OpenAI starts offering a biology-tuned LLM

OpenAI has introduced GPT-Rosalind, a large language model (LLM) specifically engineered for the life sciences. Unlike broader science-focused models from other tech companies, GPT-Rosalind is trained on 50 of the most common biological workflows and how to access major public databases. According to Yunyun Wang, OpenAI's Life Sciences Product Lead, the model aims to solve two major researcher pain points: managing massive genomic and protein datasets and bridging knowledge gaps across highly specialized subfields like genetics and neurobiology. The system is designed to connect genotype to phenotype, infer protein properties, and leverage mechanistic understanding to suggest biological pathways and prioritize potential drug targets.

A key differentiator is OpenAI's claim to have tuned GPT-Rosalind for skepticism, aiming to curb the sycophancy and overenthusiasm common in LLMs so it's more likely to identify bad drug targets. The company touts the model's "reasoning" and "expert-level" abilities on benchmarks, though its effectiveness against the persistent issue of AI hallucination remains to be seen. Due to concerns about potential harmful outputs—such as optimizing a virus's infectivity—access is heavily restricted. Only US-based entities can currently apply through a trusted access program, with a more limited Life Sciences Research Plugin planned for general availability. Its focused, biology-specific approach sets it apart from more generic science agents, but real-world utility reports are needed for a full evaluation.