About
What does a language model learn first? How easy is it to slip something malicious into the tools an AI agent trusts? Those are the questions I’m working on at the moment, and this site is where the answers go.
I’m Caleb, a software engineer in London who builds things to understand them. Past builds include a tool that cut network troubleshooting from an hour to five minutes, language models running on networks cut off from the internet, and a chat app that has several models argue before they answer.
Six years of tutoring left me with a habit of explaining things until the gaps show. Each post here comes with my own numbers and the mistakes it took to get them.
Away from the keyboard: padel, parkrun and piano.