"Useful Outsourcing Is Hard", by Gwern · Gwern.net
If you can't figure out how to use AI effectively, it's probably not because the AI is too stupid—it's because you've never learned how to outsource work to anyone, period, and your entire workflow is optimized against it.
Read OriginalMy Notes (1)
There are few valuable “AI-shaped holes” because we’ve organized everything to minimize the damage from lacking AI to fill those holes, as it were: if there were some sort of organization which had naturally large LLM-shaped holes where filling them would massively increase the organization’s output… It would’ve gone extinct long ago and been replaced by ones with human-shaped holes instead, because humans were all you could get.
Summary used for search
TLDR
• The "free human worker" test: If you couldn't outsource a task to a free human assistant, you can't outsource it to AI either—this removes the distraction of capability and reveals the real problem is outsourcing itself
• "Automation as colonization wave"—there are few "AI-shaped holes" in our work because any organization with such holes would have died decades ago and been replaced by ones with human-shaped holes instead
• Gwern's concrete example: Despite being a writer, he can only spend ~$50/month on LLMs because his workflow requires fine-tuning on his corpus, retrieval over clippings, local deployment, custom tools—the capability exists but the integration doesn't
• The discontinuous jump: LLMs may become highly capable without being useful for outsourcing to you... right until they suddenly can just replace you entirely
• Practical experiment: Try hiring a $1,000/month remote assistant—if you can't find useful work for them, that fully explains your AI adoption struggles
In Detail
Gwern argues that the primary barrier to AI usefulness isn't capability but the fundamental difficulty of outsourcing itself. He introduces a simple diagnostic: "Could you outsource it to a human being who worked for free?" If the answer is no, then it's not a matter of artificial versus natural intelligence—it's that you've never learned to outsource effectively and your workflows aren't designed for it.
The core insight is "automation as colonization wave"—major technologies take decades to have massive effects because everyone is stuck in local optima. There are few "AI-shaped holes" in our organizations because any entity with such holes would have gone extinct long ago, replaced by ones with human-shaped holes instead. This explains why even famous, time-constrained people historically didn't outsource most things. Using his own writing as an example, Gwern shows that while LLMs could theoretically write his mini-essays (they have the raw capability), actually making this work would require: fine-tuning on his entire corpus, retrieval over his clippings database, local deployment to access his files, custom tool integration for his Markdown extensions, and more. The capability exists, but the 10+ infrastructure projects needed to unlock it don't.
The dangerous implication: LLMs may become highly capable without being useful for outsourcing to individuals... right until the point where they suddenly can replace that individual entirely. The transition won't be gradual productivity gains—it will be a discontinuous jump from "barely useful" to "replaces you completely." His practical suggestion: experiment by hiring a $1,000/month remote assistant. If you struggle to find valuable work for them (and many will—some companies make you take courses on how to use an EA effectively), that fully explains why you can't spend $1,000/month on AI either. The problem isn't the intelligence available; it's that you've never reorganized your life to make any form of outsourcing possible.