75% Haven't Really Used AI. Here's What Power Users Do Differently.
Andrej Karpathy dusted off an older observation this week, and the numbers still land hard: napkin math puts roughly six billion people — about 75% of the population — as barely having touched what large language models can actually do. The gap, he said, is no longer two camps talking past each other. It is a sharp funnel.
Aaron Levie quoted him and named the everyday version of that funnel: most people still use AI as a chatbot for simple interactions that approximate next-generation search. Useful, yes. Representative of the full capability? Not even close.
I keep thinking about the other side of that funnel — the people who are getting compounding value. What are they doing that everyone else is not?
The Gap Is Not Only About Models
It is tempting to blame the gap on model versions, paid tiers, or “prompt engineering.” Those matter. They are not the main story.
The quieter difference is how people treat the conversation itself.
Casual use looks like this: open a chat, ask a question, get an answer, close the tab. The thread dies in a sidebar. Next week you ask a similar question again, from scratch.
Power use looks different. The chat is not a vending machine. It is a working document — a place where research paths, decisions, drafts, and half-formed reasoning accumulate. The people getting leverage do not throw that away. They keep the good threads, search them later, and feed pieces back in as context so the next round starts smarter than the last.
If chats are where real thinking happens, they deserve to be saved like documents — not lost in a product sidebar.
Disposable Chats Are an Invisible Tax
Every closed platform trains you to treat conversations as temporary. Switch from ChatGPT to Claude and the history stays behind. Cancel a subscription and years of “I figured this out once” can vanish. Looking for the argument that convinced you last month means scrolling, guessing, and often giving up.
That tax is small once. Over a year of serious use, it is enormous. You pay it in repeated prompts, forgotten decisions, and context that never compounds.
Search-like chats feel free because each one is short. Compounding work feels expensive until you notice you have been rebuilding the same context over and over.
What Power Users Actually Do
Strip away the mystique and the pattern is surprisingly simple:
- They keep the threads that mattered — not every joke prompt, the ones with real reasoning.
- They treat those threads as reusable context — paste, attach, or load them when starting related work.
- They build a personal library of how they think with AI, not a pile of one-off answers.
That is the jump from “AI as search” to “AI as a thinking partner with memory.” The model did not change overnight. The user’s relationship to the transcript did.
Levie’s point about underrepresenting capability sits here. If your only mental model is a single-turn Q&A box, you will never see what happens when a model works against a pile of your prior work — your product notes, your failed approaches, your preferred voice, your unfinished plans.
Portable History Is Underrated Infrastructure
Here is the part the industry still underweights: owning the transcript is infrastructure for that next step.
If your best conversations live only inside one vendor’s product, your compounding advantage is rented. Export that is awkward, incomplete, or locked to one format quietly keeps you in the disposable-chat habit — even if you are a “power user” in every other way.
What you want is boring and powerful at the same time:
- Conversations as files you can open anywhere
- A format that is readable by humans and tools
- The freedom to search, organize, and continue without asking permission from a chat UI
That is why I care about open, Markdown-based conversation files — and why I built tools around them under LocalThink.
A Practical Path: Save → Organize → Continue
LTF (LocalThink Format) is an open .ltf.md standard, released under CC0, so an AI conversation can be a file you own.
LTF Builder is a free Chrome extension (no signup) that saves chats from ChatGPT, Claude, Gemini, and Grok into that format.
LocalThink for macOS (Apple silicon) is the local library on top: search what you saved, organize it, and continue when you are ready.
The workflow is intentionally unglamorous: Save the conversation that mattered. Organize it where you can find it. Continue from there instead of starting from zero.
You do not need to become a different person overnight. You need one habit change: stop treating the best chats as disposable.
Closing the Funnel From Your Side
Karpathy’s 75% figure is about exposure. Levie’s chatbot-as-search point is about how people use what they have already tried.
Both point at the same opportunity. The people pulling ahead are not only chasing newer models. They are accumulating personal thinking — and making sure it survives the next tab, the next vendor, and the next year.
If you are already in the thin part of the funnel, protect that advantage. Save the conversations where the real work happened. Make them portable. Feed them forward.
That is how “I asked ChatGPT something once” turns into a system you actually own.