Anthropic's Dario: Young People Should Learn to Command AI, Plus 24 Free Resources
In response to "what should high school graduates do," Anthropic chief Dario said that AI can already handle the work of junior lawyers, junior financial analysts, and junior consultants, and that young people should learn to command AI. Based on this, the author compiled 24 completely free AI and agent learning resources, divided into four groups: prompting, LLM fundamentals, agent building, and Claude-specific topics, including Microsoft's 21-lesson Introduction to Generative AI course, Hugging Face's agents course, CrewAI and AutoGen framework tutorials, and Anthropic's MCP special topic. The author believes that even after finishing all the courses, you may still be eliminated, and that you should directly use tools to build your own team of AI employees.
View original postI finally understand why ordinary people could never make money making indie games before: because the compute cost of every trial and error was too expensive. I tested Singapore FUMO Lab's automatic routing fusion model Fusion Max 2.0, and after four rounds of revisions without writing a single line of code, I got a 3D chubby orange cat house-destroying stress-relief game running, and the total bill was only $5.77, directly 2.6 times cheaper than Claude Fable and GPT-6 Astra. Ordinary people can finally afford to play the probability game of making a hit, haha. In the past, people went straight at it with flagship models, burned dozens of dollars over a few revisions, and before the game ever saw real players, their wallets were already drained. Honestly, after writing this retrospective, my biggest takeaway isn't how amazing any single model is, but that once automatic routing drives development costs down to cabbage prices, ordinary people can finally afford to play the probability game of making a hit. Many people making independent products always want to save up one big move and hope for a miracle on the first try, but business reality is extremely cruel: the indie teams that actually make money have often made over a hundred prototypes, with the first hundred almost all wiped out, until suddenly around the hundred-and-something one it works. In the past, everyone simply couldn't afford to try, because the compute cost of every trial and error was too high. My approach this time with Fusion Max 2.0 was actually very crude, but also very effective: first let it write out a complete 27,000-word gameplay and numerical design document in one go, then directly generate single-file 3D web code. I went in and played like an ordinary player, and when I ran into clipping, accidental touches, or cat fur that wasn't smooth enough, I just gave revision suggestions in plain language and had it rewrite the whole thing, and even had it casually write a Blender script to swap in a chubby orange cat with 24 layers of fur. The whole time I only acted as product playtester and quality inspector. I've always felt that once fusion models push the marginal cost of producing code down to just over five dollars, indie developers' core action must shift completely from grinding on code to low-cost, high-frequency trial and error. First smash out something playable, then use real playtesting and retention data to filter directions. After all, models can help you drive development costs down to cabbage prices, but whether you can survive the platform testing period and retain real players still depends on repeatedly grinding it out through low-cost iterations.
View original
我终于明白为什么以前普通人做独立游戏根本赚不到钱了,因为每一次试错的算力成本都太贵了,我拿新加坡 FUMO Lab 的自动路由融合模型 Fusion Max 2.0 实测, 连改四轮、一行代码没写,跑通 了3D 胖橘猫拆家解压游戏,总账单居然只有 5 块 77 美金,直接比 Claude Fable 和 GPT-6 Astra 便宜 2.6 倍,普通人终于玩得起爆款的概率游戏了哈哈, 以前一上来拿旗舰模型死磕,改几轮烧掉几十美金,游戏还没见到真实玩家钱包先干了, 说实话,写完这篇复盘我最大的感触根本不是哪个单模型有多神,而是自动路由把开发成本打成白菜价之后,普通人终于能玩得起爆款的概率游戏了, 很多人做独立产品总想憋一个大招、指望一次就出奇迹,但商业现实极其残酷,真正赚到钱的独立团队往往做过一百多个原型,前一百个几乎全军覆没, 直到第一百零几个才突然跑通,过去大家根本试不起,因为每一次试错的算力成本都太高, 我这次用 Fusion Max 2.0 的打法其实特别糙、但也特别管用,先让它一口气写出两万七千字的完整玩法与数值文档, 紧接着直接生成单文件 3D 网页代码,我自己像普通玩家一样上去疯玩,遇到穿模、误触或者猫毛不够顺滑, 直接用大白话提修改意见让它整份重写,甚至让它顺手写出 Blender 脚本换上带 24 层毛发的胖橘猫,全程我只当产品试玩官和质检员, 我一直觉得当融合模型把生产代码的边际成本压到五块多美金之后,独立开发者的核心动作就必须从死磕代码,彻底切换成低成本高频试错, 先把能玩的东西砸出来,用真实试玩和留存数据去筛选方向,毕竟模型能帮你把开发成本打成白菜价, 但能不能扛过平台测试期、留住真实玩家,依然得靠一次次低成本迭代去硬碰出来。
Body · AI translation
