omarsar0: GPT and Claude can't run reliable autoresearch loops anymore — own your intelligence stack
omarsar0 · x · 2026-11-05
AI commentator omarsar0 argues that reliable groundbreaking AI research via autoresearch loops is no longer possible with GPT or Claude, citing constant inaccurate refusals — vendors should fix their classifiers. He clarifies he means simple research loops (tuning harnesses) for hill climbing, not model training. His thesis: in an RSI world, anyone planning to use these models for useful work — labs or any business using AI in products — will struggle. His takeaway from the past year: build your own thing and own your intelligence stack.
More from AGI Musings
- Survey of 200 CIOs: Enterprise AI returns concentrate by workload and rise with scale — BenBajarin · 2026-11-06
- Democrats push for day-1 ban on superintelligent AI with up to 20 years in prison — KyeGomezB · 2026-11-06
- In Banks's Culture novels, unlimited compute dissolves envy via simulated lives — yunta_tsai · 2026-11-06
- "People hate AI not because of bad PR, but because it drains meaning from life" — juliey4 · 2026-11-06
- Gary Marcus: Market spooked by idea AI companies may finally face accountability — GaryMarcus · 2026-11-06
- "Any definition of safety antagonistic to capability will never work" sparks debate — thisismyhat · 2026-11-06