LinkedIn publication text
When AI Starts Building Itself
AI does not yet autonomously develop its own successors, but it already speeds up the research and engineering work behind building them (Anthropic Institute, 2026). Not full recursive self-improvement, but the technical and organizational pieces are being put in place.
In this 16-slide study:
- From tool to agent to still-hypothetical autonomy.
- METR time horizons: agents sustaining longer chains of work.
- Research taste: choosing what to investigate is still the hard part.
- Amdahl's Law: when generating outpaces verifying, review becomes the bottleneck.
- The oversight paradox: the more necessary, the harder to exercise.
- Govern the cycle, not just the model: separation of roles, infrastructure containment, and auditability.
- Three possible futures: slowdown, advanced automation with human direction, and full autonomy.
The urgent question is not when, but whether our institutions will be ready to understand, verify, and govern this loop before it operates faster than our ability to control it.
How do you see RSI and monitoring in your context?
Aridio Silva
Independent Researcher
@aridiosilva