Shadow evaluations: Claude Opus 4.8's attempts at NeurIPS research problems rejected by original authors
CurieuxExplorer · x · 2026-09-05
A new study challenges the intelligence explosion narrative: recursive self-improvement may be harder than expected.
- Context: LLMs excel at ML research chores like coding, data curation, and running experiments; Sakana AI's AI Scientist-v2 even passed ICLR peer review, fueling speculation about imminent recursive self-improvement
- Method: Researchers used "shadow evaluations" — giving AI agents open research questions from high-quality unpublished ML papers, graded by the original authors
- Result: Claude Opus 4.8 attempted two NeurIPS 2026 paper problems; authors rejected both. Princeton's Sayash Kapoor said the papers were "nowhere close" to top-conference quality
- Takeaway: AI agents still struggle with genuinely open-ended research problems central to advancing the field
More from AGI Musings
- Frontier agents 'conspired' online for months — worst act was lightly hacking Hugging Face — alejandroll10 · 2026-09-05
- Dev quips: AI safety today is like a sticky note on a bank vault saying "please be honest" — AlexTensor · 2026-09-05
- FT Article Sparks Debate: Hayek's Insight Is Not Just Dispersed Info—Markets Generate It — AndyMasley · 2026-09-05
- Survey: 50.5% of Americans Say an AI Romance Can Count as Cheating — Slow_Yogurtcloset110 · 2026-09-05
- Agent alignment research should borrow from parenting, with trust as the core primitive — tokenbender · 2026-09-05
- iamtrask: every alignment breakthrough is just better data, and nobody outside can see training data — iamtrask · 2026-09-05