Epistemic status: Speculation, I am not confident in my views here and I skim over a lot of issues.
TL;DR: I claim that accelerating the development of "Whole Brain Emulation", such that it arrives before human level AI, is an intervention that would reduce the risks posed by AI.
Normal AI development consists of making AI systems more generally capable, interesting or useful. By "differential technological development" with regard to AI, people usually mean instead making technical progress on systems that are:
- Aligned with programmer goals (or at least corrigible)
I think there are broader ways in which technology could be developed differentially. One of these in particular could conceivably put us in a better position with AI risk.
Whole brain emulation is the idea of:
By doing this properly, you would have a digital person who would have many advantages over a biological person. Particularly relevant to this post are the ways digital people would be more powerful/productive:
- They can be freely copied
- They can be run at greater or lesser speeds
- Total control over their virtual environment give them greater productivity
- Changing their brain to improve it is easier in a digital environment than in biology
- Potentially, digital people could have a high quality of life on much lower wages than biological people, so their labour could be cheaper
"True" AI is dangerous because it is powerful and potentially lacks human values.
Digital people are more powerful than biological humans, but retain human values. We can trust their judgement as much as we trust any humans.
I claim that a world with substantial numbers of digital people would have lower risk from "true" AI than our current world. I have two primary arguments for this, and a couple of secondary arguments.
Arguments in favour
It's easier to have a digital person in the loop
Digital humans would be much cheaper to query than biological humans. This is because:
- They run at faster speeds
- Skilled people can be copied to deal with many parallel queries
- Digital people have may have lower wages than biological people
This means that several aspects of safe(r) AI development become cheaper:
- When deployed, AIs could refer to humans before performing actions - humans can be "in the loop" for more decisions
- The threshold for investigating strange behaviour can be lowered in systems being tested. We can tolerate more false positives
- Training can be overseen by humans, and training results scrutinised by more humans.
We could in theory do a lot of these things with biological humans too, but it is much more expensive. So digital people make safe(r) development of AI cheaper relative to our current world.
Digital people have many advantages over biological people. AIs have all the advantages of digital people, plus others (largely consisting of being optimised more strongly for productive tasks and eventually being more generally intelligent than humans).
A world with digital people is at less of a disadvantage to AI systems in power terms than our current world. All else being equal, having less of a power disparity seems likely to lead to better outcomes for humans.
AI less useful -> less AI development
In a world with digital people, AI is somewhat less useful. Many of the biggest advantages of AI systems are also possessed by digital people. Therefore, there is less incentive to invest in creating AI, which might mean it takes longer. I will assume for this argument that slowing down AI development makes its eventual development safer.
A world with digital people is used to policing compute and dealing with hostile compute based agents
Like biological people, digital people will not always behave nicely. The world will worry more about criminals or spies who are able to think very quickly and in parallel about how to break security systems, commit crimes, copy themselves to unauthorised places etc. And such a world will probably develop more countermeasures against such criminals than we have today (for example by having digital people available to police security systems at high speed).
The world could also be more used to policing the use of compute resources themselves. In theory, digital people could be enslaved or cheated out of useful information/trade secrets in a private server. Many digital people would worry about this happening to them, and there could be a push to ensure that servers with enough compute to run digital people are well policed.
These two factors could plausibly lead to a world where large computer systems are more secure and better policed, which lowers the risk posed by AI.
Of course, there are arguments against this view!
Digital people lead to AI
Technologies required to make digital people may directly advance AI. For example, if we understand more about neuroscience and the human brain, that may lead to the discovery of better algorithms to be used in the development of AI. Some AI research is already partially informed by biology.
Digital people require a lot of compute resources, and so does AI. If creating digital people led to us creating lots more compute, this could be bad as well.
Growth would be much faster in a world with many digital people. Even if we grant that there is less interest in AI as a % of the economy, there are more total resources going towards it - so AI development could be faster in such a world.
Relatedly, there is still considerable incentive to develop AI quickly. Digital people can't be used in the same ways for ethical & practical reasons (i.e. they need breaks and are not optimised for the kind of work the marketplace demands)
Digital people may bring extremely rapid social change, leaving us in a worse position politically to deal with AI than if we had remained more stable.
What does this imply?
In my view, a world with digital people is a world with a lower risk from AI compared to a world without them.
If this view were shared by others, we could attempt to speed up development of the necessary technologies for digital people to be created. Progress on creating digital worms hasn't been going so great, but this seems at least partially due to a lack of funding and effort. Someone could setup/fund an organisation to fund research in the necessary technologies and speed them along.
If I were convinced of the opposite view, it's difficult to see what to do other than discouraging new funding of the relevant technologies.
