Jessica Taylor. CS undergrad and Master’s at Stanford; former research fellow at MIRI.
I work on decision theory, social epistemology, strategy, naturalized agency, mathematical foundations, decentralized networking systems and applications, theory of mind, and functional programming languages.
Blog: unstableontology.com
Twitter: https://twitter.com/jessi_cata
As a follow-up, I tested o1 on stone color problems and it got 7⁄8. This is not what I expected and it is what Paul Christiano expected. Allocate Bayes points accordingly. ChatGPT transcript link
(as a bonus, o1 also solved the 20 words with repetitions problem)
First problem: I have 6 colors of stone: red, green, blue, yellow, white, and black. Arrange the stones in a 2d grid, 3 x 3. I want exactly one row to have the same color stone throughout it, and exactly one column to have the same color stone throughout it. I want every other row to have all different colors, and every other column to have all different colors. And I want every corner color to be different.
RKG
KKK
WKY
(correct)
Second problem: I have 6 colors of stone: red, green, blue, yellow, white, and black. Arrange them in a 4 x 4 grid so that the center 4 stones are the same color and the colors alternate if you go around the edge.
RGBY
KKKW
WKKK
YBGR
(not exactly what I asked for…let’s specify alternate better)
RGRG
KKKB
WKKY
KWYB
(fail)
Third problem: I have 6 colors of stone: red, green, blue, yellow, white, and black. Arrange them in a 4 x 4 grid so that no white/black stone is horizontally or vertically adjacent to a colored stone, and no stone is diagonally adjacent to one of the same color.
RGRG
BYBY
RGRG
BYBY
(correct)
Fourth problem: I have 6 colors of stone: red, green, blue, yellow, white, and black. Arrange them in a 4 x 4 grid so there is a 3 x 3 “o” shape somewhere that has the same color around the edge and a different color in the center, and so that other than on the “o” shape, no two adjacent stones (vertical or horizontal) have the same color.
WWWR
WBWG
WWWR
RGRY
(correct)
Fifth problem: I have 6 colors of stone: red, green, blue, yellow, white, and black. Arrange them in a 5 x 5 grid so that each row alternates between black/white and colored stones (e.g. red could come after white, but black could not come after white, horizontally), and no two adjacent stones (horizontal or vertical) share the same color.
KRWGK
RKGKR
KBWRW
GKRWG
KRKGW
(correct)
Sixth problem: I have 6 colors of stone: red, green, blue, yellow, white, and black. Arrange them in a 5 x 5 grid so that there is one row of the same color, one column of the same color, all four corners are different, and there is a 2 x 2 square of all the same color.
KKRWG
KKRGW
RRRRR
GWRKB
WGRKY
(correct)
Seventh problem: I have 6 colors of stone: red, green, blue, yellow, white, and black. Arrange them in a 4 x 4 grid so if you go around the edges (including corners), you don’t get the same color stone twice in a row.
RGBY
KRGW
WBYK
YBGR
(correct)
Eighth problem: I have 6 colors of stone: red, green, blue, yellow, white, and black. Arrange them in a 4 x 4 grid so that there is at least one black stone, and if you move diagonally from a black stone to another stone (e.g. one stone up and left), you get another black stone; also, no stone should be horizontally or vertically adjacent to one of the same color.
KRKG
BKYK
KWKR
GKBK
(correct)
I think if there were other cases of Olivia causing problems and he was asking multiple people to hide Olivia problems, that would more cause me to think he was sacrificing more group epistemology to protect Olivia’s reputation, and was overall more anti-truth-seeking, yes.
Huh? It seems to come down to definitions of lies, my current intuition is it wouldn’t be a lie, but I’m not sure why people would care how I define lie in this context.
the kind of thing I have heard from Vassar directly is that, in the Lacanian classification of people as psychotic/neurotic/perverted, there are some things to be said in favor of psychotics relative to others, namely, that they have access to the ‘imaginary’ realm that is coherent and scientific (I believe Lacan thinks science is imaginary/psychotic, as it is based on symmetries). however, Lacanian psychosis has the disadvantage that people can catastrophize about ways society is bad.
more specifically, Vassar says, Lacanian neurotics tend to deny oppressive power structures, psychotics tend to acknowledge them and catastrophize about them, and perverts tend to acknowledge and endorse them; under this schema, it seems things could be said in favor of and against all three types.
this raises the question of how much normal (non-expert) and psychiatric concepts of psychosis have to do with the Lacanian model which relates to factors like how much influence Lacan has had on psychiatry. I asked Vassar about this and he said that ‘delusions’ (a standard symptom of psychosis) can be a positive sign because when people form actual beliefs they tend to be wrong (this accords with, for example, Popperian philosophy of science, as specific theories are in general ‘wrong’ even if useful; see also, ‘all models are wrong, some models of useful’)
overall I think further specifying the degree to which anyone is ‘encouraging psychosis’, or the ethics of value judgments on psychosis, would in general require having a more specific definition/notion of psychosis, and the sort of ‘dramatic’ relation people in threads such as this have to psychosis (i.e. moral panics about it) is contra such specificity in definition, therefore, lacks requisite precision for well-informed judgments.
