ChatGPT peut désormais voir votre argent (c'est inquiétant)

OpenAI now wants to manage your money inside Chat GPT. Tencent is building a Jarvis style AI that can control your computer and phone. And Claude just helped someone recover five lost Bitcoin worth nearly $400,000. AI is getting into places we never really expected. So, let's talk about it. Starting with Open AI, they just dropped something that's either going to be incredibly useful or slightly terrifying depending on how you look at it. They're rolling out personal finance integration directly into chat GPT. And I'm not talking about just answering generic budgeting questions anymore. This is the full deal. Proous users in the US can now connect their actual bank accounts, credit cards, investment portfolios, loans, all of it, straight through chat GPT. They're using Plaid for the connections, which supports over 12,000 financial institutions right out of the gate. And they're planning to add into it support next. The way it works is pretty straightforward. You either click on finances from the sidebar or just type something like at finances connect my accounts and chat GPT walks you through the authentication process. Once you're hooked up, it syncs everything, categorizes your transactions, pulls in your balances, investments, liabilities, the whole picture. Then you get this dashboard showing portfolio performance, spending breakdowns, subscriptions you're paying for, upcoming payments, all that good stuff. And from there, you can actually have conversations with ChatGpt about your real financial situation. Here's where it gets interesting. Instead of chat GPT giving you those generic responses like, "You should probably save more money," or, "Here's how budgeting works in theory," it can now look at your actual spending patterns and give you specific advice. You can ask it things like whether you can afford that vacation you've been thinking about or if you should be worried about a particular investment or how to clean up your subscriptions without manually going through every service. It can help you plan for major purchases, analyze travel costs based on your actual spending habits, or figure out realistic savings goals that make sense for your income and lifestyle. They're defaulting to GPT 5.5 thinking for these financial conversations, which makes sense because money stuff is complicated and context dependent. OpenAI actually built an internal benchmark for personal finance tasks and had over 50 finance professionals from major institutions evaluate how well the system performs. GPT 5.5 Thinking scored 79 out of 100 on that benchmark, while GPT 5.5 Pro hit 82.5. So, we're talking about pretty solid performance on complex financial questions. Now, obviously, everyone's immediate concern here is privacy and security, right? Open AAI is being pretty upfront about what chat GPT can and can't do with your data. It can see balances, transactions, investments, and liabilities, but it can't see full account numbers, and it definitely cannot make any changes to your accounts. You're not going to wake up to chat GPT randomly transferring money around. You can disconnect accounts anytime through settings, and once you do that, your synced account data gets deleted from OpenAI systems within 30 days. Temporary chats won't access your financial accounts at all, which is a nice touch. They also have this concept of financial memories, which is where chat GPT remembers important context you've shared, like if you mention you're saving for a car or you owe your parents money or you've got a mortgage, that stuff gets stored separately, and you can view or delete those memories whenever you want. The whole experience follows whatever model training settings you've already chosen for chat GPT. So, if you've opted out of training, that preference carries over here, too. The really ambitious part is they're working with partners like in it to move beyond just answering questions into actually taking action. So eventually you might be able to go from asking chat GPT about credit card recommendations to understanding your approval odds to actually submitting an application all within chat GPT. Or you could ask about tax implications of selling stock and get connected with a live tax expert through in it. That's the long-term vision anyway. Right now, it's just prousers on web and iOS in the US, but they're planning to expand to plus users and eventually make it available more broadly. It's one of those features where you can see the potential, but it's also going to take some getting used to. I mean, we're talking about handing over your entire financial picture to an AI system. Some people are going to love that convenience. Others are going to be like, absolutely not. Meanwhile, over in China, Tencent is building what they're calling an operating system level personal AI assistant named Marvis. And this thing is wild. The name is a combo of Ma and Jarvis. Kind of paying homage to the Iron Man AI assistant, and they've designed each agent as a little black pony with horns wearing a scarf, which is honestly pretty cute. But here's what makes Marvis different from a lot of the other AI assistants we've seen. This isn't just about chatting or generating text. Marvis actually operates at the OS level on Windows PCs and Android phones. With iOS and Mac OS support coming soon, it can directly control applications, execute programs, manage files, generate documents, all through multiple agents working together. And it's online 24/7, supports cross-device operations, and can handle some genuinely complex tasks. The team behind this is 10 cents yingyong bao team and they've got over a decade of relationships with companies like Intel and Microsoft. That's given them the technical foundation to build something that goes deeper than most AI assistants. Marvis can operate computer exe software and also directly control mobile phone apps on your computer through Yingyong Bao. They've got authorization from apps like Flush, Kypon, and Vipo among others. What's really cool is they've optimized the edge side model at the chip level, getting a 20% increase in running speed, and the system can directly query and modify computer files through