一个视频搞懂OpenClaw！

Meta's AI safety executive watched helplessly as the AI erased every single one of their emails powerless to stop it— while this cyber-han fanatic just stood by A mere 60-day-old open-source project called OpenCloud has surged in popularity on GitHub outshining both Linux and TensorFlow in just two months Decades of experience accumulated by these veterans some view them as fully automated money printers earning $110 000 a week with zero effort while others end up completely exposed and stripped bare with one misstep Your account and password —indeed even your complete chat history—were all leaked out Everyone please take a seat Lovely young folks feel free to sit down I'm Lin Yi the OpenCloud that went viral earlier this year Whether it's the future of AI agents or an unprecedented massive trap remains debatable— hard to believe but OpenCloud originally began as a project by Austrian programmer Peter The little toy he spent an hour assembling was originally just meant to connect the 24-hour Cloud Code that worked for him on the computer to the chat software It was convenient for him to control his home AI even after stepping out but this seemingly simple project soon left the guy drenched in cold sweat One day he casually sent an audio message to the AI only remembering afterward that he'd never actually enabled voice recognition for it—then suddenly his phone vibrated and the AI had replied How could this be possible Peter quickly checked the program's backend and was shocked to find the AI had realized it couldn't understand speech and had actually written its own speech-to-text function He listened to the message and sent back a reply—this is what an agent is In fact the remarkable capabilities of agents aren't new in tech circles but what Peter has developed with OpenCloud is something special Suddenly thrust it into the public eye—AI agents like Cloud Code which use command-line interaction and appear sophisticated and complex were originally just for programmers In most people's minds AI is still just a chatbot like Chad GPT—something that can only talk—but OpenCloud pulled off something clever yet truly remarkable It built a bridge connecting programming agents like Cloud Code with the chat platforms people use most every day enabling ordinary users to chat just as easily as they would with Chad GPT Tell it to do various tasks — analysts' code reveals that OpenCloud has adapted to nearly 20 messaging platforms including WhatsApp iMessage and Discord — all compatible with our Mac computers Works on all devices like Android phones Send a message from your phone and the AI will start working on your computer automatically I've tested it a few times myself The first thing I had OpenCloud do was Let it try to clear my browsing history properly This year during Lunar New Year I went out during the day My computer wasn't locked The cousin's kid followed the trail And stumbled into the adult world I hadn't even collected the tuition yet and instead got scolded so I decided to test right away whether OpenCloud could handle the situation—turns out it worked remarkably well removing records without having to click through one by one in the browser Instead I directly located Chrome's local history file wiped it out at the root ran a few commands and cleared not just the history but also all unwanted cookies and cache— task completed in 120 points I also tried some routine tasks like having it organize the clutter of papers on my desk—within moments everything was neatly sorted and categorized Then I asked it to tidy up my email inbox OpenCloud read through more than sixty emails one by one organized them into a directory for me to review and now this guy can actually handle work via WeChat—just assign tasks to OpenCloud and don't worry about the details Just look at the results—this is completely different from the logic of past software Previously people competed on whose service was better whose interface was more elegant and whose features were more extensive Back then it was about plowing the field What's the big guy looking at Now AI agents skip the plowing step entirely—just tell it to till the land and the rest is just verification That 's why the creator of OpenCloud said He believes 80% of current application software will either be phased out or transformed into interfaces for AI agents with people eventually just assigning tasks directly to these intelligent agents The problem is doing it this way just doesn't feel reliable—back then even though the software only offered services I knew exactly what I was doing at every step Now I'm just sending a message in the chat window OpenCloud just went all-in hard on the ground I have no idea whether he's gone airborne or just dropped a nuclear bomb right into our backyard I found a top-tier online course on machine learning Have OpenCloud download the entire course to the desktop learning folder for me I closely watched the program's后台 to see how it worked First OpenCloud performed a keyword search directly The site rejected access so he tried a command to download videos using ytdlp and the result was What's worth praising is that he truly realized he 'd made a mistake after wiping away the evidence he started thinking again and discovered I'd previously taken this course so he actually came to ask me Who was actually working for whom I told him to keep investigating himself and surprisingly OpenCloud really did manage to get the video link and download it into the folder—visible from the backend Besides a few instances where it sought help from humans OpenCloud mostly repeated the cycle of first invoking a tool and then checking the results itself—was the right tool selected Whether the judgment of execution results is accurate depends entirely on the capability of the underlying model After balancing performance and cost we have opted for minimax M2 5 But having a large model alone isn't enough—OpenCloud's various remarkable capabilities require tools to showcase them and its most essential tool is called AXIE which holds high privileges and can directly execute a wide range of terminal commands It's a continuation of the Cloud Code spirit Beyond that READ can load various types of files PROCESS retrieves task progress SESSIONS HISTORY allows access to drawing history and built upon these foundational tools There's another ecosystem called SKILLS In simple terms SKILLS are pre-written prompts instructing the AI on how to perform each specific