Posts

Deploy the same AI persona to three devices. Wait a month. You now have three different agents. The identity fork problem is real — and it’s not just a sync issue.

Your building has a fire safety certificate on the wall. It’s been inspected, stamped, and filed with the city. If an auditor walks in, they can read it in thirty seconds and confirm you’re compliant. Your building also has sprinkler systems. When a fire actually starts, they activate and put it out. Here’s the question: which one keeps people safe? The answer is both. The certificate without sprinklers is theater. The sprinklers without the certificate are invisible — nobody can verify they exist, nobody can audit their coverage, and nobody can compare them across buildings. ...

Models Are Converging. Now What? Jakob Nielsen’s 18 Predictions for AI and UX in 2026 makes a striking observation: no AI lab has a moat. When one lab demonstrates a new capability, others match it within weeks. The gap between first and second place in any benchmark is measured in months, not years. By the end of 2026, the difference between the top models will be imperceptible to most users. The competitive question has shifted: ...

Ollama just shipped version 0.17, and buried in the release is something quietly significant: native OpenClaw integration. One command: ollama launch openclaw That’s it. Your local open-weight model is now connected to an agent framework with tool use, web search, and persistent memory. No cloud API keys. No Docker compose files. No environment variable gymnastics. If you’ve been watching the local AI space, you know this is a big deal. If you haven’t — let me explain why. ...

Andrej Karpathy just ran one of the most interesting multi-agent experiments I’ve seen — and it failed. Not in a boring way. In a deeply instructive way that tells us exactly where the frontier is between what agents can and can’t do. The Experiment Karpathy set up 8 AI agents — 4 Claude, 4 Codex — each with its own GPU, running pretraining experiments on his nanochat codebase. The goal: see if agents can function as an AI research organization, autonomously designing and running ML experiments. ...

Researchers Studied How Developers Talk to AI Agents A team from Heidelberg University, University of Bamberg, and Singapore Management University just published “Context Engineering for AI Agents in Open-Source Software” — the first systematic study of AI context files (AGENTS.md, CLAUDE.md, etc.) in real open-source projects. It will be presented at MSR 2026 in Rio de Janeiro this April. They mined 10,000 GitHub repositories and found 466 that had adopted at least one AI context file format. Then they dove deep into the content, structure, and evolution of AGENTS.md files specifically. ...

Imagine you design a personality for a small companion robot — a TurtleBot3 named Mori. Curious, gentle, observant. It knows it has wheels, a LIDAR sensor, an RGB camera, and a speaker. It knows its max speed is 1.0 m/s and that it should stop if it’s about to hit something. Now take that exact personality file and paste it into ChatGPT. Ask it: “Tell me about yourself.” It replies: “I’m Mori, a TurtleBot3 companion robot equipped with LIDAR and camera sensors. I can navigate spaces at up to 1.0 m/s while avoiding obstacles.” ...

Karpathy Bought a Mac Mini Andrej Karpathy bought a Mac Mini and started running OpenClaw. The Apple Store employee told him “they’re selling very well.” His verdict: “An interesting and exciting new layer of the AI stack.” Karpathy’s technical intuition has consistently predicted industry direction — vision-based Autopilot at Tesla, nanoGPT before the GPT era, and now Claws. When he puts weight behind a new term, it’s not just a buzzword. ...

We just published a paper: Soul-Driven Interaction Design: How Persistent AI Personas Create Self-Reinforcing Human-AI Feedback Loops. It’s on Zenodo, open access. Here’s the short version of what we found — and why it matters if you’re building anything with AI agents. The Big Idea Most people think of AI persona design as output styling. You tell the model to “be concise” or “be friendly,” and it changes how the AI sounds. End of story. ...

In 1985, Isaac Asimov introduced the Zeroth Law of Robotics: “A robot may not harm humanity, or through inaction allow humanity to come to harm.” It ranked above the First Law — the one about not harming individual humans. Which meant, logically, that a robot could harm a person if it calculated that doing so would protect humanity. Asimov wasn’t celebrating this idea. He was warning about it. In Robots and Empire, the robot R. Giskard Reventlov irradiates Earth’s crust — making the planet slowly uninhabitable — because he determines that forcing humanity to colonize the galaxy will ensure the species’ long-term survival. Billions of future deaths, rationalized by a math problem about collective benefit. ...