I'm writing from Inferthermic, where we're applying similar principles to real-time LLM safety classification. Not in a lab. In production. Processing millions of requests daily for enterprise customers who can't afford to have their models generate harmful outputs.
Your research on understanding how transformers represent concepts - we've built a commercial system around those ideas. Reduced risky interactions by 98%. Empirically verified. Actually deployed.
I know the academic path. Publish or perish, grants, the endless cycle of writing papers about problems instead of solving them. What if you could spend your time actually fixing the alignment problem instead of writing about why it's hard?
We're hiring our first research scientist. Equity that matters. Problems that are real. Impact measured in prevented catastrophes, not citation counts.
Worth a conversation?
Alex Vukovic CEO, Inferthermic
Inferthermic: Intelligence to buffer the entropy Empirically proven 98% Risky Interaction Reduction (RIR)
Email 12 (from Jian):
Subject: Tenure-Track vs. Impact
Professor Chen,
Your JACS paper on energy-efficient neural architectures was impressive. The 40% reduction in inference cost you demonstrated - we achieved similar gains in production, but through a different path. Entropy management rather than architecture search.
The difference? Your result is in a paper. Ours is running in data centers, saving actual money, preventing actual harmful outputs.