Open-weights advocate concedes need for tamper-resistant training algorithms
schwarzjn_ · x · 2026-09-09
schwarzjn, a strong open-weights proponent who has demonstrated full Sovereign AI strategies with open models, says it would be naive to ignore the serious safety concerns — at minimum, much stronger tamper-resistant training algorithms are needed. The post responds to deliprao's argument that concentration of power is the real issue and open models should handle most daily tasks, since we don't solve Navier-Stokes-level problems every day.
More from AGI Musings
- In favor of (some) anthropomorphization of AI systems, from a leading AI researcher — manoelribeiro · 2026-09-09
- Terry Tao Warns AI-Era Scooping Could Push Math and Science Into Silence — windx0303 · 2026-09-09
- Reviewer finds AI-written peer reviews rampant at ARR, calls for benchmarks of AI reviews — ChenhaoTan · 2026-09-09
- The Question Everyone Ignored in Social Media: Who Owns Your Data in the Personal Agent Wars? — MartinGTobias · 2026-09-09
- Ethan Mollick: with AI advancing fast, some hard research is better left for later — emollick · 2026-09-09
- Instead of pausing AI, build auditable alignment agents to watch other agents, argues founder — benfielding · 2026-09-09