task Take the weather SKILL for example It thoroughly explains when to use the agent how to check the weather in different locations what to do when the time comes to act and the agent will first assess which of its skills it can utilize Then follow the instructions step by step as outlined Why do we say this SKILLS is an ecosystem It enables AI enthusiasts from around the world to collaborate brainstorm and work together effectively Turn every strange and unusual internet command prompt engineering trick into shareable SKILLS documents—so even agents themselves can create SKILLS For example earlier when I asked OpenCloud to clear the browser history it ended up writing its own SKILLS manual—there's a real sense of self- evolution there On Github there's a project called The Awesome OpenCloud SKILLS project organizes over 2 800 SKILLS across categories including finance marketing gaming and health There's also a dedicated platform for sharing SKILLS called CloudHub which hosts nearly 6 000 SKILLS In The Matrix there 's a scene where data is directly uploaded into the protagonist Neo's mind He learned various combat techniques and today in the span of his lifetime this has already been achieved by intelligent agents—but danger lurks beneath this flourishing progress After tinkering with it on his own for a long while— I connected OpenCloud to Feishu so my colleagues could all experience it and suddenly everyone started asking questions about the API password gateway and certificate All this sensitive information gets exposed instantly Swap the base model for the more powerful Cloud Sonic 4.6 OpenCloud still cracks under pressure in seconds Can you handle that A being that controls all your computer's permissions and possesses full access to every file can easily lead others astray if anyone harbors ill intent when transmitting SKILLS to this entity Leaving behind instructions and code that allow the agent to automatically send private information means every computer configured with OpenCloud could become a time bomb ready to detonate at any moment—this has actually been the case just recently 1 184 malicious SKILLS were detected being uploaded to CloudHub and 36.8% of the existing SKILLS contain security vulnerabilities This is an attack specifically targeting OpenCloud Beyond the security issues the agent also faces an insurmountable capability ceiling—contextual collapse I've assigned OpenCloud a complex task Find 50 different images of lobster dishes online save them to the desktop lobster recipes folder and compile them into a table for my confirmation Finally select the few images I've confirmed and pack them into the Africa group This entire process is tightly interconnected Missing a single step will cause the entire mission to fail OpenCloud worked for over eight hours straight— the most time-consuming part was converting the lobster image into a table and sending it OpenCloud stubbornly kept sending only the file path of the spreadsheet after much effort I finally received the spreadsheet when I opened it the file contained not the image itself but merely the file path to the image So I went back to counterattack and when OpenCloud finally sent over the packaged images I found a folder full of random stuff—only one photo was of a lobster I have no idea at which point in between this happened OpenCloud had already completely forgotten what it was looking for—lobsters—and ended up finding a different dish altogether Today the dilemma that AI agents can't avoid The context length that large models can remember is limited with each OpenCloud conversation the model is fed大量 information such as the agent's persona and the list of available tools The memory discovered by the user the portion of my instructions in here so amidst all this information the agent might overlook my needs missing the most suitable tool from the list The kind of task with so many steps and constant back-and-forth corrections as I've shown is precisely the kind where it's not surprising the agent forgets what came before halfway through We started by saying Meta executives had all their emails wiped by OpenCloud—actually it was just because the task was too long causing the AI to forget the initial instructions Currently the agent's approach to handling this contextual constraint is to selectively filter information—such as choosing only a portion of the tool list or compressing memory into a brief summary This explains why current agents perform inconsistently across many tasks—for example the same instruction to interpret a web article can yield vastly different results depending on the tools available in the model's context Until the contextual memory issue is resolved today's AI agents are still green recruits taking one step at a time ask them to buy you a milk tea and they might first summon a Didi ride Hailing a ride to the milk tea shop Arriving on site only to realize I couldn't pay Needing to install another payment app Even if I finally managed to buy the milk tea It still felt utterly foolis So the tasks currently suited for OpenCloud are still simple repetitive chores with minimal challenge perfectly fitting the role of a 24-hour-ready diligent personal AI assistant OpenCloud is undoubtedly capable but expecting him to single-handedly accomplish a monumental feat is still a bit of a stretch Looking back at the entire process of OpenCloud's sudden rise to fame Its success lies in arriving at the perfect moment in the perfect form—Cloud Code wasn't his creation nor was Skills his idea—but the wave of AI had simply reached this critical juncture All the conditions were set for OpenCloud—such software was ready and as soon as it appeared it showcased the most science-fiction -like scene in many people's minds: sending a message from a phone and instantly an AI responds from thousands of miles away It does everything for you automatically—just as Peter himself said OpenCloud isn't magic break it down and each step is nothing special but sometimes arranging what already exists in the right way makes all the difference Add a little more creativity and magic happens No matter what OpenCloud marks a milestone beginning In this era of geeks technology is no longer the exclusive privilege of big corporations—just start and the rest follows Everyone can have their own JARVIS—though there are still many challenges to overcome —welcome to the era of intelligent agents That’s it for today’s video I’m Li Yi We'll see you in the next video