Transcription
Ready or not, here comes Mattis! China's world-first fully autonomous AI agent. The launch of DeepSeek was a massive moment for the American AI industry, described as a "Sputnik moment" for AI development. The race for AI supremacy is on; while the US leads in many ways, China is producing some amazing technology.
Not long after DeepSeek, we have Mattis. A clip showcasing its capabilities is going viral. Mattis can automate many tasks, including financial transactions and social network service (SNS) analysis. I can't verify the video's legitimacy, but industry insiders who've tested it confirm many features seem accurate. This is a real product, but this is a developing story. We can't fully confirm its features yet, but real people testing it are describing it favorably.
Leon from Mobile Capital (Grok translated this today) describes Mattis: My screen was flooded with Mattis by setting up different agents or workflows constantly called by the large language model (LLM). They execute work tasks; it's a Frankenstein's monster stitched together by one person, but beautifully done. For most, finding a suitable agent in the vast agent workshop is challenging; Mattis solves this. Previously, my workflow used DeepSeek and Dabow Slimi (two large AI models we've covered from China). That workflow could now be replaced by Mattis. Let's try integrating it into a production environment. Currently, for code output in production, it's still CLA+Tray. Today, I'm researching; Mattis is still using agents to assist with mobile tasks like social media analysis or data analysis for wallets, exchanges, and browsers. The Mattis account on Twitter/X has been disabled; the handle is @ManisAIorHQ. It might be linked to crypto scams, a hack, or autonomous bots. We're unsure.
Rowan Cheung from The Rundown AI newsletter obtained access and demonstrated its use cases. Here's Mattis's intro video (about 4 minutes).
Hi, I'm P from Manis AI. For the past year, we've quietly built what we believe is the next evolution in AI. Today, we're launching an early preview of Manis, the first general AI agent. This isn't just another chatbot workflow; it's a truly autonomous agent bridging conception and execution. While other AI stops at generating ideas, Manis delivers results. We see it as the next paradigm of human-machine collaboration and a potential glimpse into Artificial General Intelligence (AGI). Let me show you Manis in action across three tasks.
First, screening résumés. I gave Manis a zip file of 10 resumes. Each Manis session has its own computer; it works like a human, zipping, browsing, and recording information. Manis works asynchronously in the cloud; you can close your laptop, and it will notify you when finished. You can give it new instructions anytime. I asked Manis to find more résumés. After reading 15, it provided rankings, candidate profiles, and evaluation criteria. I preferred a spreadsheet; Manis created one. Manis has its own knowledge and memory; it learns for future similar tasks.
Next, researching New York properties based on multiple criteria. For complex tasks, Manis breaks them down and creates a to-do list. It searched and read articles about safe neighborhoods, researched middle schools, wrote a Python program to calculate my budget, filtered listings on real estate websites, and wrote a detailed report, compiling all resources.
Finally, a correlation analysis between stocks. Manis accesses data sources through APIs, validates data, writes code for data analysis and visualization. Coding isn't the goal, but a tool for problem-solving. It completed the analysis and visualization. Interactive visualization is cooler; I asked Manis to create a website. With permission, Manis deployed it online and provided a sharable link.
What you've seen is a small sample. On a benchmark for general AI assistance, an early checkpoint of Manis achieved state-of-the-art performance. It's solving real-world problems on platforms like Upwork and Fiverr and has proven its capabilities in Kaggle competitions. This wouldn't be possible without the open-source community; we're committed to giving back. Manis operates as a multi-agent system powered by several distinct models. Later this year, we'll open-source some, specifically the postering for Manis, inviting everyone to explore this agentic future together. The name Manis comes from "mens et manus" (mind and hand). It embodies the belief that knowledge must be applied. Manis AI extends your capabilities, amplifies your impact, and brings your mind's vision into reality.
Mr. P, founder and CTO of Manis AI, confirms the X account suspension was due to miscommunication about it being linked to another account involved in crypto scams. It might return online.
They use an Apple computer, but Manis uses its own computer in the cloud, running Ubuntu (a Linux distribution). I recently installed Linux (Ubuntu) on an old computer with CloudCoder, using the command line. I was impressed. I think more people will build on this foundation. The large language model operates on the Linux system via the command line.
