HOT RADAR
agentHEAT 0.0°

AI CLUSTERED EVENT · 10/10/2026

Epoch AI Evaluates Agent Automation in R&D: Low End-to-End Success but Emerging Algorithmic Discovery

1 reports archived0 independent sourcesupdated 10/10/2026, 06:14:40
Synthesis & Latest Updates

Epoch AI released a new evaluation assessing AI agents' performance in automating AI research and development (R&D). While agents successfully handle specific sub-tasks like GPU kernel optimization or dataset creation, they exhibit extremely low success rates in end-to-end research projects requiring complex planning. Current AI cannot yet independently replace human researchers for open-ended, multi-step scientific exploration, showing only preliminary algorithmic discovery capabilities within narrowly defined optimization metrics.

LATEST/Epoch AI released a new evaluation assessing AI agents' performance in automating AI research and development (R&D). While agents successfully handle specific sub-tasks like GPU kernel optimization or dataset creation, they exhibit extremely low success rates in end-to-end research projects requiring complex planning. Current AI cannot yet independently replace human researchers for open-ended, multi-step scientific exploration, showing only preliminary algorithmic discovery capabilities within narrowly defined optimization metrics.

HEAT TRENDHourly heat curve

48 fully observed hours
Now
0.0
Peak
13.509:00
24h change
-100%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. 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.