Muhammad Umar explains that repeating a prompt to an LLM often produces better results. He describes why LLMs aren't deterministic like traditional software: their neural weights create apparent randomness, and attention allocation across your input shapes what they generate.
Just reiterated like so I mean if you give models the same prompt twice, like for example, if you are saying okay, can you do this? Or whatever you asked your model, just repeating the prompt prompt makes the model better. So that's a hack. I mean if someone is listening and they're trying to make model work better. So think about it like you would, you would give the same prompt and models performance would better. In the second question now there are two, three things that we need to understand here. Some of the people find it hard. How is it possible? Because they are coming from the era of determinism where they ask computer to do a certain thing and it exactly did what it was supposed to do. One is like no, I mean it means just talking to a computer and it would give different results. It's not randomness, it's how neural weights works, it's how human brain works. Right? That's at a macro. We need to get used to of this non determinism in the models. But the second thing I think the more important is that models are not just dependent on the context. It depends on the conditioning, which is like it depends on allocation of attention.
FDE Epi : 16 | AI Agents, Billion-Dollar Infrastructure & the Future of Intelligence
