Large language models (LLMs) are revolutionizing how we interact with technology, but their computational costs can be a major roadblock. Imagine needing to constantly remind an intern of basic procedures and examples—it slows everything down. Similarly, feeding LLMs repetitive prompts during fine-tuning adds unnecessary expenses and latency. But what if your LLM could learn like an intern, internalizing those repetitive instructions? Researchers have introduced a clever approach called "PromptIntern" that does exactly that. By progressively embedding recurring prompt information directly into the model's parameters during fine-tuning, PromptIntern dramatically reduces the need for lengthy prompts during inference. This isn't just about trimming a few words—it's about teaching the model to understand the task's core requirements. In tests on complex coding tasks, PromptIntern slashed input tokens by over 90%, sped up inference by a staggering 4.2 times, and reduced monetary costs by a whopping 88.3%. The key is a progressive learning strategy. Initially, the model uses full prompts. As training progresses, repetitive elements are gradually removed, like an intern becoming more self-sufficient. Finally, the model performs flawlessly using only the core query. PromptIntern demonstrates that by focusing on efficient knowledge transfer, we can unlock the true potential of LLMs while keeping costs in check. This research opens doors to wider adoption of LLMs in cost-sensitive applications, paving the way for faster, more affordable AI solutions.