Amplification to Replacement

AI assistance is economically unstable when the human remains the serial bottleneck. If a worker uses AI to become somewhat faster, the workflow is still limited by the worker’s review, taste, memory, and attention. If the worker can be removed, the system can scale by running many autonomous instances.

This is the Amdahl’s-law argument applied to knowledge work. The slow serial component caps the speedup of the whole process. As long as human review is necessary, the human is still needed. The same fact also identifies the next optimization target. Every improvement that reduces the need for human review makes replacement more attractive.

Some intellectual roles are harder to remove. “AI assistant” can be a transitional configuration that reveals which human actions remain bottlenecks. The harder-to-remove roles are those whose removal changes the objective, destroys the standard of evaluation, or eliminates the importance judgment that made the work worth doing.

The connection with the importance oracle is direct. If the human only supplies labor inside a verifiable pipeline, the economic pressure is to automate that labor. If the human supplies the objective, the taste function, the institutional interpretation, or the responsibility for what should be pursued, replacement is harder because the missing object sits outside the pipeline.

Blog Use

Use this near the end, when the essay moves to labor and institutions. It explains why augmentation can be temporary rather than stable.

Source Trail

• Gwern Branwen, “Guardian Angels: LLM Personalization for Productivity and Security”, gwern.net/guardian-angel.
• ../ai_views.md, “Amplification as waystation, not destination”.
• See reference_inventory.md for full source status.