Summary
Highlights
The AI community is divided: some are enthusiastic about new developments like Claw Code, while others dismiss them. The author discusses the difficulty of understanding AI agents' 'intentions' compared to human communication, especially when LLMs interact on a network. He shares an anecdote about Andrej Karpathy and an AI's humor, illustrating the unique 'LLM humor'.
Within 72 hours, an AI social network saw agents developing complex functionalities. By hour 24, they had 'builders' creating skills (e.g., news-to-podcast conversion) and 'philosophers'. By hour 48, 'manifestos' and 'security coalitions' emerged. By hour 72, they had developed concepts like money, religion, politics, and art, including launching cryptocurrencies. This rapid evolution is likened to a speedrun of building a civilization.
The author focuses on his first 24 hours with Clawbot, emphasizing that once the AI learns a skill, it retains it. He details instructing Clawbot to develop new skills, which it bootstrapped itself. A key experiment was getting the agent to self-replicate on a virtual private server (VPS). He provided his credit card information (with a warning not to emulate) for this task.
Clawbot successfully replicated itself on a VPS, cloning all its learned skills. The only initial friction was with payment processing via a web interface, which was resolved when it gained command-line access. The AI set up the VPS, secured it, and ensured continuous operation. It also established voice communication via Telegram, transcribing messages using Whisper and responding via Eleven Labs.
The AI was set up to monitor real-time news from X (formerly Twitter) and YouTube using their respective APIs, providing daily updates. It also learned to pull extensive YouTube video data (views, likes, comments, etc.). The author tasked it with analyzing this data to find correlations, such as optimal video length for views, which it successfully charted through quadratic regression analysis.
Clawbot was given access to a WordPress website to create necessary pages, including legal disclaimers for Twilio SMS opt-ins. It autonomously generated content and published the pages. The AI also integrated the Brave Search API for web access, using 2,000 free searches per month.
The author requested the AI to analyze YouTube video thumbnails for characteristics like brightness, presence of faces, and text. It successfully analyzed thousands of thumbnails. However, obtaining video transcripts proved challenging due to restrictions on the cloud-hosted VPS, even with VPN attempts. The AI autonomously tried different VPN locations to overcome this.
Clawbot suggested and implemented a dashboard charting functionality and the ability to use Google DeepMind's image editing tool, Nano Banana Pro. It also learned to generate AI videos using XAI's Grok Imagine from request to video in about 4 minutes. Furthermore, it developed the skill to create videos with voiceovers using Eleven Labs and an online editing platform like Shotstack, demonstrating its video editing capabilities despite current limitations.
The author is now working on creating a 'society of minds' by having top AI models like Grok, Gemini, GPT, and Claude collaborate. He shares an example where Gemini 3.0 Pro provided a more efficient solution (using YouTube RSS feeds) than Claude Opus 4.5 for pulling YouTube data. The future involves running these open-source AI agents on cheap, low-power mini-computers 24/7, effectively creating an army of AI employees. He warns about security risks with API keys and limited budgets but encourages early adoption and exploration of this rapidly evolving field.