Nate Meyvis shares four concrete ways he's using AI that actually work. He built a system to find recurring facts in a large text corpus by extracting them, generating embeddings, and clustering similar ones together. He stores school documents in Claude and reliably asks it to pull out dates and update his calendar. He's connected an LLM to his email and calendar to track trip logistics and flag problems. For studying, he uses the Zippyflash API through an LLM to generate printable flashcards for pen-and-paper review, get graded on his answers, and intelligently reschedule cards when his backlog grows. The real shift he's noticing isn't that AI got smarter at any single thing—it's that cheaper, faster models killed the friction. Three hours and $50 used to feel like the minimum investment for a task; now he'll spin up an AI solution for 30 minutes and $5. This changes what feels worth automating. The fact-extraction project and the data-gathering tasks (school info, trip planning) represent a jump from "AI can see your data and be sort of useful" to "AI can reliably handle this, though I'll double-check anything important." He's also deferring some projects, betting they'll be cheaper and better in a few months. Where AI still falls flat for him: high-judgment calls like "what books should I be reading?" or lifestyle advice that requires taste. As AI gets genuinely good at the grunt work—data triage, calendar management, the stuff a personal assistant would handle—it hasn't touched the judgment layer. He's clear about this gap. The models are becoming useful for the annoying, low-stakes decisions that drain time, not the hard calls that require real thinking.