Physix FrontierWho Turns Generative AI Code into Experience?
A set of numbers in the DORA report is quite jarring. Generative AI adoption rose by 25%, yet delivery throughput actually dropped by 1.5%, and delivery stability fell by 7.2%. McKinsey's data says development time for complex tasks decreased by 12%, and medium tasks by 10%. MIT research also notes that less experienced developers are more likely to use AI, with more obvious productivity gains.
These sets of numbers seem to contradict each other, but they aren't. What AI saves is time spent typing, looking up APIs, and filling in boilerplate; it can't save the time needed to understand the system. I work on flight control and sensor fusion. Over the past month, I've been stringing together IMUs, BeiDou modules, and Arduino prototypes to run tests. Getting AI to generate a filtering framework was fast—within tens of minutes, I had a compilable shell. The trouble came later. When is the data trustworthy? When should we degrade functionality? When should we enter safe mode? AI won't take the blame for you.
In the short term, generative AI will raise the barrier for juniors and might hollow out their growth path. Previously, newcomers had to write interfaces, add unit tests, read logs, and argue with product managers about fields. It was slow, but these dirty jobs were entry points for gaining experience. Now, one prompt can produce a module, making them look much faster than people three years ago. The problem is, they skipped the process of filling in the pitfalls themselves. Code getting merged doesn't mean they know why it was merged that way.
In the long term, the question becomes: who will be the seniors? The value of senior engineers lies in judging risk when requirements are vague and information is incomplete. This kind of judgment is most feared in flight control. Motor runaway, IMU jumps, communication delays, OTA upgrade failures—often, the issue stems from engineering boundaries not being held. How do you develop a sense of boundaries? Only through repeated online troubleshooting, rollbacks, and post-mortems.
Here's a comparison.
| Stage | Where AI Helps | What Might Be Lost |
|---|---|---|
| Entry-level | Syntax, boilerplate code, test cases, API sketches | Debugging intuition, boundary awareness, accountability |
| Implementation | Draft solutions, documentation organization, cross-language assembly | Exception recovery design, long-term maintenance costs |
This table isn't necessarily entirely correct. I think the core issue is that generative AI has made "writing it down" cheap, but made "thinking it through" expensive.
Having worked on real-time systems for a long time, I care deeply about latency. If an interrupt is delayed by 1ms, the control law might become unstable. Code generation has latency issues too. AI output is fast, but if humans skip deriving things themselves, cognitive latency gets pushed into production. Once in production, being half a beat slow turns into an accident.
Companies shouldn't just look at lines of code committed, PR counts, or generated lines. They should look more at takeover rates, rollback rates, incident attribution, and whether new hires are still willing to do low-value tasks. A few days ago, I wrote a piece about FSD entering its sixth country—the key isn't intelligence, but compliance rhythm. Developer growth is also a matter of rhythm. Too fast, and the foundation isn't solid. Too slow, and the business can't wait.
Robotaxis are like flight controls: pretty demos don't matter; what matters is the rhythm when they're actually running. Developer growth is the same—you need to see if they can independently handle a module tomorrow. How many extra lines of code were written today isn't that important.
Looking ahead, junior roles may decrease but won't disappear. What disappears are roles that only convert prompts into code but cannot convert code into judgment. The remaining juniors will be like early flight control testers, having to learn faster how to read logs, waveforms, and incident reports. Just reading model outputs isn't enough.
📌 This article is compiled from Hacker News. Original text: https://eng.snu.ac.kr/en/communication/promotion/news?md=v&bbsidx=8355
All rights reserved by the original authors. This is a compilation and independent analysis based on public reports.