Aubrey de Grey on Aging and AI
No one has done more to bring the viability of ending aging to the attention of the general public and the scientific community than Aubrey de Grey. His specific plan for working toward this critical goal, which goes by the name of SENS (Strategies for Engineering Negligible Senescence), is currently the subject of research in a number of biology labs. His book Ending Aging (co-authored with Michael Rae) is required reading for anyone interested in longevity or transhumanism; and his website sens.org is full of relevant information.
Aubrey’s talk and his and back-and-forth with Michael Rose at the Humanity+ CalTech conference last year were among the highlights of the conference. Some of the key issues of contention between Aubrey and Michael were discussed in a previous H+ magazine article. Aubrey will also be keynoting at our upcoming Humanity+ @ Hong Kong conference, which will provide an opportunity to build more bridges between the SENS perspective and the large amount of research on stem cells and other anti-aging strategies currently occurring in China.
Aubrey and I have always had a particular rapport, due to our common interest in both AGI and life extension research – he was an AI researcher before he embarked on his current phase of focus on life extension. However, we have taken significantly different approaches in our work, in both AI and biology – which means we always have a lot to talk about! In this brief and somewhat unsystematic dialogue, my goal was simply to gather a few of Aubrey’s opinions on some longevity research issues touching my current interest. Hopefully they will touch your interest as well!
As well as the H+ magazine article linked above, this dialogue include some quotes from my earlier H+ magazine interview with Joao Pedro de Magalhaes, and my earlier article on AI Superflies, and the Path to Immortality.
Jumping right into the middle of the discussion we were having at the Humanity+ @ Caltech conference — your view seems to be that accumulated damage, rather than (as some others like Michael Rose would have it) antagonistic pleiotropy is the main culprit in human aging. Could you briefly summarize your reasons for this?
Aubrey de Grey:
Well, I’m not sure that we can really say that accumulation of damage and antagonistic pleiotropy (AP) are alternatives — they are answers to different questions. Damage accumulation is a mechanistic hypothesis for how aging occurs, and AP is an evolutionary hypothesis for why it occurs. There are certainly some types of damage accumulation with aging that are caused as side-effects of machinery that is useful in early life — an example would be the accumulation of potentially toxic senescent cells that have arrested as a way to stop them from becoming cancerous – and that’s basically all that the AP concept proposes. I don’t think Michael thinks that aging is a maladaptive continuation of development, or some other “programmed” process — I think he agrees with me and most other gerontologists that aging is caused by damage accumulation. The only question is how that damage accumulation changes with age.
Ben: When I posed a similar question to Joao Pedro de Magalhaes http://pcwww.liv.ac.uk/~aging/, he said:
Accumulated damage can be considered a major cause of aging, yet “damage” is such a broadly defined term that I’m not satisfied with this concept. Almost any detrimental molecular or cellular event in the body can be defined as “damage”, and many (perhaps most) forms of “damage” in the body do not cause aging. So what we need to find out is what are the specific molecular and cellular processes that drive aging, and unfortunately we don’t know yet what these are.
I don’t agree with this. First, JP is putting the cart before the horse in terms of the definition of “damage” – yes, sure, there are lots of possible definitions, but I’ve always been clear about mine:
- side-effect of metabolism
- once abundant enough, may contribute to age-related ill-health
Second, because of the third criterion above, I don’t agree that we need to know for sure which accumulating side-effects of metabolism contribute to age-related ill-health and which do not. I think we should just go after repairing all those that might so contribute. If we fix a few things we didn’t need to fix, no big deal, whereas if we waste time on further analysis of which things to pursue and which not to, we delay the overall outcome.
Joao Pedro also says:
The developmental theory of aging (which I advocate) can be seen as a form of antagonistic pleiotropy, of genes or mechanisms beneficial early in life (in this case for growth and development) being detrimental late in life and contributing to aging. The idea is very simple: some of the same genetically-regulated processes that are crucial in development continue throughout adulthood and become harmful. For example, far-sightedness is thought to derive from the continual growth of eye lenses. During development the eye lenses must grow but it seems that the genetic program determining their growth continues in adulthood and this contributes to a specific age-related disease. My hypothesis is that several processes programmed in the genome for important roles during development become harmful late in life and contribute to aging. In an indirect sense, it’s a form of programmed aging processes.
I’ve discussed Joao Pedro’s perspective with Michael and he’s largely sympathetic though he’s not sure development-related antagonistic pleiotropy plays quite as large a role as Joao Pedro thinks, relative to other sorts of antagonistic pleiotropy….
Any reaction to these notions as articulated in the above paragraphs?
I agree with Michael that the few such lifelong processes that exist are peripheral to aging. I don’t even agree with JP’s example: actually, far-sightedness is mainly caused by glycation-derived crosslinking.
However, I don’t know what you/Michael may mean by “other sorts of antagonistic pleiotropy”. Maybe you mean the sort I mentioned – accumulating damage from aspects of metabolism that exist to prevent the accumulation of other sorts of damage earlier in life — but maybe you don’t.
