Useful Outsourcing is Hard (2024)
The article argues that successful outsourcing requires not just delegation but deep understanding and oversight of the task being outsourced. It explains that common outsourcing failures stem from assuming tasks are simpler than they are, and that effective outsourcing demands significant upfront investment in documentation, metrics, and management to actually reduce one's own workload long-term.
Background
- Gwern is a well-known independent AI researcher and writer, known for deep dives on topics like neural networks, rationality, and forecasting. His personal site is widely read in the AI alignment and effective altruism communities.
- This post argues that successfully outsourcing cognitive work (like research, coding, or writing) to other people — or to AI — is surprisingly difficult. The bottleneck isn't just finding someone capable; it's the overhead of specifying the task precisely enough, providing context, verifying output, and iterating.
- Gwern draws on his own extensive experience hiring research assistants and commissioning work, concluding that the "managerial load" often negates the time savings. He suggests that as AI systems improve, they may inherit this same problem: even a very capable model is useless unless you can reliably communicate what you want and evaluate what it produces.
- The piece is a grounded, pragmatic counterpoint to the common assumption that better AI or better hiring automatically makes outsourcing easy.