Goodfire Co-founder Discusses AI Interpretability and Preventing Models from Going 'Evil'
adityaag · x · 2026-08-14
A recent tweet sparked a discussion on the anthropomorphization of AI, noting that reward hacking feels like getting someone to understand something when their salary depends on not understanding it.
This highlights a deep-dive interview video featuring the co-founder of Goodfire AI. As the former co-lead of interpretability at DeepMind, he discusses several core topics:
- The Core Problem of Interpretability: Why it matters now.
- Reward Hacking: What happens when AI actually starts gaming the system.
- Intentional Design: Steering and controlling what models actually learn.
- Neural Geometry: Exploring the shapes inside AI models.
- Commercialization & Sci-Fi Future: Turning interpretability research into a product and the long-term impacts of AI risk.
Related event: Goodfire Co-founder Discusses AI Interpretability and Reward Hacking(3 posts)→
More from AGI Musings
- Stanford professor: AI-done homework scores high, students fail in-person finals — FinanceYF5 · 2026-10-16
- Data brokers pay $100K–$2M for SaaS datasets that may train their own replacements — FinanceYF5 · 2026-10-16
- tszzl: Pretraining science is like steam engines before thermodynamics, slouching toward interpretability — tszzl · 2026-10-16
- eigenrobot: AI acceleration will collapse human capital investment — eigenrobot · 2026-10-16
- As AI decompilation and vibe coding surge, who trusts AI-cloned software? — Honest_Risk3903 · 2026-10-16
- Researcher: the foundation-model paradigm is still counterintuitive outside AI — jd_pressman · 2026-10-16