First Principles
Writing code became a commodity. Validating it didn’t.
Carlos Mattos, CTO at GFT Technologies, on the bottleneck that moved — and why the companies cutting juniors are making a mistake.
15 min watch
Carlos Mattos
Chief Technology Officer, GFT Technologies
Leads engineering across Brazil, Colombia, Costa Rica and Mexico, delivering for banks and financial institutions where a production failure is not an inconvenience.
What this conversation is about
GFT started building on generative AI in 2023, aimed at a specific problem: their clients are stuck on legacy mainframe, legacy Java, legacy .NET — systems built twenty years ago, largely undocumented, with the business rules living in the heads of developers who are now retiring. The average COBOL developer is 55. When they leave, the rules leave with them.
AI can extract those rules. That part turns out to be the easy part. The hard part is validating that what AI extracted is actually correct — and that gap, between the speed of generation and the speed of human verification, is the thing Carlos keeps circling back to.
It’s also the thing most teams are getting wrong right now, and the reason has less to do with tooling than with what happens to a nervous system under mandate.
Three things to do differently on Monday
1.
Move your value to validation, not generation.
AI can produce plausible code faster than you can meaningfully read it. Carlos’s warning is that when review can’t keep pace with generation, validation quietly degrades into a rubber stamp — someone glances, it looks right, it ships, and it breaks in production under load. In a bank, that isn’t a bad sprint. The engineer who can genuinely slow down and validate is the one the organization can’t automate away. Start reading code more than you write it.
2.
Get good at defining what AI should not do.
Carlos’s line: the best thing you can do with AI is define what it does not do for you. A bad specification produces technical debt at machine speed. The senior skill is no longer prompting harder — it’s drawing tighter fences and holding a clear picture of the boundary before the generation starts.
3.
If you’re junior, ignore the layoff theater.
Carlos is blunt about the companies announcing they no longer need juniors: he calls it a bad decision, and notes some are already hiring people back to fix what AI broke. The required skill mix genuinely changed — systems thinking, critical thinking, communication, fluency with the tools. But if you never hire juniors, no one becomes senior in five years. The seat still exists.
“We moved the bottleneck from writing code to validating code.”
— Carlos Mattos
Host’s note — Dharma Ramasamy
Carlos described something he didn’t name as biology, and I want to name it.
He said the resistance from senior developers is natural — that change forces people out of a comfort zone they’ve occupied for twenty years. That’s true, and it’s more literal than it sounds. When an experienced engineer is told they must use a specific tool, under a specific governance policy, at a specific speed, the perceived loss of autonomy registers as threat. Cortisol rises. The prefrontal cortex — the part of the brain that actually adapts to new tools — goes quiet. Dopamine anticipation, which drives voluntary skill acquisition, drops off.
What the org calls “resistance to change” is a well-organized survival response in a body that hasn’t been given a reason to feel safe. You cannot mandate your way past it, because the mandate is the thing producing it.
Which is why Carlos’s answer to my last question is the most important thing in the conversation. Asked what one signal he’d want to see about his team that he can’t see now, he didn’t ask for a productivity number. He asked what is blocking them from being open to the transformation. That’s the right question. It’s also the one almost no measurement system in any engineering organization is built to answer.
Be on the series.
Five more of these this quarter. Twenty minutes, two cameras, nothing to prepare, and you see the edit before it publishes.
Get in touch →The signal Carlos wanted.
CultureGuard reads existing collaboration metadata to surface where human capacity is actually draining — before delivery slips. No surveys, no AI, no content ever read.
How it works →