To address LLMs’ tendencies toward sycophancy and overenthusiasm, OpenAI says it has tuned the model to be more skeptical, so it’s more likely to tell you when something is a bad drug target. There was a lot of talk about GPT-Rosalind’s “reasoning” and “expert-level” abilities. We were told that the former was defined as being able to work through complex, multi-step processes, while the latter was derived from the model’s performance on a handful of benchmarks.

It’s unclear whether OpenAI has tackled the hallucination issue that has plagued a variety of LLMs and can also strike when the systems are prompted to explain the steps the company took to reach its conclusions. Given past experience, it’s likely we’ll see a mix of glowing reports about unexpected connections the AI finds, as well as instances where it produces obviously erroneous suggestions.

For the moment, however, the company is limiting access due to concerns about the model’s potential for harmful outputs if asked to do something like optimize a virus’s infectivity. Only US-based entities can apply to OpenAI’s trusted access deployment structure at the moment, and the company will limit who can use it. A more limited Life Sciences Research Plugin will be made generally available.

As noted above, a number of other companies have made science-focused agentic LLMs available, but those were much less focused than GPT-Rosalind, which is biology-specific. Until we start hearing reports on the effectiveness of this new model, it’s difficult to evaluate whether this focus improves its utility.