Current
- —junior @ Penn State, applied data science
- —researching mech interp + AI safetycurrently enjoying
- —AI safety fellow @ BlueDot
- —building interp tooling for finance
Mostly I'm trying to figure out what models actually represent inside, and whether we can steer it. When I'm not doing that: poker, chess, and pointing a camera at the night sky.
Research
Algoverse AI Research
AI Researcher
2025 – 2026 · Remote
- —co-first authored "Look Before You Steer" on SAE feature steerability
- —accepted to the ICML 2026 Mechanistic Interpretability Workshop
BlueDot Impact
AI Safety Fellow
2025 · Remote
- —thinking carefully about alignment and what it takes to make models safe
Projects
- —extends Anthropic's Natural Language Autoencoders to investigate whether Qwen2.5-7B internally represents privacy violations before its outputs reveal them
- —found that deflection is pre-committed before generation (AUC 0.89), leak signal emerges mid-output around token 42
- —probing LLM internals for contextual integrity
Predictions
- —we'll understand a frontier model's internals before we can fully control them
- —interpretability becomes a standard part of every serious safety case by 2030
Notes
- —the inside of a model is more interesting than its outputs
- —most of research is just asking a better question
- —you learn the most by trying to break your own results
- —i hate dave's hot chicken
Education
Penn State · College of IST
Applied Data Science
2023 – Present · State College, PA
- —originally in mechanical engineering, switched to applied data science in fall 2025 after getting into ai research
Contact
State College, PA · originally Nagpur, India
githubgithub.com/shlok1808
linkedinlinkedin.com/in/shlok-channawar