On: Making o1, o3, and Sonnet 3.7 hallucinate for everyone

[L-four]

The trick to coding with LLMs is not caring if the code is correct.

[andrelaszlo → L-four]

Call me cynical but coding at some companies has such perverse incentives that I kind of get it:

- "Solve" the issue assigned to me with a bunch of code that looks about right. Passes review and probably not covered by tests anyway.

- Once QA or customers notice it's not working, I can get credit for "solving" the bug as well.

- Repeat for 0 value delivered but infinite productivity points in my next performance review.

[nsoonhui → andrelaszlo]

I suppose that you are using a dynamic language? Static typed languages have less of this problem.

[andrelaszlo → nsoonhui]

This is a hypothetical "I". Personally I am deeply passionate about delivering shareholder value, producing high-quality code, enthusing stakeholders, tabs vs spaces, and so forth...