The Secrets To Making LLMs More Reliable

Hey prime, do not hallucinate!

Welcome to the era of probability coding. Let's just generate enough code until one of them works.

Happened to me on few dozens of occasions: Me: Ask a question LLM: Here's your answer Me: Are you sure? LLM: Oh, sorry, no, this is the right answer. Me: Are you sure that the first answer was not the correct one? LLM: Oh, sorry, yes, the first answer was the correct one.

Incredible advice. I also hear if you write the solution in the prompt the llm will be really good at finding the solution

My elderly neighbor asked me to look at his printer. Everything was fine including full ink cartridges. It turns out he had canceled a charge to HP a few weeks ago that he didnt understand. He paid something like $5 dollars for two years for cartridges he had never used and could not use unless he reactivated his subscription. I buy printers at work. I have not bought an HP printer since then. I don't know any large company that still uses HP.

If someone tries to mug you on the street, just say no and remind them that's against the law.

04:51 "You wanna be as clear as possible" Yeah, someone should invent a way to tell the computer what to do in a way that describes the intended technical behaviour better than normal english.

The printer rant was refreshingly cathartic and absolutely spot on.

I imagine Sam Altman to throw a Jira ticket in panic "FEED IT ALL TOWER DEFENSE STRATEGY GUIDES YOU CAN SRAPE"

User: "Don't hallucinate ! " LLM: "Done."