Posts
Comments
I think PauseAI would be more effective if it could mobilise people who aren't currently associated with AI safety, but from what I can see it largely draws from the same base as EA. It is important to involve as wide a section of society as possible in the x-risk conversation and activism could help achieve this.
The most likely scenario by far is that a mirrored bacteria would be outcompeted by other bacteria and killed by achiral defenses due to [examples of ecological factors]
I think this is the crux of the different feelings around this paper. There are a lot of unknowns here. The paper does a good job of acknowledging this and (imo) it justifies a precautionary approach, but I think the breadth of uncertainty is difficult to communicate in e.g. policy briefs or newspaper articles.
It's a good connection to draw - I wonder if increased awareness about AI is sparking increased awareness of safety concepts in related fields. It's a particularly good sign for awareness and action of the safety concepts present in the overlap between AI and biotechnology.
I think you're right that there's very little benefit compared to the risks for mirror life which is not seen as true with AI - on top of the general truth that biotech is harder to monetise.
Can you explain more about why you think [AGI requires] a shared feature of mammals and not, say, humans or other particular species?
It's very field-dependent. In ecology & evolution, advisor-student fit is very influential and most programmes are direct admit to a certain professor. The weighting seems different for CS programs, many of which make you choose an advisor after admission (my knowledge is weaker here).
In the UK it's more funding dependent - grant-funded PhDs are almost entirely dependent on the advisor's opinion, whereas DTPs/CDTs have different selection criteria and are (imo) more grades-focused.
From discussing AI politics with the general public [i.e. not experts], it seems that the public perception of AI progress is bifurcating on two parallel lines:
A) Current AI progress is sudden and warrants a response (either acceleration or regulation)
B) Current AI progress is a flash-in-the-pan or a nothingburger.
(This is independent from responding to hypothetical AI-in-concept.)
These perspectives are largely factual rather than ideological. In conversation, the active tension between these two incompatible perspectives is really obvious. It makes it hard to hold meaningful conversations without being overbearing or ?accusatory.
Where does this divide come from? Is it the image hangover from the public's interaction with the first ChatGPT? How can we bridge this when speaking to the average person?
It is worth noting that UKRI is in the process of changing their language to Doctoral Landscape Awards (replacing DTP) and Doctoral Focal Awards (CDT). The announcements for BBSRC and NERC have already been done, but I can't find what EPSRC is doing.
I agree that evolutionary arguments are frequently confused and oversimplified, but your argument is proving too much.
[the difference between] AI and genetic code is that genetic code has way less ability to error-correct than basically all AI code, and it's in a weird spot of reliability where random mutations are frequent enough to drive evolution, but not so frequent as to cause organisms to outright collapse within seconds or minutes.
This "weird spot of reliability" is itself an evolved trait, and even with the effects of mutation rate variation between species, the variation within populations is heavily constrained (see Lewontin's paradox of diversity). Even discounting purely genetic/code-based(?) factors, the amount of plasticity (?search) in behaviour is also an evolvable trait (see canalisation) - I think it's likely there are already terms for this within the AI field but it's not obvious to me how best to link the two ideas together. I'm more curious about the value drift evolutionary arguments but I don't see an a priori reason that these ideas don't apply.
It would be good if we could understand the conditions under which greater plasticity/evolvability is selected for, and whether we expect its effects to occur in a timeframe relevant to near-term alignment/safety.
Another reason is that effective AI architectures can't go through simulated evolution, since that would use up too much compute for training to work (We forget that evolution had at a lower bound 10e46 FLOPs to 10e48 FLOPs to get to humans).
It's not obvious to me that this is a sharp lower-bound, particularly when AI are already receiving the benefits of prior human computation in the form of culture. Human evolution had to achieve the hard part of reifying the world into semantic objects whereas AI has a major head-start. If language is the key idea (as some have argued), then I think there's a decent chance that the lower bound is smaller than this.
There's a connection to the idea of irony poisoning here, and I do not think it is good for the person in question to pretend to hold extremist views. This is a parallel issue with the fact that it's terrible optics and creates a difficult tension with this website's newfound interest in doing communications/policy/outreach work.
Currently I'm not convinced that the memetic analogy has done more to clarify than to occlude cultural evolution/opinion dynamics. That's not to say that work in genetics is useless, but I think that the terminology has taken precedence above what the actual concepts mean, and I read a lot of conversations that feel like people just trading the information that they read The Selfish Gene 40 years ago.
There's certainly scope for an applied "memetics" but it's really crying out for a good predictive (even if simplistic) model.
I've noticed they perform much better on graduate-level ecology/evolution questions (in a qualitative sense - they provide answers that are more 'full' as well as technically accurate). I think translating that into a "usefulness" metric is always going to be difficult though.
I would have found it helpful in your report for there to be a ROSES-type diagram or other flowchart showing the steps in your paper collation. This would bring it closer in line with other scoping reviews and would have made it easier to understand your methodology.
Linguistic Drift, Neuralese, and Steganography
In this section you use these terms implying there's a body of research underneath these terms. I'm very interested in understanding this behaviour but I wasn't aware it was being measured. Is anyone currently working on models of linguistic drift/measuring it with manuscripts you could link?
My impression is that's a little simplistic, but I also don't have the best knowledge of the market outside WGS/WES and related tools. That particular market is a bloodbath. Maybe there's better scope in proteomics/metabolomics/stuff I know nothing about.
My impression is that much of this style of innovation is happening inside research institutes and then diffusing outward. There are plenty of people doing "boring" infrastructure work at the Sanger Institute, EMBL-EBI, etc. And you all get it for free! I can however see that on-demand services for biotech are a little different.
