Hacker News AIT2·68 pts
- AI agents demonstrate capability in well-defined local optimization tasks, such as Autoresearch-style GPU kernel generation.
- Current agents still have very low success rates in end-to-end research projects that require long-term planning, hypothesis testing, and cross-domain knowledge integration.
- 'Automated AI R&D' currently functions more as an enhancement to human researcher toolchains rather than a complete replacement for human research subjects.