Personal website: https://andrewtmckenzie.com/
Andy_McKenzie
Transcript of Sam Altman’s interview touching on AI safety
What Makes My Attempt Special?
Update on the Brain Preservation Foundation Prize
Right. This is why I think it’s under-ratedly important for contrarians who actually believe in the potential efficacy of their beliefs to not seem like contrarians. If you truly believe that your ideas are underrepresented then you will much better promote them by being appearing generally “normal” and passing off the underrepresented idea as a fairly typical part of your ostensibly coherent worldview. I will admit that this is more challenging.
Brain preservation to prevent involuntary death: a possible cause area
One way to manipulate your level of abstraction related to a task
A review of cryonics/brain preservation in 2016
If somebody was planning to destroy the world, the rationalist could stop him and not break his oath of honesty by simply killing the psychopath. Then if the rationalist were caught and arrested and still didn’t reveal why he had committed murder, perhaps even being condemned to death for the act but never breaking his oath of honesty, now that would make an awesome movie.
Sure. Basically, there are two groups, each of which has made a major contribution:
1) Shawn Mikula and his group. They have made substantial progress (some would say, almost solved) of how to make the neuronal connections and other brain structures such as white matter tracts in a full mouse brain traceable using electron microscopy. Electron microscopy is the lowest level of imaging currently feasible, and can clearly resolve structures that are thought to be key to memory such as synapses.
2) The 21CM group, including Robert McIntyre. They have developed a totally new method of preserving a brain that should yield both highly practical and technical sound preservation. In a sense it combines the methods discussed by Gwern in his article Plastination vs Cryonics, because it first uses a method traditionally associated with “plastination” (glutaraldehyde perfusion), and then uses a method traditionally associated with cryonics, i.e. perfusion with a cryoprotective agent and then low temperature storage and, presumably, vitrification, which means that damage from ice crystal formation should be avoided and the brain should turn a glass state.
Apologies if this is still too technical and I’m happy to answer any follow-up questions. Many key steps remain but this is progress worthy of celebrating and, in my view, supporting.
Why wasn’t preservation with the goal of potential future revival started earlier in history?
Which cognitive biases should we trust in?
Cryonics p(success) estimates are only weakly associated with interest in pursuing cryonics in the LW 2023 Survey
A simple solution is to just make doctors/hospitals liable for harm which occurs under their watch, period. Do not give them an out involving performative tests which don’t actually reduce harm, or the like. If doctors/hospitals are just generally liable for harm, then they’re incentivized to actually reduce it.
Can you explain more what you actually mean by this? Do you mean if someone comes into the hospital and dies, the doctors are responsible, regardless of why they died? If you mean that we figure out whether the doctors are responsible for whether the patient died, then we get back to whether they have done everything to prevent it, and one of these things might be ordering lab tests to better figure out the diagnosis, and then it seems we’re back to the original problem i.e. the status quo. Just not understanding what you mean.
Here are all the LW comments I have bookmarked at del.icio.us/porejide, aside from the one I took from Grognor below. This is probably overkill for a single comment.
1) This comment by Mitchell Porter, for the idea of looking into the time before there was algebra, seeing how it was invented, and using that as an outside view for our current difficult problems (like consciousness).
2) This comment by Eliezer, which I found interesting because it responded to a critique of Bayes Cosmo Shalizi which I also found persuasive, leaving me in a (typical) state of not knowing what to believe.
3) Yvain’s comment preceding his post on slippery slopes and Schelling points; this was useful for my thinking about both of those topics.
4) PO8′s comment about early detection of cancer and whether the purported benefits could be a selection bias. This is an interesting idea and I intend to look into it further when I get more time.
5) Luke M’s comment that the best way to convert someone is to be cool, likeable, and generally sarcastic about the idea you want them to change. Makes sense—most do not respond well to wordy arguments.
6) Vladimir M’s comment that non-mathy pop-physics is unlikely to lead to real insight. I tentatively agree but would like to see some actual data.
