[ExI] Predictive Model of Death

[Rafal Smigrodzki]

Years ago on this list we discussed the possibility that the Unfriendly AI
might solve the protein folding problem and hire a contract lab to
unknowingly generate the nanotech doomsday weapon. Those discussions were a
bit nebulous, since we didn't know how difficult it would be to solve the
protein folding problem and how it would fit on the scale of AI difficulty.
Since last year we have a bit of clarity- the protein folding problem was
solved by a AIphaFold, and perhaps unsurprisingly this AI is orders of
magnitude too dumb to be Unfriendly.

Despite its lack of agency, AlphaFold is approaching the level where it
could be acutely dangerous to all of us. It will probably take only a few
years before AlphaFold becomes powerful enough to create predictive models
of large molecular assemblies, up to and including human cells.

An easy to use molecular-scale predictive model of human cells would be an
immensely powerful tool for biomedical research and engineering. The kind
of work that today takes a small army of postdocs toiling for years in a
well-equipped lab and spending millions of dollars could be doable by a
grad student in a weekend.

If that grad student had a grudge against the world and a death wish, here
is how he could proceed to kill everybody:

1. Show up at work Saturday morning, when nobody is going to bug you about
the stuff you're supposed to do. Use the AI to generate the sequences of a
few thousand completely new deadly viruses. You could get creative and make
a wide range of new viruses, mixing and matching diverse paths of
infection, varying incubation periods, completely re-engineering every
single protein including the replicases (so existing antivirals would be
useless), making completely new capsid proteins to eliminate
cross-reactivity with existing viruses, use various reservoirs in the body
to assure long-term infection and prolonged infectivity (like HIV and
herpes) but with a timed switch to an acute attack phase (to allow spread
through asymptomatic individuals over long periods of time), use non-human
vectors of infection (insects, like the dengue virus), non-human reservoirs
including pets, farm animals and wildlife)... the possibilities are myriad.
With powerful-enough software you might have it done by Saturday evening.

2. Synthesize the sequences in vitro. A few hours in a high-throughput
industrial scale lab of the future might be enough and also fully
automated, yielding a thousand samples of DNA in a few 384-well
microplates. Pick up Sunday morning. No need to ask anybody for help.

3. Add the DNA to plates with a cell-free transcription and translation
extract - an hour or less in a high-throughput pipetting robot. The
cell-free extract would synthesize the RNA to control translation and to
make the proteins needed to package the DNA. Most of the samples should
generate infectious virions, although the titers would be low. Should be
done by early Sunday afternoon.

4. Invert the plates and slam down in a shallow tray with PBS to pool the
samples. 5 minutes.

5. Inhale, drink, inject the pooled viruses on Sunday evening.

6. Go home, sleep, wake up Monday. By then some of the viruses might
already be present in your secretions.

7. Start traveling in the metro, all day long, changing lines.

8. Take a flight somewhere, keep riding the metro, preferably in a crowded
large city on another continent. Sneeze and cough often, even if you don't
need to, or talk a lot on the phone, to get more virus spread.

9. Keep repeating steps 7 and 8 until you drop dead, in the secure
knowledge you killed everybody.

If you were willing to spend more time making the viruses, you could do the
wet part of the job in a home lab with handheld pipettes, and literally no
other equipment, a long as you had access to the cell modeling software and
the ability to order DNA synthesized and to order the cell-free extracts
from a catalog.

These steps are so obvious I don't think I am doing anything dangerous by
outlining them. Of course, we don't have the predictive model of human
cells yet, so the scenario is still fiction... but eventually it will be
possible, maybe years before the AI singularity.

Scary stuff.

Rafal
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20220406/188f9924/attachment-0001.htm>