High quality quantum chemistry simulations can take days or weeks to run, even on supercomputing clusters.
This doesn’t seem very long for an AGI if they’re patient and can do this undetected. Even months could be tolerable? And if the AGI keeps up with other AGI by self-improving to avoid being replaced, maybe even years. However, at years, there could be a race between the AGIs to take over, and we could see a bunch of them make attempts that are unlikely to succeed.
That’s one simulation though. If you have to screen hundreds of candidate structures, and simulate every step of the process because you cannot run experiments, it becomes years of supercomputer time.
Although they never take the whole supercomputer, so if you have the whole supercomputer for yourself and the calculations do not depend on each other you can run many in parallel
They do in fact routinely take up the entire supercomputer. I don’t think you are comprehending the absolute mind-bogglingly huge scale of compute required to get realistic simulations of mechanosynthesis reactions. It’s utterly infeasible without radical advances beyond the present state of the art classical compute capabilities.
My job was doing quantum chemistry simulations for a few years, so I think I can comprehend the scale actually.
I had access to one of the top-50 supercomputers and codes just do not scale to that number of processors for one simulation independently of system size (even if they had let me launch a job that big, which was not possible)
Yeah just for scaling problems are an issue. These are not embarrassingly parallel problems that can be easily sharded onto different CPUs. A lot of interconnect is needed, and as you scale up that becomes the limiting factor.
What codes did you use? The issue here is that DFT gives bad results because it doesn’t model the exchange-correlation energy functional very well, which ends up being load bearing for diamondoid construction reaction pathways performed in UHV using voltage-driven scanning probe microscopy to perform mechanochemistry. You end up having to do some form of perturbation theory, of which there are many variants all giving significantly different results, and yet are all so slow that running ab initio molecular dynamics simulations are utterly out of the question so you end up spending days of supercompute time just to evaluate energy levels of a static system and then use your intuition to guess whether the reaction would proceed. Moving to an even more accurate level of theory that would indisputably give good results for these systems but would have prohibitive scaling factors [O(n^5) to O(n^7)] preventing their use for anything other than toy examples.
I am at a molecular nanotechnology startup trying to bootstrap diamondoid mechanosynthesis build tools, so these sorts of concerns are my job as well.
Ahh for MD I mostly used DFT with VASP or CP2K, but then I was not working on the same problems. For thorny issues (biggish and plain DFT fails, but no MD) I had good results using hybrid functionals and tuning the parameters to match some result of higher level methods. Did you try meta-GGAs like SCAN? Sometimes they are suprisingly decent where PBE fails catastrophically...
For the most part we’re avoiding/designing around compute constraints. Build up our experimental validation and characterization capabilities so that we can “see” what happens in the mechsyn lab ex post facto. Design reaction sequences with large thermodynamic gradients, program the probe to avoid as much as possible side reaction configurations, and then characterize the result to see if it was what you were hoping for. Use the lab as a substitute for simulation.
It honestly feels like we’ve got a better chance of just trying to build the thing and repeating what works and modifying what doesn’t, than even a theoretically optimal machine learning algorithm could do using simulation-first design. Our competition went the AI/ML approach and it killed them. That’s part of why the whole AGI-will-design-nanotech thing bugs me so much. An immense amount of money and time has been wasted on that approach, which if better invested could have gotten us working nano-machinery by now. It really is an idea that eats smart people.
You could also try to fit an ML potential to some expensive method, but it’s very easy to produce very wrong things if you don’t know what you’re doing (I wouldn’t be able for one)
These are potentially tremendously helpful. But in the context of AI x-risk it’s still not enough to be concerning. A force field that gives accurate results 90% of the time would tremendously accelerate experimental efforts. But it wouldn’t be reliable enough to one-shot nanotech as part of a deceptive turn.
Also, maybe we design scalable and efficient quantum computers with AI first, and an AGI uses those to simulate quantum chemistry more efficiently, e.g. Lloyd, 1996 and Zalka, 1996. But large quantum computers may still not be easily accessible. Hard to say.
