The article argues that while AI tools like Claude and GLM-5.3 offer a 10x speedup in code generation, the overall development process is bottlenecked by specifications, reviews, and QA, making the actual delivery improvement less significant than claimed.
We have been on agents for a good while and the code generation side, that is like a solid 10x for us. i will give it that. The 10x on its own though, it does not really tell you what is going on. That 10x is not free either, you pay for it before you ever see it. Specs take way longer now. And the amount of talking before anyone even touches the repo, way up. And the pipeline, the whole delivery setup, we had to basically rip that out and do it again for how the agents work. That one nobody really tells you. You figure it out on the way, and usually the way you figure it out is you break something and go oh, right. And the 10x depends what you are measuring against. One coding task on its own, yeah, the 10x is real. The whole thing though, idea to a release you actually trust, it is more like 10 or 15% for me. The bottleneck did not go away. It just moved somewhere else, specs, review, QA, working out whether the thing was even worth building. That is the stuff that slows you down now. Step back from it though and it is less impressive than it sounds. Software being faster does not really do anything to the economy. Not until it starts changing how actual physical things get made, or moved from one place to another. And we are nowhere close to that. GDP does not care how many apps go out a week if it all stays in the digital side of things. Tooling wise i run Claude and GLM-5.3 depending on the loop. Opus i keep for the reasoning and the planning. GLM-5.3 does the boring heavy stuff, rebuilding all the context, going through these huge files, and it actually stays coherent across the 1M window which is really why i use it for that. Depth though, anything that needs actual depth, back to Opus. It is not close on that. Agentic loops just are not mostly hard reasoning anyway, they are mostly context. So yeah the typing got a lot quicker. The thinking is pretty much where it was. And whether you ship any faster really comes down to whether your specs and reviews were already in decent shape, because the speed did not fix any of that. It just made it more obvious where you were already weak.
The author, running a small dev team, shares mixed real-world results from using AI coding tools: they speed up boilerplate and onboarding, but produce confident wrong answers on complex problems and increase code review workload, yielding modest net gains far below the often-cited 10x improvement.
A study by METR found that experienced open-source developers using AI tools (primarily Cursor Pro with Claude 3.5/3.7 Sonnet) took 19% longer to complete real-world issues, contradicting both their own expectations and expert forecasts of 24% speedup.
A measured 7.8% productivity gain from AI across hundreds of engineers contradicts the 10x hype, with many gains fading, fueling backlash over benefits not being shared with workers.
Abi Noda of DX and Brian Houck of Microsoft share early findings from DX's research on AI's impact on engineering velocity, revealing a modest 10-15% increase in PR throughput, far below the 10x hype. They discuss why coding is only a small part of developer work, the risk of 'false velocity', and opportunities for AI beyond coding.