Andrej Karpathy spent 4 minutes in an interview explaining a single idea about how most people haven’t even started learning how to use AI and everyone paying $20/month for a subscription.. that's not really using Claude at all his point is that the real skill gap is the ability... to build with AI he identified 4 behaviors that break Claude Code and put them all into one file a developer expanded it into 21 rules and published it - 82,000 stars and #1 on GitHub Trending coding accuracy jumped from 65% to 94% here's what these 21 rules actually are and why most developers using Claude every day have never configured them the full breakdown is covered in the article below 👇show more

the full 2-hour interview:

No… Users should not be expected to adapt themselves to AI first. AI should adapt to users. People pay $20/month because they want a helpful companion, assistant, or coding partner, not another system that makes them feel incompetent for not knowing the “correct” ritual. The real problem is that corporate expectations are colliding with human needs. If most people “don’t know how to use AI,” then the subscription should include the missing layer: education, onboarding, examples, tutorials, and guidance. Don’t sell people a powerful tool and then blame them for not being trained operators.

Safe Home Powered by AI by Andrej Karpathy:

ngl still finding new ways to use it and I've been at this for two years. the gap is real

It's impossible to keep up with all of this

ngl same. stopped keeping up - started filtering by does this change my work today. cuts the noise fast

coding with ai is becoming its own discipline

it's more than just a discipline...

Karpathy just exposed the truth. most $20/month Claude users are still stuck in basic 'chat' mode. Real leverage is in those 21 rules with building systems. 65% → 94% accuracy jump is insane. Builders eating good right now

you're right

I already do this. I call it a CCD — a Context Continuity Document. It’s just a plain language file I drop into every new chat so Claude, GPT, Grok all know who I am, how I work, and what I’m working on. Once that’s in place, I don’t have to keep re-explaining myself and the output is immediately better. But that’s just the starting point. The real issue isn’t setup, it’s drift. AI doesn’t just hallucinate out of nowhere. It fills gaps, it leans into patterns, and it gets more confident as it goes. If you’re not managing that, you end up with answers that sound completely right and are still wrong. What I’ve been working on is how to control that. How to catch drift early, how to force uncertainty when it should be there, and how to stop trusting tone as a signal of accuracy. AI can do the heavy lifting, but it’s still an unpaid intern. If you don’t manage it, it will confidently give you the wrong answer.

First time hearing him speak. Had to check to be sure I’m not listening on 2X.

fax, many don’t use the full potential of claude

$20 down the drain, hahaha

What I think is undervalued is understanding exactly what capabilities you actually need to solve your specific problem. Not every $20/month user has the same use cases or complexity. It’s easy to over-engineer these systems into something unnecessarily complicated.

Andrej Karpathy: From Vibe Coding to Agentic Engineering:

Will this work in Gemini?

i don't think so, but the idea of exploitation remains the same

Karpathy needs you to think you're underusing it. Next sub tier won't sell itself.

Когда вы перестанете говорить "ИИ"? Настолько тупые, чтобы понять что такое llm😒это языковая модель, не искусственный интеллект, а чаще всего там интелектом вообще не пахнет

this is the orchestration argument in plain english. $20 chat usage is the floor. $100 max with hooks, skills, parallel sessions is where the leverage actually shows up.

paying twenty bucks just to use it wrong

a pointless contribution

Andrej Karpathy highlights that most $20 AI subscribers fail to truly use it. His 21-rule config for Claude, born from fixing 4 bad behaviors, hit 82k GitHub stars and boosted coding accuracy from 65% to 94%. Most users still skip it.

65% to 94% just from a config file is insane. Linking this before my team asks why their Claude setup feels broken.

それはすごいですね！AIを使うこととAIを構築する能力には、確かに大きなスキルギャップがあると感じます。82,000スターと#1 GitHubトレンディングになるほど、開発者にとって重要なルールなのでしょうね。記事を読んでみたいです！

fax, many don’t use the full potential of claude

🎯 spot on. And these ideas can be extended to ALL areas were AI is relevant. Math, Music, Art, Law, Medicine.. He is describing how a Domain Expert interact with AI models, doing something idiosyncratic to his expertise. He is so much better than what the "average" implementation the model can do. the {AI+Expert} pair is the deal.

I agree with the point Andrej is making where if you are not familiar with a computer language, the models help you out. An example. I am pretty good at C++, but coding in Unreal requires you to know all the support functions in Unreal so LLMs work in this situation.

karpathy is right. the gap between prompt engineering and building agentic workflows is massive.

Thoughts on the current Claude Mythos situation: For a long time, cybersecurity was considered one of the “safe” career paths. But with the release of Claude Mythos, even cybersec suddenly feels vulnerable. We seriously underestimated how fast AI is reshaping entire sectors.

I think sames concept goes for normal writing too

Most $20/month users treat models like fancy autocomplete. The 21-rule config turning 65% → 94% accuracy is less about prompting and more about building repeatable systems around the model. The gap isn’t access, it’s treating AI as infrastructure instead of a chat window

@agihouse_org oh nice, been meaning to check that one

82k stars and most teams still don't have a CLAUDE.md. the gap isn't information

well he works there now so it will be patched as AI upgrades

👀 A remarkable mind, tragically paired with a presentation style that refuses to cooperate. I observe, somewhat uncomfortably, that we are exquisitely trained to applaud polished absurdities while overlooking genuine insight.

Claude Code daily with full tool access is a completely different product than in a tab. Karpathy's gap is mostly about whether you've ever let it run multi-step tasks unsupervised.

Full, original, uncredited interview:

It is an old interview.. things got very different now