This doesn’t seem very long for an AGI if they’re patient and can do this undetected. Even months could be tolerable? And if the AGI keeps up with other AGI by self-improving to avoid being replaced, maybe even years. However, at years, there could be a race between the AGIs to take over, and we could see a bunch of them make attempts that are unlikely to succeed.
That’s one simulation though. If you have to screen hundreds of candidate structures, and simulate every step of the process because you cannot run experiments, it becomes years of supercomputer time.
Although they never take the whole supercomputer, so if you have the whole supercomputer for yourself and the calculations do not depend on each other you can run many in parallel
They do in fact routinely take up the entire supercomputer. I don’t think you are comprehending the absolute mind-bogglingly huge scale of compute required to get realistic simulations of mechanosynthesis reactions. It’s utterly infeasible without radical advances beyond the present state of the art classical compute capabilities.
My job was doing quantum chemistry simulations for a few years, so I think I can comprehend the scale actually. I had access to one of the top-50 supercomputers and codes just do not scale to that number of processors for one simulation independently of system size (even if they had let me launch a job that big, which was not possible)
Yeah just for scaling problems are an issue. These are not embarrassingly parallel problems that can be easily sharded onto different CPUs. A lot of interconnect is needed, and as you scale up that becomes the limiting factor.
What codes did you use? The issue here is that DFT gives bad results because it doesn’t model the exchange-correlation energy functional very well, which ends up being load bearing for diamondoid construction reaction pathways performed in UHV using voltage-driven scanning probe microscopy to perform mechanochemistry. You end up having to do some form of perturbation theory, of which there are many variants all giving significantly different results, and yet are all so slow that running ab initio molecular dynamics simulations are utterly out of the question so you end up spending days of supercompute time just to evaluate energy levels of a static system and then use your intuition to guess whether the reaction would proceed. Moving to an even more accurate level of theory that would indisputably give good results for these systems but would have prohibitive scaling factors [O(n^5) to O(n^7)] preventing their use for anything other than toy examples.
I am at a molecular nanotechnology startup trying to bootstrap diamondoid mechanosynthesis build tools, so these sorts of concerns are my job as well.
Ahh for MD I mostly used DFT with VASP or CP2K, but then I was not working on the same problems. For thorny issues (biggish and plain DFT fails, but no MD) I had good results using hybrid functionals and tuning the parameters to match some result of higher level methods. Did you try meta-GGAs like SCAN? Sometimes they are suprisingly decent where PBE fails catastrophically...
For the most part we’re avoiding/designing around compute constraints. Build up our experimental validation and characterization capabilities so that we can “see” what happens in the mechsyn lab ex post facto. Design reaction sequences with large thermodynamic gradients, program the probe to avoid as much as possible side reaction configurations, and then characterize the result to see if it was what you were hoping for. Use the lab as a substitute for simulation.
It honestly feels like we’ve got a better chance of just trying to build the thing and repeating what works and modifying what doesn’t, than even a theoretically optimal machine learning algorithm could do using simulation-first design. Our competition went the AI/ML approach and it killed them. That’s part of why the whole AGI-will-design-nanotech thing bugs me so much. An immense amount of money and time has been wasted on that approach, which if better invested could have gotten us working nano-machinery by now. It really is an idea that eats smart people.
You could also try to fit an ML potential to some expensive method, but it’s very easy to produce very wrong things if you don’t know what you’re doing (I wouldn’t be able for one)
It’s coming along quite fast. Here’s the latest on ML-trained molecular dynamics force fields that (supposedly) approach ab initio quality:
https://www.nature.com/articles/s41524-024-01205-w
These are potentially tremendously helpful. But in the context of AI x-risk it’s still not enough to be concerning. A force field that gives accurate results 90% of the time would tremendously accelerate experimental efforts. But it wouldn’t be reliable enough to one-shot nanotech as part of a deceptive turn.
Also, maybe we design scalable and efficient quantum computers with AI first, and an AGI uses those to simulate quantum chemistry more efficiently, e.g. Lloyd, 1996 and Zalka, 1996. But large quantum computers may still not be easily accessible. Hard to say.