OpenAI is preparing to test its latest GenAI models in the bioscience wing of the Los Alamos National Laboratory, a once top-secret research base that played a central role in the development of the first nuclear weapons.
The prospect of a world-leading AI company letting its smartest creations run riot in a bioweapons lab is likely to alarm existential risk alarmists. Yet GPT-4 (or five, if it ever arrives) isn't about to be put to work crafting next-generation hypersonic nukes or breeding new strains of anthrax.
Whilst Los Alamos certainly does build weapons and is proud home of “the nation’s premiere plutonium science and manufacturing facility”, it does not appear as if GPT-4 will make it out of a lab focused on gentler - but still risky - research.
GPT-4o and its descendants will be tested in a department focused on “protecting the nation from biological, chemical, and environmental threats”, which carries out vaccine research, develops biosurveillance techniques and studies antimicrobial resistance.
A test of OpenAI's LLMs in Los Alamos “will be the first of its kind and contribute to state-of-the-art research on AI biosecurity evaluations”, the lab said in a statement.
Previous research found that ChatGPT-4 "provided a mild uplift in providing information that could lead to the creation of biological threats." Which means it could make it easier for terrorists or rogue states to build biological weapons.
As models become more powerful, their ability to produce doomsday weapons is likely to grow - meaning action at this early stage could potentially head off disaster before it happens.
The threat of “AI-enabled biological threats” is “significant”, the National Laboratory warned. “But existing work has not assessed how multimodal, frontier models could lower the barrier of entry for non-experts to create a biological threat," it continued.
"The potential upside to growing AI capabilities is endless,” said Erick LeBrun, research scientist at Los Alamos. “However, measuring and understanding any potential dangers or misuse of advanced AI related to biological threats remain largely unexplored. This work with OpenAI is an important step towards establishing a framework for evaluating current and future models, ensuring the responsible development and deployment of AI technologies.”
The rise of multi-modal models
Los Alamos will also be the first lab test of OpenAI’s multi-modal models, which can process a variety of inputs including text, images, and audio. They will work on tasks intended to "serve as a proxy for more complex tasks that pose a dual use concern.”
Dual use research is officially defined as a “subset of life sciences research that, based on current understanding, can be reasonably anticipated to provide knowledge, information, products, or technologies that could be directly misapplied to pose a significant threat with broad potential consequences to public health and safety, agricultural crops and other plants, animals, the environment, materiel, or national security.”
The experiment will involve "assessing the abilities of both experts and novices to perform and troubleshoot a safe protocol consisting of standard laboratory experimental tasks.”
Mira Murati, OpenAI’s Chief Technology Officer, said: “This partnership marks a natural progression in our mission, advancing scientific research, while also understanding and mitigating risks.”
Tasks for AI models may include “transformation”, which can involve introducing foreign genetic material into a host organism, or maintaining and propagating cells in vitro and cell separation using techniques such as centrifugation.
By examining the “uplift in task completion and accuracy enabled by GPT-4o”, OpenAI aims to “quantify and assess how frontier models can upskill both existing professionals / PhDs as well as novices in real-world biological tasks.”
Addressing AI risk in a bioweapons lab
A key part of the study will involve risk reduction - which is welcome, considering the dual-use nature of the work involved in the partnership.
Los Alamos National Laboratory has “been a pioneer in safety research”, OpenAI said, and it hopes to collaborate on developing “novel and robust safety evaluations for frontier AI models as capabilities continue to rapidly improve.”
The cooperative effort is designed to “underscore” the scientific research potential of multimodal AI models like GPT-4o and emphasise the importance of private and public sector collaboration in “leveraging innovation and ensuring safety”.
OpenAI hopes the partnership “will help set new standards for AI safety and efficacy in the sciences, paving the way for future innovations that benefit humanity."
“AI is a powerful tool that has the potential for great benefits in the field of science, but, as with any new technology, comes with risks,” said Nick Generous, deputy group leader for Information Systems and Modeling. "At Los Alamos this work will be led by the laboratory's new AI Risks Technical Assessment Group, which will help assess and better understand those risks.”
