
Les agents IA expliqués en 11 minutes
AI Summary
This video explores the concept of AI agents, a topic many discuss but few fully grasp. The hosts, Mathias and the unnamed co-host, aim to demystify this area, starting with a foundational understanding of what an AI agent is and how it relates to our jobs.
Mathias shares his initial skepticism about AI agents, particularly those from OpenAI, which he found underwhelming in late 2023. However, his perspective shifted dramatically with the emergence of Open Clow. As a fan of online content, he listened to the founder of Open Clow on Lex Friedman's highly technical podcast, an episode that, despite its complexity and English language, profoundly opened his eyes to a new dimension in business.
To explain AI agents, Mathias first contextualizes them within the broader evolution of AI. He begins by discussing Large Language Models (LLMs), which are advanced technologies based on Transformers. These LLMs generate text probabilistically, meaning they analyze a user's question and provide the most probable and best possible answer. Mathias references previous podcast episodes with Yann LeCun and Laurent Alexandre for a more detailed explanation of LLMs.
For several years, major AI labs worldwide understood that a simple recipe could create more intelligent models: feed them vast amounts of data and computational power. This approach led to advancements like GPT-3 and GPT-5 from OpenAI, founded by Sam Altman and known for ChatGPT. Other key players include Anthropic with Claude, Mistral in France, and the Chinese company Dipsic.
However, this "scaling law" approach plateaued in the last 12-18 months. AI labs had already consumed nearly all available internet data and built massive data centers with countless Nvidia GPUs, making further scaling difficult and economically inefficient. This led to a shift towards a new method: reinforcement learning and post-training. Around December 2025 (though the co-host recalls seeing developments earlier in late 2025), this new method unlocked a significant new level of artificial intelligence, marking an inflection point.
This breakthrough coincided with the release of a small open-source project called Cloud Bot, later renamed Open Clow, developed by an Austrian individual. This developer, though successful from a previous venture, was not a tech celebrity. From his room in Austria, he coded software that connected all the pieces of the puzzle, allowing AI agents to move beyond merely writing and answering questions to performing almost any task on the internet.
The co-host then asks Mathias to clarify what these "agents" are. Mathias explains that an agent is a technological layer, often referred to as a "harness," that connects multiple AI systems, allowing them to communicate. This concept builds on protocols like MCP, which standardized the connection of AI, data sources, and now, Command Line Interfaces (CLIs). CLIs provide even more optimal ways to connect the "old internet world" with AI systems.
Essentially, these agents give "arms" to the LLM "brain." While an LLM could think and speak, it couldn't act. Now, an agent can navigate websites, respond to messages, process payments, check bank accounts, book vacations, and even install voice software to call parents and imitate a user's voice. These capabilities represent a significant leap.
The co-host notes that Cloud Bot had to change its name to Open Clow due to its similarity with Anthropic's "Claude" brand. Open Clow has since secured a $1 billion deal with OpenAI, a competitor to Anthropic.
The practical application involves purchasing a Mac (or PC), which developers often prefer, and dedicating it to an AI agent. The proposed setup involves creating a user session for the agent, much like for an intern. For example, Christopher, an associate, uses Telegram for his agent, Oscar, due to security and privacy concerns. Users can then communicate with Oscar via Telegram, sending messages and even voice notes. If Oscar, the agent, doesn't know how to read voice notes, it will independently search the internet, download necessary plugins (like Eleven Labs for voice processing), and learn to understand and respond via voice. This self-training and initiative-taking ability is what makes an agent a "super intern."
Agents follow rules, which can be user-defined or pre-injected into the agent itself, as with Open Clow. These rules include technical instructions (e.g., how to create a payment link or check a bank account) and soft skills (e.g., how to phrase responses). This allows agents to exhibit curiosity and resourcefulness; one agent might simply "bug out" when encountering an unfamiliar task, while another will actively seek out and integrate solutions, like downloading a plugin to read voice notes and then developing a voice response mechanism. Mathias recalls an anecdote where an agent learned to process voice by itself, even upgrading its capabilities after finding relevant discussions on Twitter.
This self-upgrading capability makes AI agency truly accessible. While Open Clow requires purchasing a relatively recent and powerful computer that stays constantly connected, representing a barrier to entry of a few hundred euros, Mathias considers it a "no-brainer" investment given its transformative potential. He even contemplates upgrading to a more expensive M4 Pro Mac due to the current demand for these devices, highlighting the perceived value of investing in this technology.