🧠Karpathy: AGI still a decade away

Good morning, AI enthusiast. Andrej Karpathy just threw cold water on “the year of agents” hype, saying that human-replacement AI agents are still at least a decade away.

In today’s AI newsletter:

• Access leaked system prompts from top AI tools

• Andrej Karpathy says AGI is Still Decade Away

• OpenAI Bans MLK Deepfakes After Estate Complaint

• AI tools & more

HOW TO AI

💻 Access Leaked Prompts from Top AI Tools to Master How Billion-Dollar AI Systems Think and Respond

A developer leaked 10,000+ lines of system prompts from top AI tools like Claude and Perplexity, revealing how these billion-dollar AIs think. In this tutorial I’ll shows you how to explore them to learn elite prompt design and create powerful AI workflows.

🧰 Who is This For

• AI enthusiasts learning advanced prompt design

• Prompt engineers building powerful AI workflows

• Content creators automating creative projects

• Students and researchers exploring AI reasoning

STEP 1: Access Repository

Head over to the GitHub Repo.
You’ll see folders for each company, Claude, Perplexity, Cursor, Lovable, and others.

Click any folder to open it. Inside, you’ll find .txt and .json files like Agent Prompt, Chat Prompt, and Tools.json, these contain the actual system-level instructions used by AI tools.

STEP 2: Explore and Analyze the Structure

Start with “Claude Code” or “Cursor Prompts.”
Each file shows you how these companies structure their prompts:

• How they define the AI’s role and purpose

• How they manage memory, tools, and reasoning

• How they organize instructions into modular components

This gives you a real-world view of how multi-agent systems are built internally.

STEP 3: Identify Prompt Patterns

As you read through, notice patterns:

• Context layering (system → agent → user)

• Instruction hierarchies (task definition → reasoning → response format)

• The tone and clarity used in top-tier prompts

Take notes on these techniques, this is what separates amateur prompting from enterprise-grade prompt design.

STEP 4: Test and Share Your App

Now that you understand how professionals structure prompts, it’s time to apply these insights.

Here’s how you can use them in practice:

• Build structured workflows in ChatGPT, Claude, or Cursor

• Create your own AI assistants with modular prompts

• Design prompt systems for startups or internal tools

AI NEWS

🧠 Andrej Karpathy says AGI is Still Decade Away

Andrej Karpathy, in a 2.5-hour discussion with Dwarkesh Patel, argued that true human-replacement AI agents are still about 10 years away, despite hype around “the year of agents.” He emphasized that current AI tools can handle boilerplate tasks but fail at unique, complex work.

→ Agents today lack multimodal skills, continual learning, and reliable computer use
→ LLMs are “ghosts” or “spirits,” trained by imitation of humans rather than evolution
→ Claude Code and Codex CLI work for common tasks but struggle with precise, intellectually intense projects

Why is important: Karpathy’s view contrasts sharply with the current wave of “AI agent” announcements from OpenAI, Anthropic, and Google. His message: coding assistants are real, but human-level autonomous agents remain a decade away.

AI ETHICS

🛑 OpenAI Bans MLK Deepfakes After Estate Complaint

After weeks of chaos on Sora 2, OpenAI announced it will pause generations depicting Martin Luther King Jr. following complaints from his estate. The text-to-video app had become a hotspot for tasteless AI videos mocking deceased public figures, from Stephen Hawking to Elvis Presley.

→ MLK appeared in disrespectful and absurd videos, including altered speeches and rickrolls
→ OpenAI states it is now strengthening guardrails for historical figures
→ Estate requests can now block the use of a public figure’s likeness in Sora videos

Why is important: This new policy shift shows the company is moving toward more responsible use of AI for historical and cultural figures, albeit under pressure from estates and public outcry.

ESSENTIAL BITES

• At TSMC Arizona, Nvidia and TSMC unveil the first Blackwell chip wafer made in the US, representing that Blackwell has reached volume production.

• Google plans to release Gemini 3 in December; it is expected to be more performant across the board, especially in coding and multimodal generation.

• Google says North Korea hackers are using “EtherHiding” to embed malware on blockchains, the first time it has seen a nation-state threat actor using the method.

• Senate Republicans share an AI-generated video falsely depicting Chuck Schumer saying on camera comments he made in a print interview on the government shutdown.

HOT AI TOOLS

🎥 Veo 3.1: Google’s new and improved video AI model

💹 Compound: The world’s first AI finance analyst you can trust

💨 Claude 4.5 Haiku: Anthropic’s new fast and affordable AI model

⚡ Manus 1.5: Build AI web apps with logins & analytics

THAT’S IT FOR TODAY

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas