In 1976, Stanford University developed an AI system called MYCIN. It was a marvel of its time, designed to identify bacterial infections like meningitis and recommend life-saving antibiotic treatments. When tested against human experts, MYCIN achieved a 65% acceptability rating—comfortably outperforming infectious disease specialists who scored between 43% and 63%.
By every metric of "intelligence," the machine had won. Yet, MYCIN was never used on a single patient.
Its failure wasn't due to a lack of logic, but a mundane systemic friction: doctors had to spend nearly an hour manually typing patient data into a terminal before the system could provide an answer. It didn't save time; it created a new, more tedious form of labor. Today, we face a similar irony. AI has been "better than coders" at specific benchmarks since 2023, yet the demand for software engineers remains high. If the technology is so superior, why hasn't the replacement happened? To understand this, we must look past the code and into the structural bottlenecks of systems science.
Takeaway 1: The Theory of Constraints (Moving the Bottleneck)
In systems science, we often use the analogy of a water pipe to describe productivity. You can widen almost the entire length of a pipe to an immense diameter, but the total flow of water is still strictly limited by the narrowest section.
This is the Theory of Constraints. Jobs are not single, monolithic actions; they are a series of interconnected tasks. When we apply AI to a job, we aren't necessarily widening the whole pipe; we are usually just widening one segment, which inevitably shifts the bottleneck elsewhere.
In the case of MYCIN, the AI solved the "diagnosis" task, but the bottleneck shifted to "data entry." In modern software engineering, AI can generate code 10 to 100 times faster than a human, but because LLMs are prone to hallucinations, the human must now spend significantly more time in the "debugging" phase.
While newer AI systems are becoming more adept at debugging and high-level design, systems thinking teaches us that as soon as one bottleneck is cleared, the system will identify a new one. The "illusion of efficiency" suggests that we are moving faster, but the system's output remains throttled by the next narrowest task.
Takeaway 2: The "Last Mile" Problem and the Chaos of Reality
Even if we widen the internal sections of the pipe, the output valve—where the system meets the real world—often remains clogged. This is known as the Last Mile Problem.
In the 2010s, IBM’s Watson became a symbol of systemic opacity. Unlike modern generative AI, Watson was a "symbolic system," meaning it functioned on pure logic and did not hallucinate. It famously won Jeopardy and was quickly deployed to hospitals to recommend cancer treatments. However, the project was eventually sold for parts.
Why did a perfect logic machine fail? Because real-world data is chaotic. Doctors use fragmented jargon, records are incomplete, and patients often present with rare combinations of conditions that never appeared in the standardized training sets. Watson's failure proves that even a system without hallucinations can fail when confronted with the "chaos" of reality.
We see this same struggle with self-driving cars. They handle 99% of normal driving conditions with ease, but they fail catastrophically when they encounter "exceptions"—unpredictable human behavior or unusual roadway barriers. As researchers at Harvard Business Review have noted, "The 'last mile' problem may become one of the biggest bottlenecks slowing AI adoption."
Takeaway 3: Automation Can Blind You to Human Problems
Systems are human systems first, and computer systems second. When we prioritize automation over the human experience, we risk creating feedback loops that actually worsen the root problem.
Consider the case of an appliance manufacturer that sought to automate its overwhelmed customer support. They replaced their phone line with an automated intake form and a ticket-routing system. Internally, the metrics looked great: tickets were "moving" faster than ever. Externally, however, customer frustration spiked and reviews plummeted.
The company had fallen into the trap of using technology to cover up a systemic flaw. When a manager finally bypassed the automation and called customers directly, they discovered the truth: the product's instruction manuals were so confusing that customers were forced to call for help. The solution wasn't a faster ticketing system; it was a simple "quickstart" sheet with graphics.
By automating the support, the company had blinded itself to the human friction at the core of its product. This is why technologies like Sora ultimately failed to improve lives; while the tool could generate realistic video, it primarily served to increase distrust in digital media rather than solving a tangible human need.
Takeaway 4: The ATM Paradox (Demand Expansion)
History suggests that when technology makes a service significantly cheaper, it often leads to more jobs, not fewer. This is the ATM Paradox.
When ATMs were introduced in the 1990s, the common wisdom was that bank tellers would be obsolete. Instead, the number of tellers increased. Because ATMs lowered the cost of operating a branch, banks opened thousands of new locations to capture more customers.
This is the principle of Demand Expansion. When AI lowers the cost of specialized services—like legal advice, healthcare, or software development—those services suddenly become accessible to millions of people who were previously priced out of the market.
- Small Businesses: A company that could never afford a dev firm can now hire a junior coder using AI to build a custom inventory app.
- Consumers: Personal software becomes democratized. In the near future, having a custom calendar app built for your specific needs could be as common and affordable as a standard oil change.
Takeaway 5: Social Infrastructure is the Final Bottleneck
The final, and perhaps most significant, bottleneck is not technical—it is social.
In 1882, Thomas Edison built the first commercial power plant, but it took 50 years for electricity to fundamentally transform the economy. This was because early factory owners committed a classic systems error: they simply replaced their giant steam engines with giant electric motors while keeping the exact same factory layouts and workflows.
The real productivity gains didn't happen until factories were redesigned from the ground up. Today, AI is hitting similar infrastructure walls:
- Physical Infrastructure: There is a severe global shortage of electrical equipment. Power grids are currently unable to handle the constant load and "surge loads" required by massive AI data centers.
- Social Infrastructure: We lack the "legal layout" for an AI world. We have yet to establish clear safety codes, worker certifications, or legal frameworks to determine who is responsible for AI-caused damages or who owns AI-generated intellectual property.
Conclusion: Shaping the Transition
While headlines focus on AI-related layoffs, many of these are instances of "AI-washing"—corporate leaders using the hype of automation as a convenient cover for layoffs necessitated by previous over-hiring. These are often the moves of greedy and impatient CEOs, not the result of a structural shift in how work is done.
The structural bottlenecks of systems science remain. We are currently in a period of subsidized experimentation, and it remains an open question whether these productivity gains are commercially viable in the long run or if we are simply repeating the mistakes of the past by focusing on the tool instead of the system.
As we navigate this transition, we must remember that "we, the people, are the real infrastructure of our society." Technology can widen the pipe, but it is our values, our laws, and our human connections that determine where the water actually flows.
For all 2026 published articles list: click here
...till the next post, bye-bye & take care
No comments:
Post a Comment