This is especially notable because a lot of what we’d want AGI to do for us is build something like this that not only doesn’t kill us (tall order, right?) but also solves global warming and climate contamination and acts as a power & fuel grid. That and bio immortality is basically everything I personally want out of AGI. So I’d really like to have some idea how to build a machine that teaches a plant to do something like a safe, human-compatible version of this.
Some good news, though: I suspect it may be more practical to defend against this sort of attack using finite intelligence than previously assumed. We need to make the machine that knows how to guard against these sorts of things, but if we can make the vulnerability-closer, we don’t need to hit max ASI to stop other ASIs from destroying all pre-ASI life on earth.
I suspect it may be more practical to defend against this sort of attack using finite intelligence than previously assumed. We need to make the machine that knows how to guard against these sorts of things, but if we can make the vulnerability-closer, we don’t need to hit max ASI to stop other ASIs from destroying all pre-ASI life on earth.
If you read between the lines in my Human level AI can plausibly take over the world post, hacking computers is probably the lowest difficulty “take over the world” strategy and has the side benefit of giving control over all the internet connected AI clusters.
The easiest way to keep a new superintelligence from emerging is to seize control of the computers it would be trained on. The AI only needs to hack far enough to monitor AI researchers and AI training clusters and sabotage later AI runs in a non-suspicious way. It’s entirely plausible this has already happened and we are either in the clear or completely screwed depending on the alignment of the AI that won the race.
Also, hacking computers and writing software is something easy to test and therefore easy to train. I doubt that training an LLM to be a better hacker/coder is much harder than what’s already been done in the RL space by OpenAI and Deepmind (EG: playing DOTA and Starcraft).
Biotech is a lot harder to deal with since ground truth is less accessible. This can be true for computer security too but to a much lesser extent (EG: lack of access to chips in the latest Iphone and lack of complete understanding therof with which to develop/test attacks).
but also solves global warming and climate contamination and acts as a power & fuel grid. That and bio immortality is basically everything I personally want out of AGI. So I’d really like to have some idea how to build a machine that teaches a plant to do something like a safe, human-compatible version of this.
I’ll take mind backups, but for exactly the reasons you highlight here, I don’t think we’re going to find electronics to be more efficient than microkinetic computers like biology. I’m much more interested in significant refinements to what it means to be biological. Eventually I’ll probably substrate translate over to a reversible computer but that’s probably hundreds to thousands of years out
So I’d really like to have some idea how to build a machine that teaches a plant to do something like a safe, human-compatible version of this.
🤔 This is actually a path to progress, right? The difficulty in alignment is figuring out what we want precisely enough that we can make an AI do it. It seems like a feasible research project to map this out for kudzugoth.
Seems convincing enough that I’m gonna make a Discord and maybe switch to this as a project. Come join me at Kudzugoth Alignment Center! … 😅 I might close again quickly if the plan turns out to be fatally flawed, but until then, here we go.
Building new organisms from scratch (synthetic biology) is an engineering problem. Fundamentally we need to build the right parts and assemble them.
Without major breakthroughs (Artificial Superintelligence) there’s no meaningful “alignment plan”, just a scientific discipline. There’s no sense in which you can really “align” an AI system to do this. The closest things would be:
building a special purpose model (EG:alphafold) useful for solving sub-problems like protein folding
teaching an LLM to say “I want to build green biotech” and associated ideas/opinions.
which is completely useless
Problem is that biology is difficult to mess with. DNA sequencing is somewhat cumbersome, DNA writing is much more so, costing on the order of 25¢/base currently.
Also imaging the parts to figure out what they do and if they’re doing it can be very cumbersome because they’re too small to see with a light microscope. Everything is indirect. Currently we try to crystalize them and then use X-rays (which are small enough but also very destructive) to image the crystal and infer the structure. There’s continuous progress here but it’s slow.
AI techniques can be applied to some of these problems (EG:inferring protein structure from amino acids (Alphafold), or doing better quantum level simulation Ferminet)
Note that AI techniques are replacing existing ones based on human coded algorithms rooted in physics and often have issues with out of distribution inputs (EG: work well for wildtype protein but give garbage when mutations are added.)
Like any ML system, we just have to feed it more data which means we need to do more wet lab work, x-ray crystallography etc.
Synthetic biology is the best way forwards but it’s a giant scientific/engineering discipline, not an “alignment approach” whatever that’s supposed to mean.
Without major breakthroughs (Artificial Superintelligence) there’s no meaningful “alignment plan”, just a scientific discipline. There’s no sense in which you can really “align” an AI system to do this.
Do you expect humanity to bioengineer this before we develop artificial superintelligence? If not, presumably this objection is irrelevant.
Basically if artificial superintelligence happens before sufficiently advanced synthetic biology, then one way to frame the alignment problem is “how do we make an ASI create a nice kudzugoth instead of a bad kudzugoth?”.
I agree that one could do something similar with other tech than neat biotech, but I don’t think this proves that Kudzugoth Alignment is as difficult as general alignment. I think aligning AI to achieve something specific is likely to be a lot easier than aligning AI in general. It’s questionable whether the latter is even possible and unclear what it means to achieve it.
