- 17 Nov 2026Why two open chat models refuse less in HindiAsked the same harmful requests in English and in Hindi, two open chat models (Qwen3-8B and Llama-3.1-8B-Instruct) refuse far less often in Hindi. What they do instead, what changes inside them, and how far steering can bring the refusals back.
- 16 Mar 2026Applying autoresearch to backdoor discoveryAutoresearch hands the experiment loop to an AI agent: edit one file, run a fixed test, read one number, keep or revert. What it is, what carries over to new problems, and where it breaks, with a backdoor hunt on Jane Street's Dormant LLM puzzle as the case study.