Let me give you a harsh reality: Anthropic's head Dario was recently asked what young people who just graduated high school should do right now, and his answer basically condemned junior lawyers, junior financial analysts, and junior consultants to death. His exact words were very blunt: AI has already learned how to do the work of junior lawyers, junior financial analysts, and junior consultants, and we can't deny that it will eliminate a large number of jobs. In just a few minutes, he pointed young people to only one way to stay in the game: learn to command AI. When I came across the post compiling that list of 24 completely free AI and agent learning resources, my first reaction was that this batch of resources is indeed comprehensive enough, and you don't even have to spend a single cent on tuition. But on second thought, if everyone really treats these 24 links the way they used to grind through grad school and civil service exams, going from the first to the last, most of them will probably still end up among the 95% eliminated by AI. The reason is very simple: in the past, schools taught us to honestly be a senior executor, spending a year memorizing knowledge points and practicing basic skills, then competing with others for a junior position. But today, when agents can already do work in batches 24 hours a day, competing with AI over who is better at basic analysis or who writes a first draft faster is purely using your flesh to crash into a high-speed train. The 5% who truly remain are never the people who finish all the courses, but the contractors who directly use these tools to build their own AI employee teams and take charge of giving instructions and accepting results. I have reorganized these 24 zero-cost official and top open-source resources by practical use. Everyone, don't be greedy; just click into whatever area you lack and start running it right away. Group one: first train your mouthpiece well and learn to use prompts to give precise instructions 1. Anthropic prompt engineering interactive tutorial, practice and revise as you go to master instruction structure fastest https://t.co/dvC60UeX2z 2. Anthropic official system course, transitioning from the first prompt all the way to orchestrating agents https://t.co/BWQlup9ZoH 3. Anthropic practical cookbook, ready-made code and business recipes you can use and modify immediately https://t.co/5w02XZzoma 4. https://t.co/2C2ux6EgXj prompt engineering guide, the most widely circulated prompt encyclopedia on the internet https://t.co/oryBJXpKL7 5. OpenAI official cookbook, ready-made code templates covering most high-frequency daily tasks https://t.co/4wptRLfIEh Group two: fill in the underlying common knowledge of large models and engineering from zero 1. Microsoft generative AI for beginners course, 21 lessons in total, no computer science degree required at all https://t.co/IWu6Y4Diub 2. Google Cloud generative AI practical notebooks, directly run cloud cases hands-on https://t.co/rNLRKAvGtd 3. Andrej Karpathy neural networks from zero to mastery, this set is enough if you want to thoroughly understand the underlying principles https://t.co/zcIpfim408 4. https://t.co/lMNHSavZln short courses, all one-hour essence lessons taught by frontline industry experts https://t.co/VS5j0VBBND 5. A collection of excellent large model applications, hundreds of real runnable open-source projects to take apart and learn from directly https://t.co/vkuDpoFU9e Group three: directly start building your fully automated AI employee team 1. Microsoft AI agents for beginners course, step by step teaching you to build agents that can do junior work for you https://t.co/YRKeUbCads 2. Hugging Face full agent course, free learning from zero to one with official certification https://t.co/GYcHuELvgU 3. CrewAI multi-agent framework, directly get several AI roles to divide work and collaborate https://t.co/hCiaY8og0D 4. Microsoft AutoGen framework, letting multiple agents automatically converse and collaborate to tackle complex tasks https://t.co/oczlTkWfeM Group four: Claude full suite official specialized practical courses 1. Master AI basics with Claude https://t.co/VXKcrZhM4G 2. Implement AI automation with Claude https://t.co/OpbbBXrcaA 3. Collaborate with AI using Claude https://t.co/yE6AhInViw 4. Engineering development with the Claude API https://t.co/4O3kej35V2 5. Anthropic Model Context Protocol MCP advanced special topic https://t.co/V5MtuFhdpH 6. Anthropic official AI fluency teaching guide https://t.co/9zCcJ0FpMd 7. Real-world AI specialized course for everyone https://t.co/r0RaJiHwl6 8. Practical Claude AI for marketing scenarios https://t.co/axnkZBJnIa 9. Learn Claude from scratch https://t.co/4o4LD9hxBN 10. Claude Code using agents for modern software engineering https://t.co/5DTm3HOvCW To be honest, all 24 of these resources are free and cost not a single cent, but the most valuable thing has never been the links themselves. It is that today you should hand your most tedious repetitive task to an agent to run. Before graduation season and the real workplace reshuffle hit, the earlier you learn to be AI's commander, the less anxiety you will have about being replaced by algorithms.
AI Trends