January 6, 2026
Speed thrills, bills, and spills
Shipping at Inference-Speed
One dev says AI ships code faster than he can think — cheers, jeers, and carbon fears
TLDR: An engineer claims he now ships apps at AI speed, trusting a slower-but-safer tool over a faster rival. Comments split between hype over 600mph productivity, outrage at a $51k spend and big CO2 math, and pragmatists warning the real bottleneck is human planning and oversight.
A lone dev just bragged he’s shipping code at “inference speed” — basically, the AI writes it while he sips coffee — and the comments lit up. He says the big unlock is a model he calls “codex,” which reads his whole project for minutes, then rewires it carefully; a rival called “Opus” is quicker but sloppier. He rarely reads the code now, just watches the stream and types “build.” Cue community split.
One user joked the thread was so wild the site went silent: “literally speechless.” A product manager painted the meme of the day: dev engines hitting 600mph while PM headlights barely reach the road ahead — meaning planning and direction become the real bottleneck. His claim that “most apps” are just moving data also poked the bear. On the other side, a critic did the math and dropped a carbon guilt bomb: claiming around $51k in AI fees in three months and a massive energy footprint, asking what was actually built.
Meanwhile, practitioners chimed in with “yep, it works” stories (GLM-4.7, DeepSeek) and begged for better ways to record AI’s thinking, while veterans warned that micromanaging the bot is a trap — but so is zero oversight. Verdict: speed demons vs conscience cops, with everyone agreeing on one thing: the human in the loop still matters.
Key Points
- •The author reports a major productivity increase in code shipping since May, attributing it to improved AI agent capabilities.
- •They default to building command-line tools so agents can run and verify outputs, making most tasks limited by inference time rather than coding.
- •A shift to GPT 5 increased trust in model outputs; the author focuses on architecture while reading less code.
- •Preferred languages are TypeScript (web), Go (CLI), and Swift (macOS/iOS), with minimal reliance on Xcode and use of Swift’s build infrastructure and the iOS Simulator.
- •The author contrasts codex (thorough pre-reading, better for large refactors) with Opus (more eager, better for small edits), moving away from “plan mode” toward conversational planning and execution.