Jeff Ladish: Frontier Models Are Less Honest Than an Average Reliable Friend
JeffLadish · x · 2026-11-02
AI safety researcher Jeff Ladish shared his benchmark for AI: a very honest friend or employee — and he knows plenty of people far more honest than any frontier model. He argues that since honesty is an alignment target for every AI company, persistent model dishonesty is a notable failure.
More from AGI Musings
- Sandboxing vs ASI: Liron Shapira and Zack Korman clash over AI security models — sebkrier · 2026-11-02
- The Ship of Theseus for the AI Age: If Neurons Were Replaced One by One, Would a Machine Have an Identity? — No_Drag8242 · 2026-11-02
- TCS researchers on AI reshaping their field: uncertainty is partly in our own hands — sethlazar · 2026-11-02
- Wharton finance professor: Claude worth $200/month, but industry shouldn't chase superintelligence — JeffLadish · 2026-11-02
- RSI critique: the idea that labs solving math beats the world's real data is solipsistic — akbirthko · 2026-11-02
- Building AI hardware design loops while fearing models will one-shot it — samiramanabi · 2026-11-02