Obviously not? It would be absurd to walk into a thread about Rust and say “you are all morons, Rust doesn’t increase your productivity and unless you can produce a study proving it does then your own personal anecdotes are worthless.”
Why the increased demand for rigor when it comes to AI specifically?
At what point is an ‘extra’ 25% coding overhead worth it to ensure a sane human reasonably concerned about criminal consequences for impropriety read all code when making it, and every change around it? To prevent public embarrassment that can and will chase off customers? To have someone to fire and sue if need be?
[Anecdotally, the inflection point was finding tests updated to short circuit through mildly obfuscated code (introduced after several reviews). Paired with a working system developed with TDD, that mistake only becomes obvious when the system stops working but the tests don’t. I wrote it, I ran the agents, I read it, I approved it, but was looking for code quality not intentional sabotage/trickery… lesson learned.]
From a team lead perspective in an Enterprise space, using 25% more time on coding to save insane amounts of aggressive and easy to flubb review and categories of errors sounds like a smart play. CYA up front, take the pain up front.
I wonder how many 10x AI bros were 1/10th engineers slacking off most of the week before the fun new tech got them to actually work on stuff.
Obviously not all, and clearly there are huge wins to be had with AI. But I wonder sometimes..
One of the reasons software is in decline is because it's all vibes, nobody has much interest in conducting research to find anything out. It doesn't have to be some double blinded peer reviewed meta analysis, the bar can still be low, it just should be higher than "I feel like"...
Now you are moving it because your statement is provably false.
Your criticism of it is based on vibes. What specifically is wrong with the methodologies?
One of them broke randomly developers into two groups, one with access to ai and one without, timed them to complete the same task, and compared the results. That seems fine? Any measurement of performance in a lab environment comes with caveats, but since real world accounts you dismiss as vibes, that seems like the best you can do.
With assembly versus say, Go for writing a web server? That's trivially observable, good luck arguing against that one.
If you have something that needs to be done, and an agent goes and does the whole thing for you without mistakes, it is trivially observable that that is useful. That is the definition of usefulness.
Then you can research Rayleigh scattering, of which consists of a large body of academic research not just confirming that the sky is blue, but also why.
But hey, if you want to claim the sky is red because you feel like it is, go ahead. Most people won't take you seriously just like they don't take similar claims about AI seriously.
[0] https://scied.ucar.edu/image/wavelength-blue-and-red-light-i...
then they put the vibes on a graph, which presumably transforms them into data
55% faster task completion using predictive text
Quality improvements across 8 dimensions (e.g. readability, error-free, maintainability)
50% faster time-to-merge"
how is time-to-merge a vibe?
Shows that devs overestimate the impact of LLMs on their productivity. They believe they get faster when they take more time.
Since Anthropic, GitHub are fair game here’s one from Code Rabbit - https://www.coderabbit.ai/blog/state-of-ai-vs-human-code-gen...
The only objective measures I've seen people attempt to take have at best shown no productivity loss:
https://substack.com/home/post/p-172538377
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-o...
This matches my own experience using agents, although I'm actually secretly optimistic about learning to use it well