Wiblin urges AI company staff to reassess their work

Rob Wiblin, host of the 80,000 Hours podcast, published a video on 21 October addressed to the staff of AI companies. He asks them to check whether their project is the best thing they can do to delay AI's recursive self-improvement, or to make it go well, and to switch projects if it is not.

He starts from the dominant attitude inside the labs, that fully automating AI research could become possible in 2027, 2028 or 2029. He argues that chain-of-thought monitoring is losing effectiveness, that safety evaluations inspire less confidence, and that the pace of the frontier must be slowed. This is his opinion, and he acknowledges he lacks the labs' confidential information.