Andrej Karpathy, one of the most visible figures in artificial intelligence, has a simple tactic for getting work done. He sends long, rough memos to a large language model and mines the reply for signal. The approach, shared this week, spotlights how top researchers use AI not just to code but to think on paper, at speed.
Karpathy has led teams at Tesla and helped launch OpenAI. His comments land as companies race to fit AI into daily workflows. The method is simple, low cost, and easy to copy. It also raises fresh questions about privacy and the limits of machine feedback.
What He Said
Karpathy, a leading AI researcher, said he will send LLMs 10-minute long “total mess” memos, and calls the bots response “useful.”
That short statement points to a growing habit among power users. Instead of crafting perfect prompts, they dump raw thoughts and let the model sort, summarize, and spot gaps.
Why It Matters Now
LLMs are moving from novelty to office staple. Workers use them to outline briefs, tidy meeting notes, and draft emails. Karpathy’s habit reflects a shift from asking for answers to asking for structure. The model becomes a fast editor and an idea organizer.
The practice fits a broader pattern. Professionals feed messy inputs, then iterate. It mirrors how managers once relied on junior staff to shape notes into plans. The assistant is now software.
How The Method Works
Karpathy’s approach is straightforward. Record or write a quick, unfiltered memo. Paste it into an LLM. Ask for a clean outline, key points, and missing questions. Rinse and repeat.
- Start with raw notes, not polished text.
- Request an outline, risks, and next steps.
- Verify details and add sources after the draft.
This shifts the hard part from starting to refining. It reduces the blank-page problem, a common blocker in knowledge work.
The Upside, And The Catch
Supporters say the gains are speed and clarity. An LLM can pull themes from a noisy memo in seconds. It can also surface questions the writer missed. That aligns with how many teams use AI as a second pair of eyes.
Critics point to two risks. First, models can sound confident while being wrong. Second, sending raw notes to a cloud service can leak sensitive plans. Those are not small trade-offs.
Practical safeguards help. Keep confidential material off external tools. Use enterprise models with stronger privacy. Treat the first output as a draft, not a decision.
Industry Reaction And Use Cases
Product managers, consultants, and founders have echoed the tactic in public posts. They use LLMs to clean brainstorms, shape product briefs, and prep investor updates. The value is less about perfect facts and more about faster thinking.
Education and research teams report similar wins. A messy literature note becomes a reading list with themes. A field report becomes a clear status update. In each case, the human stays in charge of facts and judgment.
What This Signals For AI At Work
The trend suggests a near-term future of AI as a drafting partner. Models excel at structure, tone, and first passes. Humans handle facts, nuance, and final calls. This division plays to each side’s strengths.
Vendors are responding. Note-taking apps now auto-summarize. Meeting tools suggest action items. Office suites bundle AI that rewrites and reorders text. The market is moving from chat to embedded helpers.
Limits Worth Watching
Three limits stand out. Accuracy still varies by topic. Long uploads can miss context or import bias. And overreliance can dull a team’s writing muscles. Leaders will need norms for when AI drafts are allowed, and when a person must lead.
Karpathy’s comment lands with a clear message. Messy thoughts are not a barrier if a model can turn them into a plan. The idea is simple, and that is why it spreads. The smart move is to pair speed with care. Expect more teams to use LLMs as an early editor, while keeping final judgment, sources, and privacy in human hands.
