPT’s error rate continues to decrease to the E4 level, it will be the turn of the software industry to be affected. As mentioned in the previous article on the LLM’s limits from first principles, as the code becomes more complex, the accuracy probability of the generated code will decrease exponentially. Even if self-circulating debug is added, it cannot fundamentally solve this problem. After the code is complex, human debugging is still required (automatic debugging cannot converge), and even industry experts may find debugging complex systems very difficult. Therefore, this must become feasible only after the error rate is extremely low.

E6 Level: Lawyers and medical professionals are required to bear the responsibility as the future of AI and its impact on the workforce is a topic that continues to generate significant interest and concern. Many experts have tried to predict which industries and occupations will be the first to be affected by AI and to what extent. One approach to this question is to look at the capabilities of humans in different roles, such as their level of education, intellectual complexity, and skill or creativity. However, this may not be the best way to determine which jobs AI will replace. Instead, we should focus on the application areas of AI and the importance of the factor of error tolerance.

AI will “replace” jobs based on error tolerance

The driving force behind the replacement of jobs by AI is the level of error tolerance in different industries. In short, AI will begin to “replace” human jobs in industries with the highest error tolerance, and gradually move to those with lower error tolerance. This is because industries with high error tolerance are the easiest to automate and allow for the greatest amount of experimentation and learning. On the other hand, industries with low error tolerance require a high degree of precision, judgment, and experience, making them more difficult to automate.

The importance of error rate in determining job replacement

It is critical to understand that AI will only replace human workers when its error rate is low enough to meet the requirements of the job. For example, the reason why self-driving cars have not yet become a reality is because their error rate is not yet low enough. If self-driving cars were to cause a serious accident, the liability issues would be severe, which is why the error rate needs to be extremely low before self-driving cars can become mainstream. Similarly, in the field of accounting, even if GPT (Generative Pre-trained Transformer) produces accurate results, a human accountant still needs to review and sign off on the results as the responsible party. If the error rate of GPT is at the 1% level, humans will still need to check and verify the results. Therefore, the use of GPT in the accounting field is not likely to make a significant improvement over existing tools that can also import files automatically and perform various checks.

The job replacement timeline based on error rate

Based on the error rate from high to low, we can predict the timeline for AI “replacing” human jobs. The following are the different levels of error tolerance:

E1: The first industry to be impacted by GPT/Generative AI will be the content creation and assistance industries, such as writing, graphic design, and fortune-telling. These industries require a high level of creativity, intelligence, or professional skills, but have high error tolerance due to the lack of standard answers. These industries are also the most attractive areas where humans can generate more value by using tools.

E4: The software industry will be impacted only after GPT’s error rate continues to decrease. As I previously mentioned, in the article “The Limits of the LLM from First Principles,” as the complexity of code increases, the probability of accuracy decreases exponentially. Even with self-circulating debugging, the problem cannot be fundamentally solved. After code complexity increases, human debugging is still required, and it is challenging to debug complex systems even for industry veterans. Therefore, automation will become feasible only when the error rate is extremely low.

E6: Lawyers and doctors, where human beings need to serve as the responsible party, are the next industries to be impacted.

E8+: The industries with the lowest error tolerance are the hardest to automate, such as chip design and manufacturing, and rocket design and manufacturing. These industries are at the top of the chain in terms of being difficult to replace because they require a high level of experience, have high error costs, and debugging is extremely difficult.

Reducing error rate E8+ Level: The Most Difficult Jobs for AI to Replace

At the top of the hierarchy of difficult jobs for AI to replace are the professions that have the lowest tolerance for errors, such as chip and rocket design and manufacturing. These industries have high barriers to entry due to their complex and high-risk nature, and the cost of failure can be in the millions of dollars. For example, the cost of a failed launch of a rocket or manufacturing a chip with a feature size of 5nm or smaller is in the range of tens of millions of dollars. Therefore, these industries will be the last to be affected by AI.

AI Error Rate: The Key Factor in Determining the Extent of AI Disruption

The most important factor in determining the extent of AI disruption is the error rate. Recent advancements in GPT technology, such as RLHF, self-debugging, and CoT, are all aimed at improving error rates. As AI error rates decrease, the range of application areas expands, and the potential for AI to replace human jobs increases. Therefore, reducing AI error rates should be a top priority.

Bold Predictions About AI Error Rate Reductions

Given the current resource allocation in the industry, my bold prediction is that with normal computational development (five orders of magnitude in ten years), AI capabilities in perception, decision-making, and generation will improve by one order of magnitude every ten years, just like the progress made in image recognition over the past decade. As AI error rates decrease by one order of magnitude, the range of application areas will expand by one order of magnitude.

The path that AI takes to replace human jobs is not determined by human abilities but by the application areas of AI. The primary determining factor in this process is the error rate, with AI beginning to replace human jobs in industries with high tolerance for errors and working its way down to those with the lowest tolerance. The most challenging jobs for AI to replace are those with the lowest tolerance for errors and high barriers to entry, such as chip and rocket design and manufacturing. The key to determining the extent of AI disruption is reducing AI error rates, which can be achieved through advancements in technology such as RLHF, self-debugging, and CoT.