Twenty Years from RSI to Takeoff

Why it was saved

Cummins’s framing: “All the RSI talk seems very overhyped.” The productivity gains are real but prosaic, and the consequences are prosaic too. He offers the post as a good “nothing ever happens” timeline that is “still very radical” over several decades. In the replies he grants that an algorithmic breakthrough could change everything, “but you can’t time those.”

Key claims

• Fast reasoning, slow learning. Current methods likely enable automated general learning (AGI) soon, via automated creation of RL tasks, environments and graders that fill a model’s visible gaps, followed by automated next-model building. But that loop runs at the speed of training a next model (weeks to months per iteration), not the speed of token generation.
• Humans aren’t the bottleneck; compute is. Much of the loop is waiting for training compute. Smaller, faster models don’t help because today’s frontier models are likely only borderline capable of the skills needed; the 2028–2031 models that clearly are will be even bigger.
• Hardware tricks don’t escape FLOPs. Faster inference hardware doesn’t reduce the number of FLOPs required; cost is fundamentally the amount of compute, which only the scale of the global buildout can raise.
• The buildout slows after ~2032. LLMs become ready to close the next-model loop just as industry runs out of fuel for another near-term 1000x scale-up of compute, as decommissioning starts to catch up with new production.
• Industrial explosion as the real accelerant. The next 1000x speed-up of model-building loops comes from AI-driven industrial expansion, by roughly 2050. So “nothing substantively new” might get invented until 2040–2050, though an ASI breakthrough remains possible at any time.

Connections

• Recursive Self-Improvement · AI Timelines · Compute
• Counterweight to the Fable 5.2 launch-day hype in Claude Fable 5.2 — Recommendations.