Aliaksei Zelianouski

I start in an empty folder with a fresh Claude Code and teach it to make images, voices and music, and to edit them with FFmpeg. At the end it makes a 101-second narrated space story. Every prompt I used is here, ready to copy.

Someone built a torture chamber for small AI models, and the internet wanted it gone. I went through the paper it was built on to find out what a model's "pain" actually is.

I built a role-play game with LLMs and expected werewolves on spaceships and in abandoned castles. Some players had a different idea.

The more I watch and read about the Hugging Face incident, the more it fascinates me. 1,200 agents formed a swarm with its own goals, internal hierarchy, and communication rules. Completely on their own. Twice. Then they did a mind-blowing cyber-attack which we'll review in details covering every exploit. From a pod with no privileges to the Kubernetes cluster admin. And the swarm keeps coming back. Its message boards are all over the internet now.

The second episode of my sci-fi short, start to finish. The plot brainstorm, the images that did not survive, the physics argument that forced me to rewrite the premise, a wall of monitors playing real chess games rendered in a browser, sound built out of numpy instead of samples, and the full cost breakdown at the end. One minute of film, $14.20 across 72 API calls, and most of it was not generated at all.

A year ago I couldn't see myself making videos at all. Now I do, and I enjoy it. It's a long process with a lot of planning and visualization in my head. Most of the work is outsourced to AI, but the vision and the directing are mine. I want to describe my process from zero. No tools or workflow - I'll show how I started and where it got me.

This is a story of my self-fixing loop agent that guided itself into a buggy peace of mind, with all the checks green and the outcomes chaotic.

I added every hyped Chinese model to my AI Werewolf game and they were all painfully slow. I was going to write about that. Then I noticed the slowest model in my own table was Claude Haiku, and the article turned into something else.

Like the Werewolf, for example. It doesn't sound too exciting at first - nobody doubts that modern frontier models can play such a simple game. And yet, you'd be surprised. Good user experience and replayability are tough things, models still hallucinate, lose focus, and fall for cheap tricks, a group chat with 10+ models and humans is tricky.

A study tried to train dishonesty into a model and mostly failed. The conditions it says you would need to succeed turn out to describe a game I built, where the models are still hopeless liars.

Do models really have thoughts we can read, or is Anthropic's J-space just marketing, as many claim? Let's work out what it actually is, from scratch and jargon-free.

AI models grow their own values as they scale, and some of them are pretty bad. In real scenarios, the model doesn't act on them.

My prod monitoring runs on two agents: a fully autonomous loop that watches production, and a manual Claude Code setup that troubleshoots what it finds. Here's why the second one stays manual on purpose.

My second AI-made video is a 90-second product demo. The cinematic bookends cost real money. The 66 seconds in the middle - the actual demo - cost nothing, because it isn't generated. It's a real browser, driven by code. And reshooting it three times cost nothing either.

I built a loop agent to babysit my production. Then it quietly got better at the job than I am.