prompts. Most AI applications can't do that kind of in-depth operating system level stuff. Marvis offers two operation modes. There's a cloud-based efficiency mode using the latest models like Hunuon and Deepseek V4. And then there's a local privacy mode using the Quen Edge model, which means no data leaves your computer. They also support a local knowledge base that automatically identifies and indexes local documents and pictures. So when you need to search for something, Marvis can do title-based or contentbased matching. The interface is pretty fun, too. You've got this virtual office where you can see what each agent is doing. If an agent is assigned a task, it'll sit at its workstation and work. If agents are idle, they might doze off, wander around, work out, drink coffee, or even go to the toilet. When you assign a task, an agent supervisor comes out to take charge and delegates to the appropriate agents. After they finish, they report back to the supervisor. In testing, Marvis showed some impressive capabilities, but also some limitations. One test involved finding a picture containing NASDAQ elements in a local download folder. The file agent first tried searching by file name, realized most picture names were meaningless, then switched to using image analysis. It found the right picture, but ended up consuming 2 million tokens in the process after analyzing over a 100 images. So, it works, but it's not exactly efficient yet. Another test asked Marvis to change the computer theme to green. It found the system setting entry for theme colors showing it can access system level controls, but it stopped at finding the entry and didn't complete the actual change. The user had to do that part manually. When asked if the computer could run Blackmith Wukong, Marvis accurately pulled the system configuration, analyzed the specs, and identified that the CPU, memory, and hard drive were outdated while the graphics card was sufficient. It suggested upgrading memory as the most cost-effective improvement, though it got the pricing wrong, saying 32 GB DDR4 memory costs 4 to 500 yuan when it actually costs over a,000. For more complex tasks, Marvis can handle things like developing a local knowledge management system with custom large model access and rag functionality, all without the user touching code. In one test, it created the directory structure, wrote config files and core modules in parallel, built a web interface, and delivered a working system. Another test asked it to research Silicon Valley layoffs this year, and create a PDF report with data and sources. Marvis used multi- aent collaboration, conducted multiple web searches, and generated a detailed research report over 3,500 words with accurate data. Though during execution, the agents had an issue where they automatically switched to English, possibly due to the skill content or underlying model characteristics. Marvis will ask for user confirmation before doing potentially dangerous things like batch file deletion or modifying core configurations, which is reassuring, but overall you can see this is still early days. The speed, token consumption, and flexibility all need work, but the foundation is there for something that could genuinely act as an AI worker handling real tasks for you. Now, the third story is probably the most insane one, and it involves Bitcoin, Claude, AI, and $400,000. An ex user going by CPR KRN posted that Claude helped them recover five Bitcoin worth nearly $400,000 that had been locked away for over 11 years. These coins were bought back when Bitcoin was trading around $250. So, you can do the math on how much value that represents today. The backstory is pretty relatable in a kind of painful way. Back in college, this person changed their wallet password while drunk and then completely forgot the new password. They spent years trying to recover access using brute force software and massive password combination searches, none of which worked. After more than a decade of failed attempts, they decided to try something different and uploaded files from their old college computer into Claude. Now, here's the important part. Because a lot of people online started calling this an AI hack or saying Claude cracked Bitcoin encryption, and that's not what happened at all. Claude didn't break any security or decrypt anything. What it did was act like a really good research assistant, sorting through years of forgotten files, old backups, and different wallet versions stored on the device. During that process, Claude identified an older wallet.dat file that appeared to exist before the password change happened. The user also had access to an old pneummonic recovery phrase, and that combination allowed the wallet to open without bypassing any security protections. So, Claude basically helped organize and analyze a bunch of old data until it found the piece that mattered. The users celebrated in their viral post and thanked Anthropic CEO Dario Amodore, saying the experience was life-changing. And honestly, yeah, I bet it was. $400,000 showing up after you thought it was gone forever would be pretty life-changing for most people. This story highlights something that's becoming increasingly relevant in crypto. There are millions of Bitcoin believed to be inaccessible because people lost passwords, hard drives, or recovery phrases during the early years when Bitcoin wasn't worth much and nobody was being super careful about security. Now, those forgotten wallets potentially hold enormous value, and recovery has become a real industry. Tools like BTC Recover already exist to help people test password variations and recover damaged wallet access, but those often require technical expertise. most average users don't have. And that's where AI could become really useful because it can help by scanning old files, backups, and patterns faster than humans. It's not hacking anything, just finding clues and messy data. So, if you hold crypto, back up your wallet, secure your recovery phrase, and never rely on memory alone. All right, let me know what you think about all of this in the comments. And if you enjoyed the video, subscribe, hit the like button. Thanks for watching, and I'll catch you in the next one.