LSI identifies three classes of friction in the AI self-improvement loop: (1) narrowness of automatable research — agents are good at optimizing single metrics (test loss) but weak at balancing multiple objectives simultaneously, which is the actual job of top researchers; (2) diminishing returns from parallel agents — analogously to Amdahl's law, scaling the number of agents on a task saturates quickly because they all sample from a similar solution distribution; (3) resource and political bottlenecks — allocation of billions of dollars in compute remains under human organizational control. Each friction translates into "loss" of gain per cycle, bringing the hypothetically exponential RSI down to a more linear trajectory.
The narrative of an impending intelligence explosion via RSI ignores empirical frictions in the AI model development process. LSI provides the language to precisely discuss those frictions and forecast a trajectory that is more linear than exponential.
AI agents are good at optimizing local metrics (test loss, single reward) but not at balancing many competing objectives that characterize real research.
Adding more agents to a problem yields rapidly saturating speedup — all sample from a similar solution distribution and are bottlenecked by human supervision.
Allocation of billion-dollar compute budgets is an organizational decision, not an algorithmic one — AI cannot autonomously draw on those resources.
The more we understand intelligence, the harder further progress becomes — a law of diminishing returns for the entire research system.
Early stages of a sigmoid look exponential. LSI warns that the spectacular 2023-2026 improvements need not be sustained.
RSI scenarios assuming autonomous access to billion-dollar compute budgets are unrealistic — allocation remains under human organizational control.
A single researcher cannot meaningfully supervise hundreds of agents per day — this is a fundamental productivity-growth limit.
AI successes at optimizing local metrics (Karpathy autoresearch, kernel writing) do not automatically translate to the ability to design entire models.
Iterative training with internally generated data requires GPU for both inference (data generation) and training — often on the same cluster.
