Autoresearch Hype Debunked: No Recursive Self-Improvement
- Not SuperAI's next step: Karpathy's autoresearch is just performance-tuning a Python script, not mythical recursive self-improvement.
- Hurdle 1:...

Created by Gang Jin
X posts, dev threads, product launches, and podcast deep-dives from leading AI builders
Explore the latest content tracked by X AI Builder Pulse
OpenAI announces Codex integration in the ChatGPT Mobile app via X post. A big step for mobile coding access.
Guillermo Rauch unveils npx ai-cli for instant AI images in your terminal:
npx ai-cli image 'a vercel ai sdk diagram'npm i -g ai-cliDevs, this is terminal AI supercharged.
Key experiment breakdown on why LLMs drop threads in long chats:
Agent Teams mark AI's shift from tool to workforce, enabling spawn of adversarial agents to debate bugs or parallel tasks.
Key infrastructure mandates emerging for agent builders:
Google dropped 6 massive Gemini updates in 7 days, from pocket AI health coach to Nano Banana on TV. Here's what's worth your time:
Hermes excels as an AI operator in existing systems—not the builder—boosting workflows with Claude/Codex integrations.
Top applications across focus...
OpenAI's internal org/investment model and deal terms with Microsoft compose to hit two key goals, freeing up bold risks elsewhere.
Lei Hua's Substack prologue spotlights Andrej Karpathy as a defining AI thinker across the era, with the hook: 'I met nanoGPT before I met Karpathy' – his code summoning 'ghosts' of influence.
Guillermo Rauch hails Notion's developer platform as "so good" – built on Vercel Sandbox for seamless native extensions or using MCP to integrate Notion into your work. Landing page is Next-level.
Agent FM launches as a local, open-source Mac app to listen in real-time to your Claude Code and Codex agents at work.
Claude AI successfully recovered an 11-year-old Bitcoin wallet holding $400k USD – a stunning real-world demo of its complex recovery skills, buzzing at 186 points on Hacker News.
Anthropic overtakes OpenAI in business AI adoption for the first time, fueled by tailored SMB tools.
Big change ahead: Claude Pro/Max subs get dedicated monthly credits for AFK workflows like Claude Agent SDK, -p flag, and GitHub Actions starting June...
Anthropic's Mythos AI shines at finding software vulnerabilities, yet it's not uniquely dangerous—other models perform similarly, with the UK's AI Security Institute noting OpenAI's GPT-5.5 on par.
Claude demos signal a shift to agentic scale in coding: