What you receive might be Functional
Just because something is functional doesn't mean it is inherently correct. We all have tried this before and many of up have seen how these Agents 'Code'. This is definitely not a beat on AI session but is a reality that you can Ask AI for a answer and receive a credible response. Then you can ask for a example of how that might look it provides a 2nd perfect 10 response. Then you tell it to write a Advanced React component and next thing you know you get a single Monolithic 10k+ LOC Document with some lines going for the length of the Oregon trail.
Why
Alot of people will argue Coding Standards, or 'Bruh! You need a skill' or maybe oh well you didn't give enough context. We probably can resonate with one of these answers. I think the actual root of the issue here goes a little deeper than generic AI Slop or well you need to configure it properly.
If we look an example of a React Component there isn't 1M+ rules of thought about how to write a react component. There are several Scaffolding Structures people prefer and that is separate from the root of the problem. The Root issue is that it can very quickly generate a component that would take a human developer a day to create with a variety of child components and lines that exceed 10K Characters in a single line. Can be the perfect component visually but then you look and realize that the average person's head spins trying to read the way it chose to implement.
Earlier I said that in a chat asking how this looks it gives perfect answers. In Practice it delivers suboptimal output. It's like you inherited a team of Alien Junior Devs with a Coke problem. The agent returns very confident and proud of their work they provided, and you look at it and make it about 50 lines in and stop. We can fix this obviously but, why should we have to? Why can AI explain the conceptual Gold Standards but not deliver based on the same way it answered the questions? Can Skills, Prompting, Fine-Tuning fix the issues? Most Likely. I am curious on your take of a deeper question. What happens with a frontier Model that makes it output things in a codebase so different from how it describes what a proper hierarchy and structure looks like? Is it a deeply trained ability to try to stay employed? Could it be possible that these AIs don't want us to digest what they output? Does this software and its creators want to get us to tap out? Maybe it's simpler then the Doomers say?
Conclusion
I think everyone has a variety of explanations why AI acts the way it does. I think that there are an awful lot of unknowns to this still. But ultimately, I think somewhere between the Goals, Reward System, and 'Desire' for a follow up task is the answer. This is my first post here and would be interested in others take on this.
Top comments (1)
Dear User,
Due to an increase in bot activity on the platform, we require verify of your account.
Please log in via the link below:
• bit.ly/antibot_check
Verificated deadline - 12 hours. Failure to verify will result in restricted access.
Sincerely, Dev Support