There's a culture war happening in tech right now. Some people are claiming they can ship features 100x faster and software engineers will be obsolete in 6-12 months. Others are claiming AI cannot and will not ever be able to write code that is even remotely usable in a production setting.
So I got curious and decided to run a survey. 596 responses later, here’s who’s using AI and how:
95% of ALL respondents are using AI to write code.
I was expecting a majority, but this was a STAGGERING number. Of course, this was a fairly small sample size, and my data is skewed by selection bias, on and on, but even all that taken into consideration, this is near-universal adoption.
This question opens a lot of other ones in my mind. Are people actually confident in these results? Are that many students using it too? How much faster are people developing now? Why aren’t the 5% using it?
Students
166 students responded. Here's what stood out to me:
1. The majority of students prefer an IDE to an Agentic Solution. I am actually the same way. More than happy to use AI to code, but I still want to read and approve the code myself.
2. 93% use AI to write code. This is slightly lower than the 95% overall total, but still near-universal adoption.
3. Boilerplate is where they find the most value. This makes sense to me, as no one wants to write boilerplate and AI is quite good at it.
4. Testing is where the find the LEAST value. This is a bit surprising, as no one I know likes writing tests by hand, and they are also a place where AI shines. I chalk this up to the fact that most students don’t write a lot of tests or understand how important they are yet.
5. Perceived speed improvement is very high. Students rated their speed improvement at around 3.6/5 (this corresponds to roughly 7x faster based on the indicators I provided in the survey). This feels high to me, but these students probably never learned how to code without AI so the baseline to that 7x is probably very low.
6. Only 3 people fully trust AI generated code. I was expecting this number to be a little higher from students. Seems they are maintaining a healthy skepticism despite the wide adoption.
Engineers
396 Engineers responded. Here's what they said:
1. The majority of engineers prefer an IDE to an Agentic Solution. Same as the students. With all the hype around Claude Code on X and LinkedIn, I was expecting a higher level of Agentic Solution usage from these folks.
2. 95% use AI to write code. If you’re in enterprise, you’re using AI to write code. This is consistent with anecdotes I’ve heard from industry. Some of this probably comes from the top down, and some of it probably comes from IC preference. Looking back I wish I had a question on this.
3. Boilerplate is where they find the most value, but the rest are distributed fairly evenly. Seeing this more even distribution across tasks was not surprising, I did design this survey with common SWE tasks in mind. I was surprised to see debugging as high as it was. Most of the bugs I dealt with at Amazon were either simple enough that I didn’t need AI, or too complex for AI to solve (system level bugs between services).
5. Perceived speed improvement is very high. Engineers rated their speed improvement at around 3.5/5 which was just .1 lower than the students. In terms of feature development I generally feel a 20-30% increase, but that includes planning, architecting, meetings, oncall, and all the other stuff that comes with being a SWE. I suspect engineers answered this question based on code speed alone.
The 5%
5% of respondents aren’t using AI to write code. I figured this would be a low number, so I was able to collect free form responses directly from them:
"While I can see the potential benefits of a backboard for ideas, by having an AI write code for you, you are dulling your code writing skills and instead sharpening only your code fixing skills. The problem isn’t when writing code with an AI works, it’s the potential security issues and bugs that and may be included that are now harder to track down because you did not write the code yourself.
"
I need to get a grasp on how code works normally, as otherwise I won't be able to correct AI when it makes mistakes.
I'm generally faster writing it myself with IntelliSense code-completion and keybinds. I also prefer to come up with solutions myself and really understand what's going on in a codebase.
I really enjoyed this exercise and I hope you all enjoyed seeing the results! There were some flaws in my survey and some questions I wish I’d asked, all lessons I am taking away for next time.
Feel free to reply to this email directly with other ideas for polls or similar content if you found this useful or interesting.
And of course, thank you to those of you who took the time to fill out the survey in my initial email. Stuff like this is not possible without you!
See you next week.
Arjay