Ramez Naam

@ramez · Twitter · 2026-09-10 19:33 UTC

We simply don't have modeling tools that can tell you that a viral sequence is going to turn into a global pandemic with 90% lethality. Not even close to being able to do this. 100 shots on goal is likely many many orders of magnitude too few. It might take millions of attempts, and even then just might not work at this scale. It might kill some people! (Most likely by far: It kills the teenager before he ever finishes his work.) But getting this magic rainbow virus is just ... maybe impossible (we don't know if ther'es such a point in the parameter space), and certainly not something that any current sofware can predict to any useful degree. LLM progress is unlikely to help here. We're missing the fundamental training data, and getting that data requires experimentation - in actual humans. Human trials of your pandemic are likely to end with a visit from Seal Team 6. The overhwelming majority of those experiments will fail to get the effect you want, regardless. This is the flipside of why AI isn't going to solve aging without enormous amounts of experimentation. Again, you can very likely make a virus that will kill some people or infect some people. (And there are many ways to kill people today.) But the scenario you're imagining is many standard deviations away from anything that exists. The viral parameter space is just huge if you're attempting to create a pandemic. A virus that's too lethal relative to its incubation time & spread rate. A virus that spreads fast with a short incubation time gets detected quickly. (The spread is normally related to symptoms.) A virus with a long incubation time is hard to make very lethal. A virus that's asymptomatic is hard to make spread fast. And those are just population level trade-offs. The virus itself is a machine with many physical trade-offs. Every bit of the sequence in it does multiple things. There's very seldom a clean way to change one trait without having side effects on other traits. Then there's evolution. Viruses that kill their hosts off don't spread as effectively. Those that just make their hosts sick enough to spread the virus outcompete the ones that kill their hosts. So as soon as you release this, evolution is selecting for variants of the virus that are less lethal. There are just layers and layers of sci-fi here, and magical AI capabilities that can both engineer viruses with this accuracy and predict / model / engineer pandemics at this accuracy, without the many trade-offs. And then, in the real world, most labs won't actually ship you something that has a sequence that resembles a known pathogen. And that number will (and should!) shrink fast over the coming years. That's one AI safety measure I fully support!