The observation
Paul Ford, writing in The New York Times, notes that the expected wave of AI-generated killer applications has not appeared. Software projects still depend on people who can think together and apply practiced skill. Widespread access to AI code generators has instead made it easier to see why many teams should avoid shipping software in the first place.
Earlier assumptions
For a stretch of time the industry treated the arrival of tireless AI systems as a direct threat to roles like Ford’s. The premise was that automated generation would remove the need for coordinated human effort. In practice the outcome has differed. AI produces usable code in many cases, yet it also lowers the cost of performing someone else’s work at a lower standard. That shift helps explain the continued rate of project failure rather than a sudden increase in durable new applications.
What the work still requires
Ford points out that cutting-edge software continues to need humans who can align varied skill sets and maintain disciplines that current models do not replicate. When anyone can generate code, the gap between competent and incompetent results grows more visible. Projects do not collapse mainly from a shortage of code volume. They fail from the lack of coherent direction on integration, intent, and long-term upkeep. AI speeds the creation of surface features while leaving those harder problems untouched.
The piece ends with a short personal remark from Ford: he would like one killer app before a killer robot arrives. The line captures his view that current AI output remains derivative and does not yet supply foundational new systems on its own.
Reactions
No direct counter-statements from other developers or companies appear in the source material. The column stands as one practitioner’s assessment of the gap between earlier predictions and present results.
Why it matters
Tooling that generates code does not by itself create value. Organizations that treat the output as a replacement for design, review, and architectural decisions will keep producing systems that are difficult to maintain. Teams that keep clear human responsibility for trade-offs and overall direction stand a better chance of shipping software that lasts. The bottleneck was never the number of lines that could be written; it remains the ability to decide what should be built and why. AI multiplies both successful efforts and expensive missteps at the same rate.
This pattern affects hiring, project staffing, and risk assessment inside companies that build or commission software. When generation becomes trivial, the scarce resource shifts back to judgment about scope, integration, and long-term cost. Teams that continue to treat AI output as the primary deliverable will see the same failure modes they saw before, only reached faster. Teams that treat the tools as one input among several will continue to need the same coordination and craft that Ford describes. The difference shows up in the quality of the final system rather than in the speed of the first commit.
Ford’s column therefore functions less as a defense of existing jobs and more as a reminder of where the real constraints lie. Until those constraints move, the industry will see more code produced with less corresponding progress on the applications that users actually rely on.
---
Sources:
{"word_count": 612, "sources_used": 1}
No comments yet