Much of AI doom assumes one unified, goal-directed agent that wants things. Eric Drexler's Comprehensive AI Services (CAIS) offers a calmer picture: maybe superintelligence arrives as a vast ecosystem of narrow, bounded tools — each superhuman at one task, none of them a self-preserving agent with a master goal.
Why it's more controllable
- A translator, a protein-folder, and a theorem-prover have no reason to resist shutdown — they aren't pursuing open-ended goals, so Instrumental Convergence doesn't bite.
- Capability is decoupled from agency. You get the power of superintelligence without a single will behind it.
The future might not be one god-machine, but a workshop of brilliant, obedient instruments.
The open worry: tools can be composed into agents, and competition may push us to build agentic systems anyway. Still, CAIS is a serious reason the scariest scenarios aren't inevitable.
Related: Intelligence Amplification · The Control Problem · The Abundance Argument