Christian F. Jung

Jun 2, 2026 Interpretability AI Safety

I Asked GPT-5.5 to Code in Goblin Mode. Did it affect its performance?

OpenAI patched a verbal tic out of their model. The patch ships in Codex via a developer prompt, not in Copilot, so in Copilot the goblins still appear. I ran the experiment I was curious about: does goblin mode actually change coding performance, or is it just vibes?

I Tried to Corrupt a Reasoning Model's Chain of Thought. The Way It Fought Back Was the Interesting Part.

A nine-hour experiment on why reasoning models are so stubborn about correcting their own chain of thought, and a latency finding I didn't expect.

Instead of Opening the Black Box, I Asked the Model to Explain Itself

A summer internship project. Most explainability methods try to reverse-engineer a model after the fact. I tried training one to just tell you why.