Introducing Devin, the first AI software engineer
This release has been in the rounds for a few days scaring a lot of junior devs and raising some dollar signs in the eyes of some managers that have not seen the other arguably better projects like GPT Pilot and now think that junior devs can be replaced.
Based on the demos that Cognition Labs has shown us this seems like a more polished well designed version of tools like GPT Pilot or even Auto GPT.
Devin does have features that are not as simply available in the open source alternatives which makes devin’s demos feel so much more impactful and practical, seeing how it can take care of deploying the projects it builds or “learn” new tech through documentation.
The Demos
The demos are impressive but seem very hand picked and seem to not show Devin actually generating the responses, especially in demos like:
Our AI software engineer fixes a bug in Python algebra system
Here Devin is trying to contribute to an existing project, here they start off the demo with an already pre run example, we don’t see Devin in action but just see the end result.
This can make it seem that Devin just casually fixed bugs in your existing codebase, but by the timestamps we can see that Devin took almost an hour to find the bug.
This project that they chose for this is one of the ideal cases for models like GPT 4 that are very good with python, so the demo seems like it was setup to make sure that Devin could actually find a bug in an existing codebase.
So basically the demos were very well picked for their marketing and I hope they can work on it more to make it faster cuz it seems like an actually useful tool to develop with, like how GPT Pilot too has been doing.
You can checkout more on GPT Pilot with one of the best creators for AI right now Matthew Berman
And he also created a follow up video on the new features of pilot’s new version few days before Devin released here:
What Devin cannot do
There are a lot of things that Devin can do by just becoming more faster, like iterating faster and building smaller projects of fixing things in existing projects much faster than any human can.
But there are few things that only real humans seem to be able to do right now which is understand what humans want, which is arguably something that we ourself dont always know, AI relies on the user to prompt it in a specific way to get what we want.
Humans are cheaper for Software Engineering
At least as of right now humans can be trained for cheaper than AI to do complex tasks that are in context of your company or and project (this cost estimate is an assumption as, if AI was actually cheaper then this would have already been implemented by the bigger companies to save on cost, so this could be wrong.)
But we can be sure that this will change over time, and we can only hope that becoming better software engineers can help us stay relevant.
Conclusion
As of now all projects that are being used in production have been built by humans to solve human problems, so unless AI can understand humans it doesn’t seem likely that it will be able to build a solution for a human problem better than a human.
Software engineering is more than just writing code and there are many aspects that AI needs to be fine-tuned for apart from coding to replace a full fledged engineer.
We right now use AI to build our projects faster but this can change to AI using us to build projects faster, this should not stop anyone new to software engineering from learning it but should be a reason to be more excited about its future of Software engineering.