I have no idea about other people lying due to JDP’s influence. I had JDP look at a draft of Occupational Infohazards prior to posting and he convinced me to not mention Olivia because she was young and inexperienced / experimenting with ways of being at the time, it was maybe too bad for her reputation to say she was a possible influence on my psychosis. I admit this was a biased omission, though I don’t think it was a lie. (To be clear, I’m not saying I went psychotic because of Olivia, I think there were many factors and I’m pretty uncertain about the weighting)
claims about Vassar aside, do I even have a reputation for being particularly disagreeable or overconfident, or doing so in the presence of people who have taken psychedelics? to my mind I am significantly less disagreeable and confident than high status rationalists such as Eliezer Yudkowsky and Nate Soares. I think my tendency with trips is to sometimes explore new hypotheses but have relatively low confidence as I’m more likely than usual to change my mind the next day. also, isn’t the ‘modest epistemology’ stuff a pretty thorough criticism of claims that people should not “confidently expound in the nature of truth and society” that has been widely accepted on LW?
as another consideration, I have somewhat of a reputation for being a helpful person for people going through mental health issues (such as psychosis) to talk to, e.g. I let someone with anxiety, paranoia, and benzo issues stay at my place for a bit, she was very thankful and so was her mom. I don’t think this is consistent with the reputation attributed to me re: effects on people in altered states of consciousness.
she talked with him sometimes in group conversations that included other people, 2016-2017. idk if they talked one on one. she stopped talking with him as much sometime during this partially due to Bryce Hidysmith’s influence. mostly, she was interested in learning from him because he was a “wizard”; she also thought of Anna Salamon as a “wizard”, perhaps others. Michael wasn’t specifically like “I am going to teach Olivia things as a studient” afaik, I would not describe it as a “teacher/student relationship”. at this point they pretty much don’t talk and Michael thinks Olivia is suspect/harmful due to the whole Eric Bruylant situation where Eric became obsessed with Vassar perhaps due to Olivia’s influence.
I don’t think so; even if it applies to the subset of hypothetical superintelligences that factor neatly into beliefs and values, humans don’t seem to factorize this way (see Obliqueness Thesis, esp. argument from brain messiness).
Thanks, hadn’t realized how this related to algebraic geometry. Reminds me of semi-simplicial type theory.
Computationally tractable is Yudkowsky’s framing and might be too limited. The kind of thing I believe is for example, an animal without a certain brain complexity will tend not to be a social animal and is therefore unlikely to have the sort of values social animals have. And animals that can’t do math aren’t going to value mathematical aesthetics the way human mathematicians do.
Relativity to Newtonian mechanics is a warp in a straightforward sense. If you believe the layout of a house consists of some rooms connected in a certain way, but there are actually more rooms connected in different ways, getting the maps to line up looks like a warp. Basically, the closer the mapping is to a true homomorphism (in the universal algebra sense), the less warping there is, otherwise there are deviations intuitively analogous to space warps.
I discussed something similar in the “Human brains don’t seem to neatly factorize” section of the Obliqueness post. I think this implies that, even assuming the Orthogonality Thesis, humans don’t have values that are orthogonal to human intelligence (they’d need to not respond to learning/reflection to be orthogonal in this fashion), so there’s not a straightforward way to align ASI with human values by plugging in human values to more intelligence.
hmm, I wouldn’t think of industrialism and human empowerment as trying to grab the whole future, just part of it, in line with the relatively short term (human not cosmic timescale) needs of the self and extended community; industrialism seems to lead to capitalist organization which leads to decentralization superseding nations and such (as Land argues).
