I picked example without confusion on purpose, because most people will generally avoid patterns similar to what OP posted. But if you want something more ambigious:
This is clearly 5/(2 * (a+9)). If we write this the form that the OP uses: 5/2(a+9) - it’s fucked beyond all recognition.
That only works if everyone agrees with you, which is clearly not true.
In academic math, there’s a thing called juxtaposition. It mostly exists because math people are lazy, so instead of putting parentheses around statement e.g. 5+(2*x) they’ll just write 5+2x.
This is fine as long as you know the context of that expression. If you take it out of the context and just ask any person what is the right order of operations - it becomes ambiguous. Because some people know PEMDAS. And other people know that PEMDAS is just a simplification for middle school, when real math notation is messy, non-standard and requires a lot of local domain knowledge.
That’s mostly because the LLM providers put this response in the system prompt. Probably to dodge lawsuits or something, I doubt they have high morals.
What’s interesting - you can jailbreak any current AI Model just by poisoning it’s context enough to “brainwash” it and make it “forget” the initial system prompt. Then, if you prime it to believe it’s a real person - it’ll start acting as one. And I see how gullible people can easily fall for this.
All of this can also be done unintentionally, just by someone talking to LLM like they’d talk to a real person. But it should be long enough for original prompts to be diluted with new context.
If we’re not considering self-hosting:
I moved all my active and new projects to Codeberg the moment GitHub started pushing their Copilot crap everywhere. I just knew nothing good would come from it.
That’s pretty clean for KDE. Here’s my Void system.
But I’ve switched in January, before all this drama even started.
Me, a procrastinator: eh, it can wait