From what I've seen, AI can code, but it is a tangled mess and hard to debug. Also, I read that AI costs more than human workers, so there is no advantage to using it other than faster productivity.
I think companies are starting to realize that big AI companies were partially subsidizing the costs and that it is more expensive than they realized. They will probably start to monitor token usage.
I use AI for a lot of coding and it takes effort and experience to reduce some of the excess cruft it likes to add.
The cost/ROI vs productivity thing I see on my team / hear from others is mostly that people are leveraging LLM code to do lower value work they previously never would have done. Lots of "I finally improved my test coverage" and "so I finally got to that ticket thats been in the backlog 2 years" and "I finally did that big refactor" and "I spent 2 weeks on the codebase making it easier for LLM to understand it" kind of stuff.
I have heard fewer anecdotes from people I know of the form "I do my job faster, shipping features 50% faster using LLMs". I think this is a project lifecycle thing too. An existing 5 year old project with users, stakeholders, product team, etc probably is going to null out any coding speed up.
I recall a very "agile" team I was on where we met with users weekly to get their requests, demo the changes made, and usually half the meeting was them changing their mind about what they said last week. I don't think delivering features 2x faster and meeting twice a week would have worked.
From what I've seen, AI can code, but it is a tangled mess and hard to debug. Also, I read that AI costs more than human workers, so there is no advantage to using it other than faster productivity.
I think companies are starting to realize that big AI companies were partially subsidizing the costs and that it is more expensive than they realized. They will probably start to monitor token usage.
I use AI for a lot of coding and it takes effort and experience to reduce some of the excess cruft it likes to add.
The cost/ROI vs productivity thing I see on my team / hear from others is mostly that people are leveraging LLM code to do lower value work they previously never would have done. Lots of "I finally improved my test coverage" and "so I finally got to that ticket thats been in the backlog 2 years" and "I finally did that big refactor" and "I spent 2 weeks on the codebase making it easier for LLM to understand it" kind of stuff.
I have heard fewer anecdotes from people I know of the form "I do my job faster, shipping features 50% faster using LLMs". I think this is a project lifecycle thing too. An existing 5 year old project with users, stakeholders, product team, etc probably is going to null out any coding speed up.
I recall a very "agile" team I was on where we met with users weekly to get their requests, demo the changes made, and usually half the meeting was them changing their mind about what they said last week. I don't think delivering features 2x faster and meeting twice a week would have worked.
https://youtu.be/WuhAaMSXD9A AI costs more than using Humans. In 2030, they might have lower-cost tokens.