The open-source ecosystem for AI is expanding faster than expected. DeepSeek and Mattis (untested by me) seem to have all the parts; we'll see how well it works. Building on Linux gives users significant power; it's an open-source operating system. They're planning to open-source some fine-tuned models later this year.
If Mattis is as good as it seems (less than 24 hours since it became known in the US), we might have open-source, fully autonomous AI agents before accessing large proprietary models or their AI agents. This could be massive.
Rowan Cheung posted about Mattis in his newsletter; Mr. P sent him an invitation code. Rowan asked Mattis to create his biography and deploy a website. It was insanely impressive; it browsed his social channels, articles, and deployed an accurate, up-to-date website. The video on X shows two windows: the left shows what Mattis is doing (similar to a chat interface), and the right shows the action (running on Linux). It receives instructions, creates a project folder, and begins working. It creates a to-do file with subtasks (research, biography writing, website development). It imports tools (APIs), does research, saves it in JSON format, and creates a biography file. It creates a directory for the biography, writes the biography, and creates Rowan's portfolio. It installs dependencies, a Git repository, and downloads software. It troubleshoots issues. This is what an autonomous AI agent should be. It creates the webpage in HTML, designs and codes it, runs `npm run dev` (a development server), and shows the page locally. It fixes build errors. It completes most tasks and subtasks, and deploys the website. It handles issues like needing permission to deploy publicly. It then posts the live website.
Next, Rowan asked Mattis to find a room to rent in San Francisco for 6 months, with low crime. It creates a project folder, breaks down subtasks, analyzes crime statistics, matches them with criteria, and writes a full report.
Another request: build a full AI development course. It builds a complete outline and modules. It beats many open-AI Deep Research results on the Gaia Benchmark. It seems that Manis was hit hard by massive demand; they lacked sufficient servers for a basic demonstration. It went viral in China, and now the US is catching on. This all happened in the last 24 hours.
I'll post footage of CloudCoder (from Anthropic) if I can. It operates similarly, running commands on Linux, installing packages, creating files, and creating a development server to view applications in real time.
The implications are significant: job displacement is obvious. But in software design and sales, Mattis could build a website and host it online. I recently needed a simple application and considered using Mattis instead of searching online. In 20-30 seconds, I had the code and application. For simple software, it's faster to have a chatbot create it than search online. As AI improves, software-as-a-service might be impacted; people may create custom software with only the functions they need. You might not even need a user interface; you just input a file and get the output. Adobe makes $5.6 billion; few use every feature. Most use 20% of tools. With better AI assistants, you might speak in natural language and have custom software created. This won't happen overnight, but it will impact software-as-a-service companies. It will also impact Google and research-based information sites. It seems better than Deep Research (which I use daily). This is a phenomenal step forward.
Manis's website (not great in dark mode) shows its capabilities: stock analysis, insurance policy comparison, trip planning, B2B supplier sourcing, analysis of Y Combinator companies, online store operation analysis, and a professional text prompter.
Something like this is coming. They'll open-source it (hopefully this year). It'll run on open-source operating systems. It's mind-blowing. OpenAI announced AI agents at $20,000/month; how will that compete with open-source options? Maybe custom solutions will remain for large enterprises.
To get ahead of the curve, install Linux on an old laptop (not Apple or Chromebook). Linux is free and open-source. 5-10 years ago, it was harder to use; now it's easier to install, and chatbots like ChatGPT can guide you. There are various distributions; I used Ubuntu (like Manis). Download the ISO file (6-7GB) and use a USB drive. Rufus (free, open-source) formats the USB drive. ChatGPT can guide you through the process. It creates a bootable drive; insert it, restart your computer, and boot from the USB drive. ChatGPT helps with the rest. It takes about 10 minutes. The desktop is similar to others. You'll mostly use the command line. ChatGPT provides commands. You can install it alongside Windows. You can install CloudCoder (Anthropic); it’s simple and lets you run commands like a chatbot. I have a video coming on things to try.
What do you think about Manis? Let me know in the comments. Thanks for watching; I'm West, and I'll see you next time.