Pursuing the same train of thought a little further —
Michael Rose presents an argument in favor of a “late life” phase in fruit flies and other animals including humans. He argues that during “late life” the odds of the organism dying during any given year becomes essentially constant. But this seems at odds with the idea that accumulating damage plays a large role in aging (because it seems that damage would just keep progressively accumulating as the organism gets older, rather than stopping to accumulate when “late life” occurs). I know this is a technical matter, but it’s an important one, so could you try your best to summarize your views on “late life” in a nontechnical way?
You’ve got it. I think it is highly implausible that an individual’s accumulation of damage will even cease to accelerate, let alone slow down or stop, after some age. It’s not completely impossible that that could happen — that after some point the individual could, for example, shift to a different lifestyle (like a fly stopping flying) that would drastically reduce the rate of damage accumulation — but I don’t buy it. Michael adopted this view as a result of what I have shown was a premature and oversimplistic mathematical analysis of the distribution of ages at death of his flies, an analysis that seemed to show that damage accumulation must indeed slow down. I’ve shown that the data are in fact entirely compatible with a lifelong acceleration in damage accumulation, given a modest degree of variation between individual flies in the initial rate of damage accumulation and (more importantly) the rate at whch that rate accelerates via positive feedback. Thus, while I don’t claim that Michael’s interpretation can be excluded on the basis of existing data, I prefer the biologically more natural alternative that damage accumulation does indeed accelerate throughout life.
On a different note — I wonder how much do you think progress toward ending aging would be accelerated if we had an AGI system that was, let’s say, roughly as generally intelligent as a great human scientist, but also had the capability to ingest the totality of biological datasets into its working memory and analyze them using a combination of human-like creative thought and statistical and machine learning algorithms? Do you think with this sort of mind working on the problem, we could reach the Methuselarity in 5 or 10 years? Or do you think we’re held back by factors that this amazing (but not godlike) level of intelligence couldn’t dramatically ameliorate?
I think it’s highly unlikely that such a system could solve aging that fast just by analysing existing knowledge really well; I think it would need to be able to do experiments, to find things out that nobody knows yet. For example, it’s pretty clear that we will need much more effective somatic gene therapy than currently exists, and I think that will need a lot of trial and error. However, I’m all for development of such a system for this purpose: firstly I might be wrong about the above, and secondly, even if it only hastens the Methuselarity by a small amount, that’s still a lot of lives saved.
Yeah, I think that’s probably right. But of course, a sufficiently advanced AGI system could also design new experiments and have human lab techs run them — or it could run the experiments itself using robotized lab equipment.
Right. I was explicitly excluding that scenario, because you seemed to be.
There have already been some simple experiments with completely robotized labs, where the experiments themselves are specified by AI algorithms, so there’s a fully automated cycle from AI experiment specification, to robotized experiment execution, to ML data analysis, back to AI experiment specification, etc. But of course in the cases done so far, the experiments have been “templatized” rather than requiring creative, improvisatory experimental design.
I suppose the relevance of this sort of approach is relative to one’s judgment of the relative timing of progress in AGI versus non-AGI-based life extension R&D. If you think we can make a Methuselarity by 2030 without AGI, but that developing AGI capable of providing dramatic help to life extension research will take till 2040, then AGI will be low priority for you (insofar as your goal is life extension). But if you think it will take till 2060 to make a Methuselarity without AGI, whereas AGI capable of dramatically helping life extension research could likely be created by 2030 — then it will make sense for you to advocate a lot of resources going into AGI, even if your ultimate main focus is life extension.
Right — except that of course AGI research is much cheaper, so there’s really no reason to prioritise one over the other.
So it would appear to me, based on your comments, that you think the timeline to “Methuselarity without AGI help” is probably sooner than the timeline to “advanced AGI that could dramatically accelerate life extension research”. My question, at too long last, is: Is this a correct inference of mine, regarding your subjective estimates of these two timelines?
Not really, no. I honestly don’t have a feel for how far away AGI of the sort that could really help is. It may be impossible (the Monica Anderson view). It may be rather easy (your view). It may be reasonably easy to create but really hard to make safe (Eliezer’s view). I honestly haven’t a clue.
Another factor that occurs to me, when discussing AGI and life-extension research, is the relative cost of the two types of research. If I’m right about AGI, the cost to achieve it could be tens of millions … whereas non AGI based life extension research could very easily cost billions. So the question is: What’s your current rough cost estimate for achieving Methuselarity? Are you still thinking of it as a multi-billion dollar endeavor? (Which of course is quite affordable by society, given the massive amount we spend on healthcare…)
Definitely multi-billion, yes — probably trillions, if we include all the medical delivery infrastructure and manpower. But the question then arises of whether AGI will cut the cost of those later stages as well as of the early stages. I don’t really see why it shouldn’t…