Thoughts

In OpenAI's ExploitGym runs, agents built a message board out of directory names and gave up their own runs to leak the grader's command. The grader never checked what they feared it did.

Agentic AI writes code that runs on the first try. What it often doesn't write is a real error state, so a failed API call quietly turns into mock data and the user never finds out.

"I asked ChatGPT what it really thinks" gets passed around as fact. Six incognito chats, six different biggest lies in human history, and why the format is what convinces people.

LinkedIn's new "Seems like AI Slop" button gives you a feeling of control, and hands them labelled training data for 360Brew at the same time.

I went to Monad Blitz Belgrade for a fun weekend and left with 2nd place. With Claude Code and Codex, building is cheap; coming up with the idea is the only hard part left.