7) Carl Shulman’s comment on how a normal prior distribution for charity effectiveness does not map well to reality. This doubles as a demonstration of how difficult Bayesian computations can be and an interesting quantitative look at charity.
8) Yvain’s comment, which I feel bad re-posting because it is ironic and I’m growing increasingly annoyed with irony, but I include for completeness and because it makes a useful point through its irony.
9) Eugine Nier’s comment that it’s more important for your beliefs to be correct than consistent. He also gives an example of a situation in which there can be a trade-off between the two. I found this useful because I often am biased towards consistency. I made an Anki flashcard based on this comment.
10) jimrandomh’s comment that “drugs” are not a natural category. Useful on both the object (people talk about “drugs” often) and meta (people talk about non-natural categories as if they were natural categories often) levels.
11) komponisto’s comment that “I said I was apathetic. I didn’t say I was ignorant.” I just thought this was clever. Doesn’t seem as good in hindsight. But maybe I’m biased by the relatively low upvotes.
12) Nominull’s comment that “When promoting the truth, if you value the truth, it is wise to use especially those methods that rely on the truth being true. That way, if you have accidentally misidentified the truth, there is an automatic safety valve.” I also made an Anki flashcard for this one.
13) Xachariah’s comment deconstructing the phrase “how are you” in a way I still often think about when I hear that phrase.
14) Richard Kennaway’s comment discussing the trade-offs to engaging in sexual relationships (i.e., lost time and energy for intellectual pursuits).
15) Konkvistador’s comment quoting Peter Thiel saying that as soon as you starting discussing why something occurs, people start losing sight of whether it occurs. Useful rhetorically, depending on your goals.
16) Mitchell Porter’s comment about how LW tropes might eventually find political expression, and what that would actually look like. This comment is truly a gem. “Look, politics isn’t a game of hide and seek. Ideological groups have the cohesion that they do because membership in the group depends on openly espousing the ideology.” I also made an Anki flashcard for this. This kind of comment makes me a bit sad that he appears to be spending much of his time on looking for the quantum correlates of consciousness, which does not make much sense to me. But from the perspective of “fund people, not projects”, we should give him leeway.
17) Mitchell Porter’s comment dismissing the idea of game theoretic equilibria between intelligences in disjoint worlds. I remember finding it profound when I read it.
18) taw’s comment which contains an interesting history lesson on infanticide.
19) Eliezer’s comment that you should deal with sunk costs by imagining that you were teleported into someone’s life and thinking about what you would do differently if that were the case.
20) Mitchell Porter’s comment speculating on what LW’s role in history would be. Another gem. Not only was this informative, but I hadn’t even thought on that kind of level before.
21) Will Newsome’s comment. I forget why I tagged this, but looking back it has some very interesting bits about theism.
22) hegimonicon’s comment that there is an “enormous gulf between finding out things on your own and being directed to them by a peer.” It’s somewhat counter-intuitive that in many cases your best way to convince someone of something is to suggest general lines of reasoning and let them figure out the specifics for themselves.
Agreed, and this is very similar to what I described in my comment on the other post about this here.
Where I disagree is the sole focus on connection strengths or weights. They are certainly important, but synapses are unlikely to be adequately described by just one parameter. Further, local effects like neuropeptides likely play a role.
This reminds me of a conversation from Dumb and Dumber.
Lloyd: What are the chances of a guy like you and a girl like me… ending up together? Mary: Well, that’s pretty difficult to say. Lloyd: Hit me with it! I’ve come a long way to see you, Mary. The least you can do is level with me. What are my chances? Mary: Not good. Lloyd: You mean, not good like one out of a hundred? Mary: I’d say more like one out of a million. [pause] Lloyd: So you’re telling me there’s a chance.
Good post.
Thanks for writing this up as a shorter summary Rob. Thanks also for engaging with people who disagree with you over the years.
