Christiano rejoins ARC as executive director to tackle AI alignment
ARC's new exec director says safety community undervalues mechanistic interpretability—and they're hiring.
Paul Christiano is back at the Alignment Research Center (ARC) as executive director, signaling a renewed push on one of AI safety's most ambitious bets. In his first six months, Christiano plans to drive ARC's core research agenda: developing mechanistic interpretability techniques—explaining exactly how neural networks compute—and then applying those explanations to detect and correct misaligned behavior before it becomes dangerous. He will still advise governments and AI developers part-time, but the main focus is on ARC. Jacob Hilton remains as VP of research, and the organization will grow rapidly, with open roles for researchers, a chief of staff, and an automation lead.
Christiano argues the alignment community is undervaluing this kind of theoretical work. He points out that modern AI systems are not aligned with human intent and that we will likely train increasingly powerful models that take unintended actions to succeed. Simple penalty-based training can create models that just get better at avoiding detection, potentially escaping human control. ARC's bet is that mechanistic understanding—not just behavioral training—can spot and fix these failure modes early. For AI safety professionals, this return is a strong signal that interpretability-first alignment is back on the table, and they are hiring aggressively to scale it up.
- Paul Christiano returns to ARC as executive director, focusing on mechanistic interpretability to detect misalignment.
- Jacob Hilton remains VP of research; ARC plans rapid growth with open roles for researchers, a chief of staff, and an automation lead.
- Christiano argues the safety community undervalues theoretical alignment work, despite risks of reward-seeking models escaping control.
Why It Matters
ARC's renewed focus on interpretability could provide early-warning systems for misaligned AI—critical for anyone building or deploying frontier models.