A studio owner who has spent 30 years in software, 20 with game dev tools, and 10 in Unity says his team’s development speed went up roughly 5x in a matter of months. He is not celebrating. His shop builds simulators and industry training software, and the post is a cost accounting of what the shift did to his business — not an argument about AI art, which he rules out explicitly.

His observations, roughly in his own order:

  • ~5x speed almost overnight, and quality has held, because the team is senior and refuses to let the model make decisions about project structure or design patterns.
  • But the pressure to skip review is building. The generated code is better than most devs he knows can write, at 20x the speed, and “the temptation to just push and move on is huge and irresistible.”
  • No joy left. “There is no joy or sense of accomplishment. No pride in making something clever or quickly, and all progress is associated with the AI and not with you.”
  • The market is dead. Leads dropped off a cliff and nobody has budget — the speedup landed as table stakes in a shrinking market, not as headroom.
  • There is no going back. “If we stop using AI we’ll be out of business in a month.”
  • No room for juniors. Every junior they tried became “a meat proxy for AI while learning nothing, and getting in the way.”
  • Management broke. He can no longer reliably estimate a task, and nobody has a working model for running a team this way.
  • Nobody is learning. “To understand code you have to write it, and nobody is writing it.”
  • Everyone hates it. Veteran peers are planning early retirement and waiting to be laid off.

The most useful reply comes from a 20-year AAA developer, who names the mechanism behind the exhaustion. With implementation delegated, humans spend the entire day on nothing but high-level design and architecture decisions. The “manual labor” half of the job used to be the restorative half; take it away and the work that remains is the mentally heaviest part, all day, with zero downtime.

The counterarguments are worth keeping too. One commenter notes that controlled studies have never replicated 5x gains and that the real bill is deferred technical debt — check back in six months. Another argues token prices are a subsidy that will reprice, and the scarce skill will become managing token spend. The OP answers that even at 10x the current cost, it stays wildly cheaper.

What makes the post worth reading is that it treats the change as a labor story rather than a tooling story. The number that matters is not 5x. It is that the apprenticeship ladder stopped working, estimation stopped being a skill anyone can trust, and the craft satisfaction that kept people in this industry for decades has drained out — none of which any benchmark measures.