Hi,
Recently my org onboarded us with claude code and there is a proper guideline passed that developers are not expected to code now and just review the AI written code, also the story points will be reduced to half for lets say a task took 3 SP only 1.5 SP will be given now.
The codebase is growing messy the developers around me just slap everything into the claude code and cant even make a line change without it.
What are your thoughts on this? And what is the future of developers? How can be optimize ourselves with the trend while also being technically sound and not slapping everything into AI.
You can't review any code without writing one prior. Future Sr Devs are doomed.
if not cse then what?
Just waiting for a huge ai mishap at this point..
this sounds a bit extreme tbh, like going from coding to just reviewing overnight rarely works well. if people stop understanding the code, things will break later for sure. ai should speed you up, not replace your thinking completely. i've seen teams use tools like claude, chatgpt, runable etc. but still expect devs to write and own parts of code. future is probably more "ai assisted dev" not "ai only dev"
The problem is if you can view and understand code then you are fine
Because AI didn't write messy code and making changes is easy
you show slight hesitance in the beginning, then give in, so when things finally shit themselves, you can be the one to subtly say I told you so
The clients will soon say we can give prompts to Claude ourselves and fire these companies.
AI tools for coding are getting added everywhere but you still need to review and correct the AI because even after clear instructions it can make mistakes also as a dev you still own the changes you are adding.
Let nature take its own course.
The current frontier models (Opus 4.6 and GPT 5.4) were trained on Nvidia H100 Hopper architecture chips. The next generation (Mythos / Spud) are trained on Blackwell chips which are 4x faster and 20x more efficient. They are expected to be 10T parameter models while the current ones are 2-3 T max.
Also, you may say that although the models are getting better, the inference costs are also rising for output tokens which is valid. But new research in memory efficiency with research papers like the recent Turboquant paper and other papers from the open source chinese labs are working to reduce memory usage of models by 4-6x. This will greatly drive down inference costs, making frontier models cheaper and slightly less capable open source models capable of running on our personal hardware.
All of this was just in this year by the way, and we aren't even half way done. The progress in AI is accelerating and just like the printing press or the sewing machine, writing code will be the domain of AI models and we will be the operators, except for specialised operations requiring niche languages or extreme memory efficiency which will still be written by hand.