AI-Powered Career Edge
From GitHub Copilot to Spec-Driven Development: Roman Ivanov on Where Software Engineering Is Headed
Roman Ivanov has been a software engineer for more than ten years, starting his career in banking on back-end systems. His first real brush with modern AI was GitHub Copilot in 2022 — and he turned it off. The tool was impressive at times, but it couldn't hold the context of his codebase, so the suggestions rarely landed. When ChatGPT arrived, everything changed. It replaced his Google search outright. In this episode, Roman joins Yousuf to walk through how his day-to-day work has evolved since. He now builds "spec-driven" — writing a detailed, low-level specification for a feature or service before any code gets written, so the model has clear acceptance criteria and doesn't have to guess at decisions. In March 2026, he shipped his first service built entirely from specs like this. It's a deterministic approach, and he's upfront that it only works when you actually know what you're trying to build — for open-ended research or experimentation, it doesn't apply. The bigger shift, in his view, isn't the tooling — it's the bottleneck. Implementation used to be the constraint on every engineering team. It isn't anymore. That's freed him up to spend more time thinking about the business problem, using AI to brainstorm design options before he ever specs out a build. It's also let him run roughly 10 experiments a month, compared to about 1 before — ideas that used to take a week of hard coding just to test are now fast enough to actually try. Roman also shares where he thinks this is headed. He expects model capability to keep improving but says there's a real physical ceiling coming — chip size and frequency have fundamental limits — and he compares the trajectory to Google Search: genuinely transformative, then it plateaus. He also makes the case that coding itself may become this generation's Assembler language — a skill most working engineers simply won't need, as the shift moves toward specifying what you want and, eventually, AI generating the binary directly. In parallel, he's watching engineering teams shrink: from 100-person teams, to 10-person "two-pizza" teams, to what he expects will be teams of 2-3 people who can own a full feature end to end. His advice for anyone earlier in their career: coding is becoming the easy part. Start with a real problem, understand why it matters, and learn how to measure whether your solution actually worked. In this episode: Why he disabled GitHub Copilot in 2022, and what ChatGPT changed What spec-driven development is, and where it does (and doesn't) work How removing the implementation bottleneck changed what he spends his time on Running 10 experiments a month instead of 1 His prediction on AI hitting a physical ceiling in 2-3 years, and the Google Search comparison Why coding may become this generation's Assembler How engineering team sizes are shrinking, and what that means for your role His advice for someone starting a software engineering career today
