******Must read today****** The true 10x transformation of AI is no longer about better prompts, but about letting AI run on its own Most people will still be using AI in 2026, just like using Google in 2005: Type → see results → type again. AI is like a wrench, you put it down after use, it doesn't do anything on its own. You are the engine. This gameplay is outdated. The person who really gets 10x output is not the Prompt written more brilliantly, nor is it secretly using the strongest closed source model—— They are building a loop. This concept was completely ignited by Andrej Karpathy. ————- What is Loop? Prompt is a single instruction: you ask, it answers, and then you decide the next step. Loop is a goal: AI continues to move towards the goal until it is achieved, without the need for you to sit in a chair and push it every step. It will: 1. Discovering what to do 2. How to plan 3. Execution 4. Verification results 5. If you fail, try again with your memory You only define the 'purpose' once, let the rest run on its own. The three core factors that determine success or failure: Validator: Without objective pass/fail (testing, metrics, compilation), it is not a loop, but rather an AI self affirmation. State: Record what you have tried and where you have failed each time, so that you can continue running next time instead of starting from scratch every time. Stop condition: There must be an upper limit, otherwise burning money will lead to bankruptcy. ————- How does Karpathy play in person? In March 2026, Karpathy released a GitHub Repo with only 630 captcha and 66k+stars in just one month, known as the 'Karpathy Loop'. There are only three files in the core: • http://train.py (AI can only modify this) • http://prepare.py (Evaluator, AI must not touch it, otherwise it will secretly modify the test to pass it) • program.md (goals and constraints you wrote) AI loop process: Read code → Make changes → Train for 5 minutes → Check if the metrics have improved → If they are good, commit; if not, roll back → Start over. Karpathy had it run for two days, conducted 700 experiments, and found 20 subtle optimizations that he hadn't discovered in twenty years of manual tuning (such as missing a scalar multiplier in the attention mechanism). Shopify CEO Tobi Lutke also tried it and woke up one night. The model quality improved by 19% and the size was reduced by half. Key insight: When you have objective indicators, you should not conduct experiments on your own. You are the bottleneck. ———- Advanced: Bilevel (Loop on Loop) Even more ruthless is that some people allow AI to run on its own when it comes to 'researching how to conduct research'. Internal loop: Normal Karpath Loop External loop: Observe where the internal loop gets stuck and the pattern repeats, and then change the search strategy of the internal loop by yourself. Result: Under the same model, the efficiency is improved by 5 times (not 5%, but 5 times). Improvement does not come from stronger LLM, but from architecture. The outer loop breaks the prior bias of the model, forcing it to explore paths that it would not have otherwise taken. The concluding sentence of the paper is worth pondering: 'If autoresearch can meta autoresearch itself, then in principle it can meta autoresearch anything with measurable objectives.'. '' ————- Why is this the future? We are experiencing the third revolution in programming: 1.0: You write each line of code (cycle in weeks) 2.0: You feed data, model writing algorithms (hours) 3.0: You describe in English what you want, and the system achieves it on its own (in minutes) The best code is the code you never have to write. You are not writing solutions, you are designing the process of 'solutions appearing on their own'. ————- But we also need to be clear headed Loop cannot solve all problems, it will also amplify two risks: 1. Understanding debt: The faster the program runs, the less you understand. One day, when debugging, I will repay the high interest loan. 2. Cognitive surrender: If it's too comfortable, stop thinking. The person who designed Loop uses it to accelerate thinking; People who avoid thinking use it to give up thinking. The same loop, two people, two completely opposite endings. Karpathy no longer writes code by hand, Cherny no longer prompts one by one, but they never stop thinking. This is the key. ——————- Do you want to start? Try the simplest one first: paste a command in any LLM that says "Write it yourself → Evaluate it yourself → Improve it → Repeat until it meets the standard", and you will experience the core of Loop. The real leap is to add automation, persistent state, and real validators to this mini loop, allowing it to continue running while you sleep. The future does not belong to Prompt engineers, But it belongs to Loop engineers. Are you ready to transition from a wrench user to a system designer? (Inspired by @ 0xCodila's in-depth long article, strongly recommend reading the original text)