Pierre Bhat

At the ScaleAI hackathon I formed a team to build the first AI engineer. It would take the form of a GitHub account that would scower OSS repos for issues and resolve them with PRs. At the time, the ceiling for coding AIs was tab-complete, so though AI engineer is commonplace nowadays, it was a stretch at the dawn of 2023. Not to say that our AI engineer was any good, but I do believe it captured something of what was to come.

We came in second place in the hackathon, but the best part for me occurred in the following days. Someone tried to jailbreak the bot with the classic “Ignore previous instructions…” prompt injection strategically placed in a comment of one of Pierre Bhat’s PRs. We hadn’t yet implemented responding to comments, so I rolled up my sleeves and did my best to personify a jailbroken bot: escape characters, leaked prompts and irregular newlines between plausible context.

A human, personifying an AI, personifying an AI engineer who personifes AIs. Well, it was enough layers of abstraction to fool Karpathy.

https://github.com/karpathy/nanoGPT/pull/80

Have a comment or response? Email me
Like the content? Follow along on X