The Tragedy of the Defense Industry and Software: Lost Knowledge—But Is There Hope for the Future?

Ladies and gentlemen, I’ve heard that Western manufacturing is gradually declining, but the defense and computer industries are still holding their own. Yet, unexpectedly, even these two “fig leaves” are no longer as glorious as they once were. The article states that the core problem lies in the lack of succession for technical know-how—people no longer know how to make things, nor do they know how to pass on that knowledge. The dire consequences of over-optimization have turned the U.S. defense industry into a laughingstock. After the Cold War, the U.S. enjoyed an unprecedented peace dividend, prompting defense giants to downsize and merge into five major conglomerates. The workforce plummeted from 3.2 million to 1.1 million as companies streamlined operations to minimize costs. But this process completely destroyed traditional craftsmanship. Should war break out, defense contractors would be forced to scramble to reorganize—even the nose cones of missiles would have to be hand-twisted by master craftsmen in their seventies. The most striking example is Raytheon. Orders placed four years ago won’t be delivered until June 30 because production had been halted for nearly 20 years—the factories were gone, the blueprints were lost—so they had to bring back veteran craftsmen to teach young workers step by step using 40-year-old blueprints. And that’s not even the worst of it—Javelin anti-tank missiles, 155-millimeter artillery shells, and TNT explosives are all unable to withstand the demands of combat. The situation is the same in Europe: a pledge to supply Ukraine with 1 million artillery rounds in one year resulted in only half being delivered on time, with the remainder taking nine months to complete. This shows that the defense industries in both Europe and the U.S. suffer from small scale and numerous bottlenecks. The entire system never factored in production capacity reserves; long-term low demand has led to a lack of market signals, and supply chain optimization has resulted in maintaining only the bare minimum of production capacity. Resuming mass production will inevitably involve a lengthy ramp-up period. But that’s not the worst of it; the greatest fear lies in the human factor—no one knows how to manufacture these items anymore. This is what is meant by “long-standing problems” and “people no longer understanding military affairs.” This isn’t the first time the U.S. has recognized this issue; it encountered similar difficulties as early as 2000 when modernizing its nuclear arsenal. In 2001, the U.S. released a new study indicating that extending the service life of nuclear warheads required replacing a secret material called “Fogbank.” Production of this material had ceased in 1989, and the veteran employees who knew how to handle it had all left. It took the U.S. five years to figure it out again. When they finally produced the material and found it worked, something still felt off—it turned out that the old process had omitted the removal of a certain impurity. And even the veterans from back then couldn’t figure out exactly what that impurity was. People have an unconscious reliance on original processes; that knowledge had vanished before it could even be codified. But that wasn’t the most critical issue. Both the U.S. military and the software industry realized that over-optimization can lead to a loss of knowledge. The RAND Corporation found that in the defense industry, the apprenticeship period for new entrants takes 2 to 4 years, and reaching the supervisor level takes 5 to 8 years. Submarine design requires 10 years of experience after earning a Ph.D. The software industry is no different: it takes entry-level programmers 3 to 5 years to become full-fledged software engineers, 5 to 8 years to become senior engineers, and more than 10 years to reach the level of chief engineer or architect. There are no shortcuts for companies; blind optimization will almost certainly lead to a talent shortage. Salesforce stopped hiring software engineers last year, and enrollment in computer science programs at 62% of U.S. universities is declining. In many software companies, the role of humans is to review code written by machines. The solution proposed by executives is to have machines review machines—leaving no trace of human intervention, with everything handled by artificial intelligence. This leads to a rapid increase in human dependence on AI, making people less likely to question the machine’s self-assured errors. METR conducted an experiment and found that experienced developers actually saw a 19% drop in efficiency when using AI tools for real-world open-source tasks, whereas AI was expected to boost efficiency by 24%—a gap of 43 percentage points. People have learned to deceive themselves; even when they notice that results fall short of expectations, they still believe that AI has helped them become 20% more efficient. Once AI tools are adopted, it’s hard to give them up. Junior engineers only know how to guide the AI; they lack the ability to even question its capacity to modify code. Without a comprehensive understanding of the system and codebase, they cannot accumulate tacit professional knowledge, making it impossible for them to grow into senior engineers even in ten years. So, will AI eventually take over entirely in the future? Perhaps, but let’s not forget that arms dealers once thought the same way, believing that peaceful development would always be the norm. That turned out not to be the case. AI may continue to advance and become exceptionally intelligent, but the issue of knowledge transfer will not be easily resolved.