The Last-Mile Challenge for LLMs

In February 2026, Elon Musk stated, “Code itself will disappear.” He suggested that a world is coming in which AI directly completes programs, replacing source code written by humans.

Today’s software consists of two layers: human-readable code and machine-executable instructions. Code serves as the blueprint, while the binary is the finished product. Musk implied that even this blueprint may no longer be necessary.

AI is indeed evolving rapidly. OpenAI and Google DeepMind describe a shift toward “agentic” systems—AI that can make decisions and carry out tasks autonomously. For simple apps and tools, we have already entered an era in which software can be generated automatically from natural language.

However, reality is not so simple. While AI can write code, complex modifications and accountable management remain difficult. Even in practical benchmarks such as SWE-bench, performance has improved but is far from perfect. “Hallucinations,” in which AI outputs incorrect information with confidence, still persist.

In short, a world that moves directly from human language to a finished product is unlikely to arrive anytime soon. In practice, an intermediate design layer—one that structures and clarifies human intent—will likely remain necessary.

At the same time, there are already cases in which AI optimizes internal structures. Google DeepMind’s AlphaDev discovered algorithms faster than those designed under conventional human assumptions, and these have been adopted in real-world software. The era in which AI finds more efficient internal designs than humans has already begun.

What matters here is transparency. If intermediate design steps become invisible and only the final output remains, how do we verify its safety? How do we ensure it has not been tampered with? Where does responsibility lie?

The more advanced technology becomes, the heavier the burden of “proof of trust.” Just as food products require labels indicating their origin, software will increasingly need mechanisms to show its history and provenance.

There is also the question of whether jobs will disappear. The International Labour Organization (ILO) suggests that AI is more likely to augment work than replace it outright. What will become crucial is the ability to define what should be built—and the ability to explain and verify it.

Strictly speaking, coding will not “die”; its role will change. The center of gravity will shift from writing to assuring, from implementing to taking responsibility. This may also create a role for blockchain technology.

Musk concluded by asking, “What would you create if there were no constraints?” If the cost of implementation approaches zero, what remains is human intention. The more AI becomes a magic wand, the fewer excuses we have. If the day comes when coding disappears, it may also be the day when the full responsibility for creation returns to humans.

Related Articles

The movement toward the integration of social media and finance accelerates.

Interactive Brokers Launches USDC Deposit Service