Why AI is accelerating software development faster than our ability to understand it
Over the past few days, I listened to a podcast that revisited a concept Dario Amodei, CEO of Anthropic, has been repeating for quite some time.
An idea that sounds simple, yet completely changes how we should think about artificial intelligence:
humans reason in a linear way, while technological progress grows exponentially.
It’s also a recurring theme throughout his essay The Adolescence of Technology.
This is why we consistently underestimate what happens in the early stages of an exponential curve.
As long as changes appear incremental, the human brain interprets them as continuity.
Then, suddenly, everything accelerates at once:
- capabilities
- automation
- adoption
- economic impact
The turning point in software
In the world of software development, this phenomenon is already visible.
With modern AI coding tools, a single developer can now produce in a few hours an amount of code that, until recently, would have required days or even weeks.
But this is where a problem highlighted by Martin Fowler — one of the world’s leading experts in software architecture and modern software engineering — begins to emerge in his recent article The VibeSec Reckoning:
the speed of generation is surpassing the human capacity for understanding and review.
For years, we treated human code review as the primary security mechanism.
It worked because the pace of software development was still compatible with human cognitive limits.
But in the context of so-called “vibe coding,” that balance breaks down.
AI is now generating:
- code
- cloud configurations
- pipelines
- infrastructure permissions
at a speed that makes relying solely on manual supervision increasingly unrealistic.
The core argument in Fowler’s analysis is straightforward: prompts are not security controls.
Telling a model:
“write secure code”
does not guarantee a reliable outcome.
AI models optimize for the statistical probability of a response, not compliance with:
- company policies
- least-privilege principles
- infrastructure security standards
From the platform side
It is no coincidence that Anthropic recently introduced a security guidance plugin for Claude Code:
https://code.claude.com/docs/en/security-guidance
The goal is clear: adding deterministic guardrails to systems that are, by nature, probabilistic.
This represents an important cultural shift.
Security can no longer be delegated exclusively to developers or final-stage code reviews.
It must become part of the infrastructure itself:
- automated scanners
- policy enforcement
- continuous validation
- isolated sandboxes
- runtime controls
The industry’s new reality
The industry is beginning to accept an uncomfortable reality:
if software production grows exponentially, security must also become automated and exponential.
Because the real risk is not that AI writes code.
It’s that it does so much faster than humans can meaningfully understand it.