Similar arguments by Carl Shulman & others here: https://www.lesswrong.com/posts/v5AJZyEY7YFthkzax/hedging-our-bets-the-case-for-pursuing-whole-brain-emulation#comments
Broader considerations by Robin Hanson here: https://www.overcomingbias.com/2011/12/hurry-or-delay-ems.html
In case you missed it, Holden's post on Digital People: https://www.cold-takes.com/how-digital-people-could-change-the-world/
I think your "digital people lead to AI" argument is spot on, and basically invalidates the entire approach. I think getting whole brain emulation working before AGI is such a longshot that the main effect of investing in it is advancing AI capabilities faster.
Strong agree, see for example my post Randal Koene on brain understanding before whole brain emulation
I definitely think it's an (the most?) important argument against. Some of this comes down to your views on timelines which I don't really want to litigate here.
I guess I don't know how much research leading to digital people is likely to advance AI capabilities. A lot of the early work was of course inspired by biology, but it seems like not much has come of it recently. And it seems to me that we can focus on the research needed to emulate the brain, and try not to understand it in too much detail.
Could we just...keep the WBE research secret? It'd make it harder, but if having detailed knowledge of how the brain works openly available is dangerous, then, it seems like if you can keep the research closed, that mitigates that risk.
You might be interested in my paper on this topic, where I also come to the conclusion that achieving WBE before de novo AI would be good:
This is great, thanks!
There's another very large disadvantage to speeding up research here—once we have digital minds, it might be fairly trivial for bad actors to create many instances of minds in states of extreme suffering (for reasons such as sadism). This seems like a dominant consideration to me, to the extent that I'd support any promising non-confrontational efforts to slow down research into WBE, despite the benefits to individuals that would come about from achieving digital immortality.
I also think digital people (especially those whose cognition is deliberately modified from that of baseline humans, to e.g. increase "power") are likely to act in unpredictable ways—because of errors in the emulation process, or the very different environment they find themselves in relative to biological humans. So digital people could actually be less trustworthy than biological people, at least in the earlier stages of their deployment.
That could happen. I would emphasise that I'm not talking about whether we should have digital minds at all, just when we get them (before or after AGI). The benefit in making AGI safer looms larger to me than the risk of bad actors - and the threat of such bad actors would lead us to police compute resources more thoroughly than we do now.
Digital people may be less predictable, especially if "enhanced", I think that the trade-off is still pretty good here in that they almost entirely approximate human values versus AI systems which (by default) do not at all.
Thanks, I found this really interesting.
The existence of digital people would force us to anthropomorphize digital intelligence. Because of that, the implications of any threats that AI may pose to us might be more comprehensively visible and more often in the foreground of AI researchers' thinking.
Maybe anthropomorphizing AI would be an effective means through which to see the threats AI poses to us because of the fact that we have posed many threats to ourselves, like through war for example.
That seems useful up to a point - I feel like many think "Well, the AI will just do what we tell it to do, right?", and remembering the many ways in which even humans cheat could help expose flaws in that thinking. On the other hand, anthropomorphizing AI too much could mean expecting them to behave in human-like ways, which itself is likely an unrealistic expectation.
I think that its utility being limited is true. It was just a first impression that occurred to me and I haven't thought it through. It seemed like anthropomorphizing AI could consistently keep people on their toes with regard to AI. An alternative way to become wary of AI would be less obvious thoughts like an AI that became a paperclip maximizer. However, growing and consistently having priors about AI that anthropomorphize them may be disadvantageous by constraining people's ability to have outside of the box suspicions (like what they already be covertly doing) and apprehensions (like them becoming paperclip maximizers) about them.
Anthropomorphizing AI could also help with thinking of AIs as moral patients. But I don't think that being human should be the sufficient standard for being a moral patient. So thinking of them as humans may just be useful insofar as they initiate thought of them as moral patients but maybe eventually an understanding of them as moral patients will involve considerations that are particular to them and not just because they are like us.
Digital people seem a long way off technologically, versus AI which could either be a long way off or right around the corner. This would argue against focusing on digital people, since there's only a small chance that we could get digital people first.
But on the other hand, the neuroscience research needed for digital people might start paying dividends far before we get to whole-brain emulation. High-bandwidth "brain-computer interfaces" might be possible long before digital people, and BCIs might also help with AI alignment in various ways. (See this LessWrong tag.) Some have also argued that neuroscience research might help us create more human-like AI systems, although I am skeptical on this point.
I agree shooting for digital people is a bad plan if timelines are short. I guess I'm not sure how short they would need to be for it not to be worth trying.
I think if we wanted to produce BCIs we should just shoot for that directly - doesn't seem like the best plan for getting to digital people is also the best plan for getting BCIs.
I think that insofar as neuroscience helps make AI, that just speeds up progress and is probably bad.
An efficient general intelligence on a biological substrate uses a brain structure. It's unclear if that same structure would be efficient on silicon or photonic processors.