Any chance you could rewrite this and update it based on the latest releases of Claude and OpenAI? In particular how it impacts AI completion rates. Thank you!
The TDD part is an interesting wrinkle. Part of me says that generating tests is cheating the process. But what you say rings true, the process of prethought is important. I can imagine the tests are really valuable as one looses a LLM to make changes on a codebase leading to lots of regressions.
Yeah it sounds like cheating to me too, but rationally it makes sense. Writing tests is ~50% of the dev effort, so if AI creates them for ~free, even if they are 80% as good as human ones, it looks like a good deal.
I think eventually 90% of the code will be written by AI (as Anthropic's CEO said last week), and engineers will work on higher level tasks: good system design, translating requirements into specs, technical strategy, buy vs build, etc.
But I don't believe this will happen anytime soon — even if the tech was there, companies are super slow to adjust
This resonates a lot. A number of the observations match what we've experienced first-hand and are learning at CodeYam and in chats with engineering leaders and their teams. "Augmentation" / Iron Man is a good framing; we've talked about building a JARVIS for software developers on our team with the goal of ultimately giving you super-powers (augmentation), not eliminating you. The challenges of AI navigating complexity and TDD quiet renaissance are also points that hits home. Thanks for sharing these insights.
Incredible article. Thank you!
Any chance you could rewrite this and update it based on the latest releases of Claude and OpenAI? In particular how it impacts AI completion rates. Thank you!
will definitely write more about this as soon as updated benchmarks come out!
btw from my personal use I am not seeing a big difference (at least for coding) between sonnet 3.5 and 3.7
i feel we are getting in the land of diminishing returns
The TDD part is an interesting wrinkle. Part of me says that generating tests is cheating the process. But what you say rings true, the process of prethought is important. I can imagine the tests are really valuable as one looses a LLM to make changes on a codebase leading to lots of regressions.
Yeah it sounds like cheating to me too, but rationally it makes sense. Writing tests is ~50% of the dev effort, so if AI creates them for ~free, even if they are 80% as good as human ones, it looks like a good deal.
Amazing Article!
sure, ai often fails due to lack of context. but once there is infinite context is solved - im assuming in <6 months - what happens to software dev?
see Google’s Chain-of-Agents Framework (January 2025), InfiniRetri: Infinite Context Processing (February 2025), Google’s Titan and Internal Gemini Build (January–March 2025)
I think eventually 90% of the code will be written by AI (as Anthropic's CEO said last week), and engineers will work on higher level tasks: good system design, translating requirements into specs, technical strategy, buy vs build, etc.
But I don't believe this will happen anytime soon — even if the tech was there, companies are super slow to adjust
TDD, Docs & Quality angle are great. I can see where I can improve my workflow. Thanks for sharing!
This resonates a lot. A number of the observations match what we've experienced first-hand and are learning at CodeYam and in chats with engineering leaders and their teams. "Augmentation" / Iron Man is a good framing; we've talked about building a JARVIS for software developers on our team with the goal of ultimately giving you super-powers (augmentation), not eliminating you. The challenges of AI navigating complexity and TDD quiet renaissance are also points that hits home. Thanks for sharing these insights.
Thank you Nadia!
Amazing article. When I first used Cursor AI I called it the metrics. I liked your name the Iron Man more and you inspired me to start writing also.
Went and wrote about it for the first time. https://open.substack.com/pub/roiezra/p/the-iron-man-mentality?r=supoi&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true