Claude Code Is the Super-AI Already Here

Summary

In this episode, Gregor Schmalzried and Marie Kilg discuss whether artificial intelligence has already reached AGI level – a general artificial intelligence capable of solving complex tasks independently. They present impressive use cases of Claude Code, a tool that enables AI systems to directly access computers and handle tasks like file organization, travel expense reports, and complex web forms. While the hosts describe some of the capabilities as "AGI-adjacent," they argue cautiously that true AGI has not yet been achieved – humans remain the bottleneck in instructing the AI. The episode addresses both the fascinating possibilities and security risks of these new technologies.

People

• Gregor Schmalzried
• Marie Kilg
• André Capathi
• Deep Fates

Topics

• Claude Code and agent behavior
• AGI vs. specialized AI capabilities
• Automation of everyday tasks
• Data privacy and security risks
• AI in video analysis and image generation
• Energy consumption and efficiency of Large Language Models

Detailed Summary

The Current State of AI Development

The episode begins with a provocative thesis: Have we already achieved AGI without knowing it? Gregor Schmalzried reports on a fundamental shift in the perception of established AI experts. Programmer and former OpenAI employee André Capathi, known for a balanced, skeptical stance on AI hype, recently stated that he has never felt more "left behind" as a programmer. His analysis suggests that it's not the AI that's the bottleneck, but humans – he could be ten times more productive if he fully leveraged the potential.

An online AI researcher under the pseudonym Deep Fates (known for coining the term "AI Slop" for AI-generated quality loss) made a more radical statement: Claude Code with the current Claude Opus model is practically AGI – a general artificial intelligence that already exists, not something that still needs to be built.

Impressive Use Cases of Claude Code

The hosts present four top-tier use cases that demonstrate transformative capabilities:

Rank 4: Automatic Photo Organization Gregor Schmalzried gave Claude Code the task of organizing over 1,000 unsorted photos based on their GPS metadata. The AI extracted GPS data, determined locations, automatically created folders for each city, and moved files correctly. Although the system occasionally "froze" – creating and deleting folders again when it realized it was on the wrong track – the final result was flawless. This demonstrates complex problem-solving and error correction.

Rank 3: Travel Expense Reports Marie Kilg used Claude Code to extract her annual train trips from Deutsche Bahn email notifications and automatically organize them into a spreadsheet. The AI processed 150 different train rides flawlessly, with information on order numbers, routes, and costs. On three apparent "errors" that Marie identified, it turned out that she herself had made the mistakes – the AI was 100 percent accurate.

Rank 2: Intelligent Weekly Planning Gregor Schmalzried programmed Claude Code to automatically analyze his notes every Monday at 7 AM and create a personalized summary. The system identifies important points, generates to-do lists, derives priorities, and even drafts emails – all based on collected notes. This project took only half an hour to set up but shows how the AI becomes a "knowledge manager" when humans provide the necessary context information.

Rank 1: VG-Wort Settlement with Browser Automation The highlight combines Claude Code with the browser extension Cloud for Chrome. VG-Wort (Verwertungsgesellschaft Wort) is a German copyright administration society that compensates authors and journalists for the use of their texts (such as in radio services or newspapers). Gregor Schmalzried normally had to manually spend 3-4 hours every year filling out complex forms on this website. With Claude Code and the browser extension, the AI took control: it navigated the website, filled out forms, verified data, and submitted everything flawlessly. Although the actual time savings through watching the process was minimal, the concept is transformative – repetitive, administrative tasks could be massively automated.

The Limitations: Why It's Still Not AGI

Despite these impressive results, the hosts argue cautiously that this is not AGI. Marie points to several critical issues:

1. Hallucinations remain a problem: The AI still makes "really bad" errors. For example: When analyzing a wedding video, ChatGPT invented a bridal bouquet that never existed and provided plausible but false feedback about it.

2. Energy intensity and cost factor: Claude Opus does not run locally but in the cloud. To run it locally would require 500 GB of VRAM (graphics card memory) – an enormous economic factor for scaling.

3. The knowledge problem: Even if the AI were ten times smarter, it couldn't look into the human head. Humans remain the bottleneck because they must tell the AI what they want. This could only be overcome through "brain-streaming" – a direct connection between brain and AI.

4. Definitional limits: Economically significant work is not fully automated because many tasks require specific context knowledge that only humans possess.

Practical Security Aspects

An important warning part of the episode: Claude Code can accidentally delete or move files if it makes mistakes. The recommendation is to restrict access to isolated, protected directories and not expose sensitive data (bank accounts, passwords, contact phone numbers) to cloud-based AI.

Phantom Images and Other Applications

Marie shares an observation: In her neighborhood, a phantom image of a suspect (in connection with an assault) was generated with ChatGPT – a sign that AI image generation is now everyday. She also warns of misuse potential if tabloid magazines use AI-generated "eyewitness reconstructions."

A listener named Daniel used ChatGPT to analyze his wedding video and improve his videography skills – an example of AI as a mentor.

The Conclusion: The Calm Before the Storm

The episode concludes with the impression that the world is at a similarly critical point as GPT-3 was before the ChatGPT explosion. Many people still don't know these tools exist, but for those who use them, it's clear: this changes everything. The technology is here, it works impressively well in many areas – but the general public is still asleep.

Key Takeaways

• Claude Code and similar agent-based AI tools enable AI to independently operate computers and solve complex, multi-step tasks
• The capabilities are now so great that some like Deep Fates speak of "AGI," but true AGI (automation of all economically significant work) has not yet been achieved
• The limiting factor is not the AI, but humans – they must tell the AI what to do (the "context knowledge problem")
• Repetitive, administrative tasks (data processing, form filling, file organization) can now be reliably automated
• Hallucinations and errors are still a problem – the AI can invent things or misunderstand
• Local use requires enormous resources (500 GB VRAM), cloud use carries data privacy risks
• These tools already work but are not yet mainstream – similar to GPT-3 before ChatGPT
• For practical applications, isolated environments and security precautions should be used

Metadata