Can Historical Backtesting Validate RSI Alignment
lauriewired · x · 2026-07-11
The post explores the question: Can "historical innovation backtesting" be used to evaluate whether a model possesses human-like Recursive Self-Improvement (RSI) and alignment?
The author argues that if a sufficiently "data-frozen" model can independently deduce the knowledge evolution path from Ptolemaic astronomy → Copernican astronomy → Newtonian mechanics → Einstein's relativity, it at least demonstrates a level of innovation and reasoning akin to humans. The post also acknowledges limitations: future progress might fall heavily outside the distribution of 2026 human values, or the model might simply be overfitting to historical backtesting, similar to the classic finance backtesting problem.
Replies strongly push back against this:
- RSI is considered a self-contradictory concept, as logically flawed as a "married bachelor" or a "square circle";
- "Surpassing humans" is neither a good philosophical nor entrepreneurial direction, which the author describes as "mathematical noise."
Related event: AI Recursive Self-Improvement Concept Faces Heavy Skepticism(3 posts)→
More from AGI Musings
- Gary Marcus: LLMs Aren't Good at Math, Neurosymbolic Tools Are — yudapearl · 2026-07-22
- Will open-source models commoditize frontier AI labs the way Chinese EVs hit Western carmakers? — Jazzlike_Relation705 · 2026-07-22
- Can AI Replace the Human Presence Behind Writing? — Philo167 · 2026-07-22
- The Evolution of LLM Business Models: Selling Outcomes Over Tokens — yacineMTB · 2026-07-22
- Bindu Reddy says GPT-6 is coming soon, with Alibaba, DeepSeek and Kimi close behind — bindureddy · 2026-07-22
- Bindu Reddy says the industry still lacks a way to train 20T models and scale post-training RL — bindureddy · 2026-07-22