I disagree with Eliezer’s comments on inclusive genetic fitness (~25:30) on Dwarkesh Patel’s podcast—particularly his thought experiment of replacing DNA with some other substrate to make you healthier, smarter, and happier.
Eliezer claims that evolution is a process optimizing for inclusive genetic fitness, (IGF). He explains that human agents, evolved with impulses and values that correlate with but are not identical to IGF, tend to escape evolution’s constraints and satisfy those impulses directly: they adopt kids, they use contraception, they fail to take maximum advantage of their ability to become a sperm donor, and so on. Smart people would even be willing to have their kid’s genes replaced with something other than DNA, and as long as the kid would straightforwaredly benefit, they’d be fine with it.
Once we get to the point of this thought experiment, though, I do not think it is correct to say that evolution is optimizing for IGF. Just as the body is merely the vehicle for the gene, the gene is merely the vehicle for information. Evolution is optimizing for self-replicating information, and DNA just happens to be a convenient molecule for this purpose. And evolution isn’t adaptive or forward-looking. The optimization process it generates is purely mechanical in nature, like a rock rolling downhill.
If we could keep our survival-and-replication information the same, while replacing our DNA with some other storage mechanism, that would be perfectly in accord with the optimization process evolution is running. Eliezer seems to think that a total rewrite to the DNA would be losing information. I would argue that if the vehicle is still able to perform and replicate as effectively as before, then the information relevant to evolution has been perfectly preserved—only the genetic sequence is lost, and that’s the mere vehicle for the replication-relevant information. Evolution doesn’t care about the DNA sequence. It acts on the replication-relevant information.
Does that matter for the AI safety argument?
I think it shows the limits of evolution as a metaphore for this topic. The key issue with agentic intelligence is that it is adaptive and forward-looking, in a way that natural selection simply is not (even though natural selection can be steered by the agentic processes that it has given rise to, and the rise of intelligence is a consequence of natural selection). So we can’t say “we broke free of evolution, therefore AI can break free of us.” We are agentic, foreward-thinking, and so we can adaptively anticipate what AI will do and pre-empt any such breakaway in a way that’s simply beyond the capacity of evolution. Just because we can doesn’t mean we will, but we should.
Ultimately, then I think we have to look at this aspect of the empirical result of evolution not so much as a piece of evidence about what will happen with AI alignment, as an intuition-building metaphore. The empirical result that is actually informative about the results of agency is the way we’ve taken control over the world and killed off our competitors. Whatever information can self-replicate most effectively—whether contained in human DNA, in a computer virus, or in an AI agent—will do so, up until it runs into a hard limit in its ability to do so or gets outcompeted by some other even more effectively-replicating piece of information. Or, in other words, evolution will be perfectly happy to continue optimizing for the reproductive fitness of information that uses computer code as the gene and machines as the vehicle.
I disagree with Eliezer’s comments on inclusive genetic fitness (~25:30) on Dwarkesh Patel’s podcast—particularly his thought experiment of replacing DNA with some other substrate to make you healthier, smarter, and happier.
Eliezer claims that evolution is a process optimizing for inclusive genetic fitness, (IGF). He explains that human agents, evolved with impulses and values that correlate with but are not identical to IGF, tend to escape evolution’s constraints and satisfy those impulses directly: they adopt kids, they use contraception, they fail to take maximum advantage of their ability to become a sperm donor, and so on. Smart people would even be willing to have their kid’s genes replaced with something other than DNA, and as long as the kid would straightforwaredly benefit, they’d be fine with it.
Once we get to the point of this thought experiment, though, I do not think it is correct to say that evolution is optimizing for IGF. Just as the body is merely the vehicle for the gene, the gene is merely the vehicle for information. Evolution is optimizing for self-replicating information, and DNA just happens to be a convenient molecule for this purpose. And evolution isn’t adaptive or forward-looking. The optimization process it generates is purely mechanical in nature, like a rock rolling downhill.
If we could keep our survival-and-replication information the same, while replacing our DNA with some other storage mechanism, that would be perfectly in accord with the optimization process evolution is running. Eliezer seems to think that a total rewrite to the DNA would be losing information. I would argue that if the vehicle is still able to perform and replicate as effectively as before, then the information relevant to evolution has been perfectly preserved—only the genetic sequence is lost, and that’s the mere vehicle for the replication-relevant information. Evolution doesn’t care about the DNA sequence. It acts on the replication-relevant information.
Does that matter for the AI safety argument?
I think it shows the limits of evolution as a metaphore for this topic. The key issue with agentic intelligence is that it is adaptive and forward-looking, in a way that natural selection simply is not (even though natural selection can be steered by the agentic processes that it has given rise to, and the rise of intelligence is a consequence of natural selection). So we can’t say “we broke free of evolution, therefore AI can break free of us.” We are agentic, foreward-thinking, and so we can adaptively anticipate what AI will do and pre-empt any such breakaway in a way that’s simply beyond the capacity of evolution. Just because we can doesn’t mean we will, but we should.
Ultimately, then I think we have to look at this aspect of the empirical result of evolution not so much as a piece of evidence about what will happen with AI alignment, as an intuition-building metaphore. The empirical result that is actually informative about the results of agency is the way we’ve taken control over the world and killed off our competitors. Whatever information can self-replicate most effectively—whether contained in human DNA, in a computer virus, or in an AI agent—will do so, up until it runs into a hard limit in its ability to do so or gets outcompeted by some other even more effectively-replicating piece of information. Or, in other words, evolution will be perfectly happy to continue optimizing for the reproductive fitness of information that uses computer code as the gene and machines as the vehicle.