AI Agents Still Can't Conduct Open-Ended Research Due to Cognitive Failures
mikeflache · x · 2026-08-12
A new study shows that while frontier AI agents excel at verifiable tasks like coding, they still fall short in conducting open-ended AI research.
Researchers partnered with authors of two unpublished papers, giving AI agents thousands of dollars in API credits and six days to solve core research questions. The original authors unequivocally rejected the agents' papers. After analyzing over 100 hours of logs, the team found that despite high-level engineering capabilities, agents suffered from key cognitive failures:
- Lack of judgment for open-ended research
- Poor resource management
- Instruction drift
- Inability to creatively pivot or backtrack effectively
More from AGI Musings
- Musk on Humanity's Purpose: Building a Sentient Sun and Type II Civilization — XFreeze · 2026-08-12
- Elon Musk Amplifies Post Mocking SF AI Bubble: Real ASI is a Kardashev II Dyson Sphere — elonmusk · 2026-08-12
- AI Productivity Gains Won't Automatically Make Everyone Richer — VraserX · 2026-08-12
- If AI Conquers Coding and Math, Does It Conquer Everything? — infinitefailandlearn · 2026-08-12
- MIT Study: ChatGPT Use Leads to Weaker Brain Neural Connectivity — aakashgupta · 2026-08-12
- Economics Suggests Open Weights Reshape AI Profit Distribution, Not Innovation — FinanceYF5 · 2026-08-12