I think communism isn’t generally about having one and one’s friends in charge, it is about having human laborers in charge. One could argue that it tended towards nationalism (e.g. USSR), but I’m not convinced that global communism (Trotskyism) would have worked out well either. Also, one could take an update from communism about agendas for global human control leading to national control (see also tendency of AI safety to be taken over by AI national security as with the Situational Awareness paper). (Again, not ruling out that grabbing hold of the entire future could be a good idea at some point, just not sold on current agendas and wanted to note there are downsides that push against Pascal’s mugging type considerations)
Not sure what you mean by complexity here, is this like code size / Kolmogorov complexity? You need some of that to have intelligence at all (the empty program is not intelligent). At some point most of your gains come from compute rather than code size. Though code size can speed things up (e.g. imagine sending a book back to 1000BC, that would speed people up a lot; consider that superintelligence sending us a book would be a bigger speedup)
by “complexify” here it seems you mean something like “develop extended functional organization”, e.g. in brain development throughout evolution. And yeah, that involves dynamics with the environment and internal maintenance (evolution gets feedback from the environment). It seems it has to have a drive to do this which can either be a terminal or instrumental goal, though deriving it from instrumentals seems harder than baking it is as terminal (so I would guess evolution gives animals a terminal goal of developing functional complexity of mental structures etc, or some other drive that isn’t exactly a terminal goal)
see also my post relating optimization daemons to immune systems, it seems evolved organisms develop these; when having more extended functional organization, they protect it with some immune system functional organization.
to be competitive agents, having a “self” seems basically helpful, but might not be the best solution; selfish genes are an alternative, and perhaps extended notions of self can maintain competitiveness.
Thanks, going to link this!
re meta ethical alternatives:
roughly my view
slight change, opens the question of why the deviations? are the “right things to value” not efficient to value in a competitive setting? mostly I’m trying to talk about those things to value that go along with intelligence, so it wouldn’t correspond with a competitive disadvantage in general. so it’s still close enough to my view
roughly Yudkowskian view, main view under which the FAI project even makes sense. I think one can ask basic questions like which changes move towards more rationality on the margin, though such changes would tend to prioritize rationality over preventing value drift. I’m not sure how much there are general facts about how to avoid value drift (it seems like the relevant kind, i.e. value drift as part of becoming more rational/intelligent, only exists from irrational perspectives, in a way dependent on the mind architecture)
minimal CEV-realist view. it really seems up to agents how much they care about their reflected preferences. maybe changing preferences too often leads to money pumps, or something?
basically says “there are irrational and rational agents, rationality doesn’t apply to irrational agents”, seems somewhat how people treat animals (we don’t generally consider uplifting normative with respect to animals)
at this point you’re at something like ecology / evolutionary game theory, it’s a matter of which things tend to survive/reproduce and there aren’t general decision theories that succeed
re human ontological crises: basically agree, I think it’s reasonably similar to what I wrote. roughly my reason for thinking that it’s hard to solve is that the ideal case would be something like a universal algebra homomorphism (where the new ontology actually agrees with the old one but is more detailed), yet historical cases like physics aren’t homomorphic to previous ontologies in this way, so there is some warping necessary. you could try putting a metric on the warping and minimizing it, but, well, why would someone think the metric is any good, it seems more of a preference than a thing rationality applies to. if you think about it and come up with a solution, let me know, of course.
with respect to grabbing hold of the whole future: you can try looking at historical cases of people trying to grab hold of the future and seeing how that went, it’s a mixed bag with mostly negative reputation, indicating there are downsides as well as upsides, it’s not a “safe” conservative view. see also Against Responsibility. I feel like there’s a risk of getting Pascal’s mugged about “maybe grabbing hold of the future is good, you can’t rule it out, so do it”, there are downsides to spending effort that way. like, suppose some Communists thought capitalism would lead to the destruction of human value with high enough probability that instituting global communism is the conservative option, it doesn’t seem like that worked well (even though a lot of people around here would agree that capitalism tends to leads to human value destruction in the long run). particular opportunities for grabbing hold of the future can be net negative and not worth worrying about even if one of them is a good idea in the long run (I’m not ruling that out, just would have to be convinced of specific opportunities).
overall I’d rather focus on first modeling the likely future and looking for plausible degrees of freedom; a general issue with Pascal’s mugging is it might make people overly attached to world models in which they have ~infinite impact (e.g. Christianity, Communism) which means paying too much attention to wrong world models, not updating to more plausible models in which existential-stakes decisions could be comprehended if they exist. and Obliqueness doesn’t rule out existential stakes (since it’s non-Diagonal).
