We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Dec 16·5 min read
Measuring AI’s capability to accelerate biological research
TL;DR
Biological researchers lack practical frameworks to measure whether AI actually speeds up wet lab work (hands-on experimentation). OpenAI built a real-world evaluation framework and used GPT-5 to optimize molecular cloning protocols—the process of copying DNA segments.
✦ Why It Matters
Engineers can adopt this evaluation framework to measure AI's real-world impact on domain-specific lab workflows beyond theoretical metrics.
Key Takeaways
- Build evaluation frameworks grounded in actual lab workflows, not just benchmark datasets.
- Measure AI impact through concrete metrics: cost reduction, success rates, and time savings.
- Identify and document safety risks when deploying AI to guide hands-on experimentation.
- Require expert review checkpoints before AI recommendations are executed in physical labs.
Full Summary
Biological research relies on wet lab experimentation, where scientists physically conduct chemical and biological tests. While AI shows promise in accelerating research, few rigorous frameworks exist to measure real-world impact on actual lab workflows.
OpenAI developed an evaluation framework grounded in practical experimentation and deployed GPT-5—their latest language model—to optimize molecular cloning protocols, which are standard procedures for duplicating and manipulating DNA sequences. The team measured how AI suggestions improved protocol efficiency, cost, and success rates compared to baseline approaches.
Results demonstrated measurable acceleration in specific tasks, though the work also surfaced risks including potential errors in protocol design and safety concerns when AI recommendations bypass expert review. These findings establish a methodology for quantifying AI's genuine contribution to biological research rather than relying on theoretical benchmarks alone.