kalomaze calls ES/genetic-algorithm training pointless unless backprop is truly intractable
kalomaze · x · 2026-09-30
Researcher kalomaze argues that evolutionary-strategy / genetic-algorithm training has always felt boring to him — "Sakana-tier" in terms of pointless nerdsniping. His bar for abandoning backprop: only do it in regimes where naive backprop is intractable without BPTT, citing synth gradients / DNI-style work as the right kind of setting. A notable technical take in the ongoing debate over alternatives to backpropagation.
More from AGI Musings
- 'Machine rights' debate is a distraction from engineering and corporate accountability — AlexTensor · 2026-09-30
- Williams College philosophy chair questions whether writing is the highest form of thinking — soumitrashukla9 · 2026-09-30
- Commenter pushes back on 'rogue AI' framing: Anthropic's risk lies with its creators — AlexTensor · 2026-09-30
- "Against the AI Omnicause": New Essay Says Safety Orgs Dodge Extinction Risk — zetalyrae · 2026-09-30
- AI researcher: years of grading essays show thinking isn't central to writing — dioscuri · 2026-09-30
- Renaming AI to 'Super Intelligence' is like rebranding the caterpillar industry as butterflies — jam3scampbell · 2026-09-30