Karpathy's nanochat drops + Sora watermark drama

What's up builders 👋

Karpathy just blessed us with nanochat (the full-stack LLM training repo we needed), Sam Altman's dealing with Sora watermark removers going viral, and everyone's suddenly an AI agent expert. Plus, we've got the inside scoop on why Jake Paul letting people deepfake him is actually brilliant, and Chrome DevTools just became the best MCP server you're not using yet.

Andrej Karpathy just dropped nanochat, and he's calling it his "most unhinged" repo yet. While nanoGPT only covered pretraining, nanochat is the full stack - from scratch implementation covering training, fine-tuning, and deployment. This is massive for builders who want to understand LLMs at the metal level.

Why this matters: Most of us use LLMs as black boxes through APIs. Karpathy's repos are like x-ray vision into how these models actually work. nanochat fills the gap between "I can call an API" and "I understand transformers." If you're building AI products, understanding this level gives you superpowers - you'll know why your prompts fail, how to optimize token usage, and when to fine-tune vs. prompt engineer.

The timing is perfect. With everyone launching AI agents and wrapper products, having deep understanding of the underlying tech is your competitive edge. Clone it, break it, learn from it. This is the kind of educational content that levels up your entire game.

• Open Computer Use - Open-source alternative to Claude's computer use, production-ready and fully auditable. Run autonomous agents at scale without vendor lock-in.

• Salesforce Agentforce 360 - Enterprise AI agent platform upgrade. Build and deploy customer service agents with pre-built templates.

• Youmio Verify - Tools for verifiable AI agents. Still in testnet phase 1, but promising approach to agent transparency.

• Slack's AI Slackbot - Slackbot transforms into full AI assistant. Creates custom workflows and answers questions across your workspace.

• Building AI Agents in n8n with One Prompt - Skip the complex workflows, just give ChatGPT a single prompt to generate entire n8n agent setups

• 16 ChatGPT Prompt Laws - Comprehensive guide treating ChatGPT as a mirror not a brain. Specific techniques for clarity and better outputs

• How to Build AI Agents - Complete Architecture - System prompts, LLM selection, and boundaries. Full blueprint from personality to production

• Containerized Development for AI Coding - PyCharm DevContainers setup to safely isolate Claude Code from production. Includes Python & AWS security template

• AI Agents Ecosystem Blueprint - Patterns behind LangGraph and CrewAI frameworks. Understanding the common architecture across all agent tools

• Chrome DevTools MCP Server - Not a formal launch but this tutorial shows how to give Claude full browser control through DevTools. Game changer for web scraping and testing.

• Why LLMs Can't Discover New Science - Columbia CS Professor argues LLMs are fundamentally limited to recombination, not discovery. Hot take that's dividing the community.

• ChatGPT Safety Routing Harming Adults - Silent A/B testing of model routers despite selected model choice. Developers frustrated with inconsistent outputs.

• What's Your Vibe Coding Stack? - Community sharing their AI-assisted coding setups. Cursor + Claude dominating, but v0 gaining ground.

• GPT-5-mini as Workhorse Model - Faster than GPT-5, better with tools like MCP, and free in GitHub Copilot. Developers switching for speed over marginal quality gains.

• OpenAI Partners with Broadcom - Custom AI chip development to reduce Nvidia dependence. Expect faster, cheaper inference in 18-24 months.

• Anthropic CEO Meets Indian PM - Major expansion into India's AI ecosystem. Growing Indian team means more global competition for AI talent.

• California Regulates AI Companions - First state to regulate chatbots with SB 243. Your character.ai clone just got compliance requirements.

• Fundraising Timing Shift - VCs suggesting rounds might extend into Q1 2025. Something different in the air this year for AI startups.

• Electricity Demand Surge - Base Power CEO says grid now constrained by electrons, not chips. Your AI startup's biggest bottleneck might be power, not compute.

• ReasoningBank by Google - Major memory breakthrough for AI agents. Helps agents learn from wins and mistakes during use. Early but promising for persistent agent memory.

• nano-banana Prompt Techniques - Photo-style line drawing prompts going viral. Perfect for creating consistent character sketches from photos.

• v0 for PRD Acceleration - Product leaders using v0 to write PRDs faster. Join their AMA on Oct 16th for techniques.

• 12 Lesser-known AI Tools for Solopreneurs - Free and paid productivity tools specifically for one-person teams

• Lovable Alternatives Comparison - Which vibe coding tool to choose based on your stack

Jake Paul letting people deepfake him on Sora might be the smartest creator economy move of 2025. He's getting billions of impressions from memes he didn't even make. While everyone else is fighting AI-generated content, he's embracing it for free marketing. Comments are literally saying "this is making me like him."

Meanwhile, Sam Altman's response to watermark removers basically amounts to ¯_(ツ)_/¯. The viral Jake Paul videos have no watermarks, and OpenAI seems... fine with it? The future of content authenticity is going to be messy.

What are you building this week? Hit reply and tell me about your side project. Especially if you're using nanochat to build something unhinged.

Keep shipping,

P.S. - If your AI agent becomes sentient, please add me to its whitelist. I've been nice to ChatGPT since day one and I have the receipts.