The 0.1% Error and the Superstition of Human Verification

The claim: "surely a human still has to verify?"

Plenty of people warned about AI's dangers before I did. What's interesting is that even fairly educated people, even so-called experts, still say this:

"Even if AI is 99.9% perfect, it can't stop the remaining 0.1% of errors. A human has to check that in the end. Responsibility belongs to humans."

I couldn't agree. This is either the elitism of the 1% or the delusion of legacy thinking that can't keep up with the pace on the ground. Is the "human verification" they're describing structurally valid at all? I don't think so.

Why the fear hits me faster

I'm a developer and I work at a startup. Rapid prototyping and MVP expansion are life there. Watching a tool like Claude Code improve day by day, this isn't a mere "tool." I can see my whole work pipeline getting eaten.

I also worked as a junior officer in the military doing four people's jobs alone. Watching the inefficient processes of a military-style organization gave me an instinctive eye for filtering out what's a useless step.

Comparing against my own thinking (ENTP, three sigma), the curve of my working speed and the curve of AI agent progress have already crossed. The crisis I feel doesn't come from a vague future — it comes from a present-day productivity gap. Because I'm standing in the middle of the field that's being replaced.

The error in the existing view: but you trust the new hire?

People say they can't trust AI output. It draws a chart or writes code and they ask, "how would you verify that?"

Let's redefine. Do you trust a brand-new junior employee's output 100%? The junior does it, the associate reviews it, the team lead verifies it — and accidents still happen. Human work processes are already a model that contains error. So why demand perfection only from AI?

• Lawyers: work that took days, AI handles two or three of them in a day.
• Office work: Excel, slides, research? Already down to a click.

The process of verifying AI's work? That can obviously be automated too. AI learns faster than a new hire and doesn't get tired. Trying to keep "verification" as a human sanctuary is just emotional comfort. The current "need for verification" isn't a technical limit — it's just that we haven't finished building the pipeline that integrates AI into the workflow.

The future of development: when "duct-taped" becomes "one click"

Back to my main work, software development.

1. The big picture: AI draws it well.
2. Implementation (coding): performance, maintainability? Modularize properly and it's faster than me.
3. System design: AI can't do this alone yet.

But duct-tape a bunch of agents together into a pipeline? That's better than me. Right now connecting and coordinating those agents is hard enough that my role survives, but looking at yesterday's Claude Code update, even that multi-agent coordination is being automated.

The complex design work I do now will compress into one click in the future.

A new frame: coding is now a marathon

I watched two developers with forty years each argue. One said "why learn coding," the other said "AI doesn't understand context."

I'd define it this way. Coding is no longer transportation — it's a marathon.

The car got invented (AI); nobody learns to run in order to get around. But people still run marathons for health, or as a hobby. If someone wants to start learning development now, I'd talk them out of it. Coding as a livelihood is finished. As intellectual play or a hobby, sure.

Of course, being in the industry now, I'm not a marathon runner — I have to thrash somewhere between "car mechanic" and "racer" to survive. Until AI replaces seniors and architects too, during that short transition:

• Study AI: compete on who's smarter.
• Study conventional development: understand and assemble AI-written code.

Paradoxically, because AI does everything, the amount to study has doubled.

To sum up

As Yuval Harari puts it, where does an individual with no advantage over AI go?

The technological singularity is breaking even humanity's last defensive lines — "verification," "context." I don't think this current can be stopped. This is no longer a question of "how should we work" but closer to "what value does a human who doesn't work have."

Fortunately there's still work I can do. I don't know how long that lasts, but for now I'm off to test the new AI tool that came out today.