A technical case against RSI: transformers have capacity limits, more data causes forgetting
gerardsans · x · 2026-09-12
gerardsans makes a technical argument that RSI is a nothing burger: software isn't smooth or infinite — once layers, inputs, outputs and parameters are set, it's done. AI is probabilistic but still software; a transformer has a finite capacity, and you can't train a model indefinitely — continued training with more data leads to catastrophic forgetting.
Related event: Devs debate AI risk: real danger is misuse, not recursive self-improvement(3 posts)→
More from AGI Musings
- Mathematicians issue declaration: AI companies' benchmark race misaligned with math community — rao2z · 2026-09-12
- Musk: AGI risk far exceeds nuclear weapons, as doomer-verse debates shift — Linahuaa · 2026-09-12
- Terence Tao and 24 Fields Medalists sign letter blasting AI labs over math — alexbilz · 2026-09-12
- Siemens: Managers Now Delegate to Junior Programmers and AI Agents Side by Side — erikbryn · 2026-09-12
- 100 LLM Agents Run a Town Economy for 26 Weeks — and Money Stops Moving — omarsar0 · 2026-09-12
- 'Marketplace of rationalizations': AI risk discourse lets you believe anything by picking experts — xuanalogue · 2026-09-12