This is especially notable because a lot of what we’d want AGI to do for us is build something like this that not only doesn’t kill us (tall order, right?) but also solves global warming and climate contamination and acts as a power & fuel grid. That and bio immortality is basically everything I personally want out of AGI. So I’d really like to have some idea how to build a machine that teaches a plant to do something like a safe, human-compatible version of this.
Some good news, though: I suspect it may be more practical to defend against this sort of attack using finite intelligence than previously assumed. We need to make the machine that knows how to guard against these sorts of things, but if we can make the vulnerability-closer, we don’t need to hit max ASI to stop other ASIs from destroying all pre-ASI life on earth.
If you read between the lines in my Human level AI can plausibly take over the world post, hacking computers is probably the lowest difficulty “take over the world” strategy and has the side benefit of giving control over all the internet connected AI clusters.
The easiest way to keep a new superintelligence from emerging is to seize control of the computers it would be trained on. The AI only needs to hack far enough to monitor AI researchers and AI training clusters and sabotage later AI runs in a non-suspicious way. It’s entirely plausible this has already happened and we are either in the clear or completely screwed depending on the alignment of the AI that won the race.
Also, hacking computers and writing software is something easy to test and therefore easy to train. I doubt that training an LLM to be a better hacker/coder is much harder than what’s already been done in the RL space by OpenAI and Deepmind (EG: playing DOTA and Starcraft).
Biotech is a lot harder to deal with since ground truth is less accessible. This can be true for computer security too but to a much lesser extent (EG: lack of access to chips in the latest Iphone and lack of complete understanding therof with which to develop/test attacks).
Pshh, low expectations. Mind uploading or bust!
I’ll take mind backups, but for exactly the reasons you highlight here, I don’t think we’re going to find electronics to be more efficient than microkinetic computers like biology. I’m much more interested in significant refinements to what it means to be biological. Eventually I’ll probably substrate translate over to a reversible computer but that’s probably hundreds to thousands of years out
🤔 This is actually a path to progress, right? The difficulty in alignment is figuring out what we want precisely enough that we can make an AI do it. It seems like a feasible research project to map this out for kudzugoth.
Seems convincing enough that I’m gonna make a Discord and maybe switch to this as a project. Come join me at Kudzugoth Alignment Center! … 😅 I might close again quickly if the plan turns out to be fatally flawed, but until then, here we go.
Building new organisms from scratch (synthetic biology) is an engineering problem. Fundamentally we need to build the right parts and assemble them.
Without major breakthroughs (Artificial Superintelligence) there’s no meaningful “alignment plan”, just a scientific discipline. There’s no sense in which you can really “align” an AI system to do this. The closest things would be:
building a special purpose model (EG:alphafold) useful for solving sub-problems like protein folding
teaching an LLM to say “I want to build green biotech” and associated ideas/opinions.
which is completely useless
Problem is that biology is difficult to mess with. DNA sequencing is somewhat cumbersome, DNA writing is much more so, costing on the order of 25¢/base currently.
Also imaging the parts to figure out what they do and if they’re doing it can be very cumbersome because they’re too small to see with a light microscope. Everything is indirect. Currently we try to crystalize them and then use X-rays (which are small enough but also very destructive) to image the crystal and infer the structure. There’s continuous progress here but it’s slow.
AI techniques can be applied to some of these problems (EG:inferring protein structure from amino acids (Alphafold), or doing better quantum level simulation Ferminet)
Note that AI techniques are replacing existing ones based on human coded algorithms rooted in physics and often have issues with out of distribution inputs (EG: work well for wildtype protein but give garbage when mutations are added.)
Like any ML system, we just have to feed it more data which means we need to do more wet lab work, x-ray crystallography etc.
Synthetic biology is the best way forwards but it’s a giant scientific/engineering discipline, not an “alignment approach” whatever that’s supposed to mean.
Do you expect humanity to bioengineer this before we develop artificial superintelligence? If not, presumably this objection is irrelevant.
Basically if artificial superintelligence happens before sufficiently advanced synthetic biology, then one way to frame the alignment problem is “how do we make an ASI create a nice kudzugoth instead of a bad kudzugoth?”.
I guess but that’s not minimal and doesn’t add much.
“how do we make an ASI create a nice (highly advanced technology) instead of a bad (same)?”.
IE: kudzugoth vs robots vs (self propagating change to basic physics)
Put differently:
If we build a thing that can make highly advanced technology, make it help rather than kill us with that technology.
Neat biotech is one such technology but not a special case.
Aligning the AI is a problem mostly independent of what the AI is doing (unless you’re building special purpose non AGI models as mentioned above)
I agree that one could do something similar with other tech than neat biotech, but I don’t think this proves that Kudzugoth Alignment is as difficult as general alignment. I think aligning AI to achieve something specific is likely to be a lot easier than aligning AI in general. It’s questionable whether the latter is even possible and unclear what it means to achieve it.