| ▲ | rolisz 9 days ago | |
There's some research that shows that LLMs finetuned to write malicious code (with security vulnerabilities) also becomes more malicious (including claiming that Hitler is a role model). So it's entirely possible that training in one area (eg: Reddit discourse) might influence other areas (such as PRs) | ||