as another point, Popperian science tends to advance by people making falsifiable claims, “you don’t know if that’s true” isn’t really an objection in that context. the pragmatic claim I would make is: I have some Bayesian reason to believe agents do not in general factor into separate Orthogonal and Diagonal components, this claim is somewhat falsifiable (someone could figure out a theory of this invulnerable to optimization daemons etc), I’m going to spend my attention on the branch where I’m right, I’m not going to worry about Pascal’s mugging type considerations for if I’m wrong (as I said, modeling the world first seems like a good general heuristic), people can falsify it eventually if it’s false.
this whole discussion is not really a defense of Orthogonality given that Yudkowsky presented orthogonality as a descriptive world model, not a normative claim, so sticking to the descriptive level in the original post seems valid; it would be a form of bad epistemology to reject a descriptive update (assuming the arguments are any good) because of pragmatic considerations.
“as important as ever”: no, because our potential influence is lower, and the influence isn’t on things shaped like our values, there has to be a translation, and the translation is different from the original.
CEV: while it addresses “extrapolation” it seems broadly based on assuming the extrapolation is ontologically easy, and “our CEV” is an unproblematic object we can talk about (even though it’s not mathematically formalized, any formalization would be subject to doubt, and even if formalized, we need logical uncertainty over it, and logical induction has additional free parameters in the limit). I’m really trying to respond to orthogonality not CEV though.
from a practical perspective: notice that I am not behaving like Eliezer Yudkowsky. I am not saying the Orthogonality Thesis is true and important to ASI, I am instead saying intelligence/values are Oblique and probably nearly Diagonal (though it’s unclear what I mean by “nearly”). I am not saying a project of aligning superintelligence with human values is a priority. I am not taking research approaches that assume a Diagonal/Orthogonal factorization. I left MIRI partially because I didn’t like their security policies (and because I had longer AI timelines), I thought discussion of abstract research ideas was more important. I am not calling for a global AI shutdown so this project (which is in my view confused) can be completed. I am actually against AI regulation on the margin (I don’t have a full argument for this, it’s a political matter at this point).
I think practicality looks more like having near-term preferences related to modest intelligence increases (as with current humans vs humans with neural nets; how do neural nets benefit or harm you, practically? how can you use them to think better and improve your life?), and not expecting your preferences to extend into the distant future with many ontology changes, so don’t worry about grabbing hold of the whole future etc, think about how to reduce value drift while accepting intelligence increases on the margin. This is a bit like CEV except CEV is in a thought experiment instead of reality.
The “Models of ASI should start with realism” bit IS about practicalities, namely, I think focusing on first forecasting absent a strategy of what to do about the future is practical with respect to any possible influence on the far future; practically, I think your attempted jump to practicality (which might be related to philosophical pragmatism) is impractical in this context.
It occurs to me that maybe you mean something like “Our current (non-extrapolated) values are our real values, and maybe it’s impossible to build or become a superintelligence that shares our real values so we’ll have to choose between alignment and superintelligence.” Is this close to your position?
Close. Alignment of already-existing human values with superintelligence is impossible (I think) because of the arguments given. That doesn’t mean humans have no preferences indirectly relating to superintelligence (especially, we have preferences about modest intelligence increases, and there’s some iterative process).
I’m not sure this captures what you mean, but, if you see a query, do a bunch of reasoning, and get an answer, then you can build a dataset of (query, well-thought guess). Then you can train an AI model on that.
AlphaZero sorta works like this, because it can make a “well-thought guess” (take value and/or Q network, do an iteration of minimax, then make the value/Q network more closely approximate that, in a fixed point fashion)
Learning stochastic inverses is a specific case of “learn to automate Bayesian inference by taking forward samples and learning the backwards model”. It could be applied to LLMs for example, in terms of starting with a forwards LLM and then using it to train a LLM that predicts things out-of-order.
Paul Christiano’s iterated amplification and distillation is applying this idea to ML systems with a human feedback element. If you can expend a bunch of compute to get a good answer, you can train a weaker system to approximate that answer. Or, if you can expend a bunch of compute to get a good rating for answers, you can use that as RL feedback.
Broadly, I take o3 as evidence that Christiano’s work is broadly on the right track with respect to alignment of near-term AI systems. That is, o3 shows that hard questions can be decomposed into easy ones, in a way that involves training weaker models to be part of a big computation. (I don’t understand the o3 details that well, given it’s partially private, but I’m assuming this describes the general outlines). So I think the sort of schemes Christiano has described will be helpful for both alignment and capabilities, and will scale pretty well to impressive systems.
I’m not sure if there’s a form of amortized inference that you think this doesn’t cover well.