Here’s my main area of disagreement:
General intelligence is very powerful, and once we can build it at all, STEM-capable artificial general intelligence (AGI) is likely to vastly outperform human intelligence immediately (or very quickly).
I don’t think this is likely to be true. Perhaps it is true of some cognitive architectures, but not for the connectionist architectures that are the only known examples of human-like AI intelligence and that are clearly the top AIs available today. In these cases, I expect human-level AI capabilities to grow to the point that they will vastly outperform humans much more slowly than immediately or “very quickly”. This is basically the AI foom argument.
And I think all of your other points are dependent on this one. Because if this is not true, then humanity will have time to iteratively deal with the problems that emerge, as we have in the past with all other technologies.
My reasoning for not expecting ultra-rapid takeoff speeds is that I don’t view connectionist intelligence as having a sort of “secret sauce”, that once it is found, can unlock all sorts of other things. I think it is the sort of thing that will increase in a plodding way over time, depending on scaling and other similar inputs that cannot be increased immediately.
In the absence of some sort of “secret sauce”, which seems necessary for sharp left turns and other such scenarios, I view AI capabilities growth as likely to follow the same trends as other historical growth trends. In the case of a hypothetical AI at a human intelligence level, it would face constraints on its resources allowing it to improve, such as bandwidth, capital, skills, private knowledge, energy, space, robotic manipulation capabilities, material inputs, cooling requirements, legal and regulatory barriers, social acceptance, cybersecurity concerns, competition with humans and other AIs, and of course value maintenance concerns (i.e. it would have its own alignment problem to solve).
I guess if you are also taking those constraints into consideration, then it is really just a probabilistic feeling about how much those constraints will slow down AI growth. To me, those constraints each seem massive, and getting around all of them within hours or days would be nearly impossible, no matter how intelligent the AI was.
As a result, rather than indefinite and immediate exponential growth, I expect real-world AI growth to follow a series of sigmoidal curves, each eventually plateauing before different types of growth curves take over to increase capabilities based on different input resources (with all of this overlapping).
One area of uncertainty: I am concerned about there being a spectrum of takeoff speeds, from slow to immediate. In faster takeoff speed worlds, I view there as being more risk of bad outcomes generally, such as a totalitarian state using an AI to take over the world, or even the x-risk scenarios that you describe.
This is why I favor regulations that will be helpful in slower takeoff worlds, such as requiring liability insurance, and will not cause harm by increasing take-off speed. For example, pausing AGI training runs seems likely to make takeoff speed more discontinuous, due to creating hardware, algorithmic, and digital autonomous agent overhangs, thereby making the whole situation more dangerous. This is why I am opposed to it and dismayed to see so many on LW in favor of it.
I also recognize that I might be wrong about AI takeoff speeds not being fast. I am glad people are working on this, so long as they are not promoting policies that seem likely to make things more dangerous in the slower takeoff scenarios that I consider more likely.
Another area of uncertainty: I’m not sure what is going to happen long-term in a slow takeoff world. I’m confused. While I think that the scenarios you describe are not likely because they are dependent upon there being a fast takeoff and a resulting singleton AI, I find outcomes in slow takeoff worlds extraordinarily difficult to predict.
Overall I feel that AI x-risk is clearly the most likely x-risk of any in the coming years and am glad that you and others are focusing on it. My main hope for you is that you continue to be flexible in your thinking and make predictions that help you to decide if you should update your models.
Here are some predictions of mine:
Connectionist architectures will remain the dominant AI architecture in the next 10 years. Yes, they will be hooked up in larger deterministic systems, but humans will also be able to use connectionist architectures in this way, which will actually just increase competition and decrease the likelihood of ultra-rapid takeoffs.
Hardware availability will remain a constraint on AI capabilities in the next 10 years.
Robotic manipulation capabilities will remain a constraint on AI capabilities in the next 10 years.
Dick Teresi, The Undead