This fail-state is particularly worrying to me, although it is not obvious whether there is enough time for such an effect to actually intervene on the future outcome.
Are you aware of anyone else working on the same topic?
I was reading the UK National Risk Register earlier today and thinking about this. Notable to me that the top-level disaster severity has a very low cap of ~thousands of casualties, or billions of economic loss. Although it does note in the register that AI is a chronic risk that is being managed under a new framework (that I can't find precedent for).
I do think this comes back to the messages in On Green and also why the post went down like a cup of cold sick - rationality is about winning. Obviously nobody on LW wants to "win" in the sense you describe, but more winning over more harmony on the margin, I think.
The future will probably contain less of the way of life I value (or something entirely orthogonal), but then that's the nature of things.
Your general argument rings true to my ears - except the part about AI safety. It is very hard to interact with AI safety without entering the x-risk sphere, as shown by this piece of research by the Cosmos Institute, where the x-risk sphere is almost 2/3rds total funding (I have some doubts about the accounting). Your argument about Mustafa Suleyman strikes me as a "just-so" story - I do wish it were replicable, but I would be surprised, particularly with AI safety's sense of urgency.
I'm here because truly there is no better place, and I mean that in both a praiseworthy and an upsetting sense. If you think it's misguided, we, on the same side, need to show the strength of our alternative, don't we?
I think you're getting at something fairly close to the Piranha theorem from a different (ecological?) angle.
Advice for journalists was a bit more polemic which I think naturally leads to more engagement. But I'd like to say that I strongly upvoted the mapping discussions post and played around with the site quite a bit when it was first posted - it's really valuable to me.
Karma's a bit of a blunt tool - yes I think it's good to have posts with broad appeal but some posts are going to be comparatively more useful to a smaller group of people, and that's OK too.
Your points are true and insightful, but you've written them in a way that won't gain much cachet here.
I wrote a similar piece to the Cosmos Institute competition, which hopefully I can share here when that is finished, and maybe we can bounce the idea off each other?
I think this effect will be more wide-spread than targeting only already-vulnerable people, and it is particularly hard to measure because the causes will be decentralised and the effects will be diffuse. I predict it being a larger problem if, in the run-up between narrow AI and ASI, we have a longer period of necessary public discourse and decision-making. If the period is very short then it doesn't matter. It may not affect many people given how much penetration AI chatbots have in the market before takeoff too.
This is not an obvious continuation of the prompt to me - maybe there are just a lot more examples of explicit refusal on the internet than there are in (e.g.) real life.
thinking maybe we owe something to our former selves, but future people probably won't think this
This is a very strong assertion. Aren't most people on this forum, when making present claims about what they would like to happen in the future, trying to form this contract? (This comes back to the value lock-in debate.)
Is there a reason to expect this kind of behaviour to appear from base models with no fine-tuning?
Unfortunately you did nerdsnipe me with the 'biologists think' statement so I am forced to keep replying!
It's worth noting that the original derivations of natural selection do use absolute fitness - relative fitness is simply a reparameterization when you have constant N (source: any population genetics textbook). This was why I brought up density-dependent selection, as under that framework N (and s) is changing, and selection in those circumstances is more complicated.
In fact, even under typical models, relative fitness and absolute fitness show interesting relations. See this paper by Orr where alleles (which in this model only affect relative fitness by increasing absolute fitness) show diminishing returns on excess fecundity. The first paper I sent you also explicitly says [that absolute fitness is required under N-varying s or T-varying N] in the abstract.
I thought you were making a more subtle point about the additional demands of life history theory vs. the pure allele-eye view, which I agree is interesting. I hope I have convinced you that biologists are already doing fruitful work in this area. I don't understand the mesa-optimization arguments well enough to tell whether such an analogy is useful (to people who work in AI), but I do think it is true in at least a trivial sense.
I don't think contemporary theory has ignored this - see recent theories of density-dependent selection here: (article making the same point), (review). The fundamental issue you're hinging on is that absolute population growth (most effective exploitation of resources) is an ecological concept, not an evolutionary one, and population ecology theory is less well-known outside its field than population genetic theory.
My understanding was the typical explanation was antagonistic pleiotropy, but I don't know whether that's the consensus view.
This seems to have the name 'pathogen control hypothesis' in the literature - see review. I think it has all the hallmarks of a good predictive hypothesis, but I'd really want to see some simulations of which parameter scenarios induce selection this way.
I first learned about Arcadia from https://dynamicecology.wordpress.com/ blog as a "evolutionary biology" startup. When I looked they only had their fungal capsid negative result published.
I'm quite optimistic about the potential for data mining from phylogenomic inference, but I wouldn't have described any of their current projects as "blue-sky" or "high variance" like mentioned in the post. I'm not sure that generating data, competing with large government-funded research hubs, is effective. Maybe there's scope away from human health research areas which are overrepresented, but that's also probably the most directly marketable area.
Does Altos Labs come under this umbrella? They seem successful and/or very good at marketing.
There's an implied assumption that when you lose parts of society through a bottleneck that you can always recreate them with high fidelity. It seems plausible that some bottleneck events could "limit humanity's potential", since choices may rely on those lost values, and not all choices are exchangeable in time. (This has connections both to the long reflection and to the rich shaping the world in their own image).
As an aside, the bottleneck paper you're referring to is pretty contentious. I personally find it unlikely that no other demographic model detects a bottleneck of >0.99 in the real data, but all of them can do it on simulated data. If such an event did occur in the modern day, the effects would be profound and serious.