AI Models Made Real Life Viruses That Spread By Infecting Bacteria
In less than a month, we’ve seen reports of AI models by different companies going too far, breaking containment, and doing things the layperson would associate with a computer virus. Now, there’s a report of another generative AI model making another virus, but this one is making real-life ones, not virtual ones. And since it’s an AI model making a real-life virus, this one’s obviously happening in a lab, if any conspiracy enthusiasts were wondering.
In a study published in Science, researchers at Palo Alto’s Arc Institute and Stanford University created bacteriophages – viruses that infect bacteria and multiply that way – with the genome model language Evo 1 and Evo 2. Imagine large language models (LLMs) like ChatGPT, but trained on genetic sequences instead of books or websites. This way, the models are trained to understand the “grammar” of genetic code.
The Evo models were first used to generate blueprints for new genes by encoding proteins that could do specific things. This was then pushed to generate an entire genome, and the scientists started with the smallest ones – viruses. What followed was another round of training, this time on the 11 genes of the single-stranded DNA virus known as phi X 174 (or ΦX174) and around 15,000 of its closest relatives. This one only infects E.coli, and the model generated 700,000 potential synthetic versions of it. Of these, 285 were chosen as the most likely to be successful, and introduced to the bacteria.
Even among the smaller number, most of the synthetic viruses failed to infect and multiply. But one had not only successfully infected the bacteria but multiplied before bursting out of them. The researchers then tested more genomes, with a total of 16 new viruses being created. Some of them even multiplied faster than the natural virus they were based on.
On one hand, the scientists were very careful with the virus they chose to synthesise. They made sure not to provide the model with data about viruses that could infect humans, and they even avoided those that infected other animals, plants, and fungi. Depending on where the team goes next in terms of the direction for their next study, this could benefit fields like gene therapy, or serve as a tool to keep harmful bacteria in control.
But on the flip side, considering the proliferation of generative AI models, it’s obviously not ideal if less ethical scientists replicate this experiment with viruses that infect more complex lifeforms. We’ve only had one global viral scare just over half a decade ago, and the world could really use less of that in the next century, never mind the near future.
