arstechnica.com
|
ksl
|
|
OpenAI released GPT-Rosalind, a model trained on 50 common biological workflows and connected to major public databases for tasks like prioritizing drug targets and inferring protein properties. Named after Rosalind Franklin, the model was specifically tuned to be more skeptical than standard GPT outputs – pushing back when something looks like a bad target rather than defaulting to enthusiasm. Access is restricted to US-based institutions and individually vetted, citing dual-use risks around pathogen optimization. Most science-focused models from Google DeepMind and others have stayed generalist; OpenAI going narrow on biology signals a bet that domain-specific fine-tuning will matter more than breadth for research applications. The gated access model mirrors what OpenAI did with GPT-5.4-Cyber just days earlier
