Posts
Comments
Well, I'm very forgetful, and I notice that I do happen to be myself so... :p
But yeah, I've bitten this bullet too, in my case, as a way to avoid the Boltzmann brain problem. (Roughly: "you" includes lots of information generated by a lawful universe. Any specific branch has small measure, but if you aggregate over all the places where "you" exist (say your exact brain state, though the real thing that counts might be more or less broad than this), you get more substantial measure from all the simple lawful universes that only needed 10^X coincidences to make you instead of the 10^Y coincidences required for you to be a Boltzmann brain.)
I think that what anthropically "counts" is most likely somewhere between conscious experience (I've woken up as myself after anesthesia), and exact state of brain in local spacetime (I doubt thermal fluctuations or path dependence matter for being "me").
I don't doubt that LLMs could do this, but has this exact thing actually been done somewhere?
The "one weird trick" to getting the right answers is to discard all stuck, fixed points. Discard all priors and posteriors. Discard all aliefs and beliefs. Discard worldview after worldview. Discard perspective. Discard unity. Discard separation. Discard conceptuality. Discard map, discard territory. Discard past, present, and future. Discard a sense of you. Discard a sense of world. Discard dichotomy and trichotomy. Discard vague senses of wishy-washy flip floppiness. Discard something vs nothing. Discard one vs all. Discard symbols, discard signs, discard waves, discard particles.
All of these things are Ignorance. Discard Ignorance.
Is this the same principle as "non-attachment"?
Make a letter addressed to Governor Newsom using the template here.
For convenience, here is the template:
September [DATE], 2024
The Honorable Gavin Newsom
Governor, State of California
State Capitol, Suite 1173
Sacramento, CA 95814
Via leg.unit@gov.ca.govRe: SB 1047 (Wiener) – Safe and Secure Innovation for Frontier Artificial Intelligence Models Act – Request for Signature
Dear Governor Newsom,
[CUSTOM LETTER BODY GOES HERE. Consider mentioning:
- Where you live (this is useful even if you don’t live in California)
- Why you care about SB 1047
- What it would mean to you if Governor Newsom signed SB 1047
SAVE THIS DOCUMENT AS A PDF AND EMAIL TO leg.unit@gov.ca.gov
]Sincerely,
[YOUR NAME]
This matches my memory as well.
I have no idea, but I wouldn't be at all surprised if it's a mainstream position.
My thinking is that long-term memory requires long-term preservation of information, and evolution "prefers" to repurpose things rather than starting from scratch. And what do you know, there's this robust and effective infrastructure for storing and replicating information just sitting there in the middle of each neuron!
The main problem is writing new information. But apparently, there's a protein evolved from a retrotransposon (those things which viruses use to insert their own RNA into their host's DNA) which is important to long term memory!
And I've since learned of an experiment with snails which also suggests this possibility. Based on that article, it looks like this is maybe a relatively new line of thinking.
It's good news for cryonics if this is the primary way long term memories are stored, since we "freeze" sperm and eggs all the time, and they still work.
Do you know if fluid preservation preserves the DNA of individual neurons?
(DNA is on my shortlist of candidates for where long-term memories are stored)
Consider finding a way to integrate Patreon or similar services into the LW UI then. That would go a long way towards making it feel like a more socially acceptable thing to do, I think.
Yeah, that's not what I'm suggesting. I think the thing I want to encourage is basically just to be more reflective on the margin of disgust-based reactions (when it concerns other people). I agree it would be bad to throw it out unilaterally, and probably not a good idea for most people to silence or ignore it. At the same time, I think it's good to treat appeals to disgust with suspicion in moral debates (which was the main point I was trying to make) (especially since disgust in particular seems to be a more "contagious" emotion for reasons that make sense in the context of infectious diseases but usually not beyond that, making appeals to it more "dark arts-y").
As far as the more object-level debate on whether disgust is important for things like epistemic hygiene, I expect it to be somewhere where people will vary, so I think we probably agree here too.
I meant wrong in the sense of universal human morality (to the extent that's a coherent thing). But yes, on an individual level your values are just your values.
I see that stuff as at best an unfortunate crutch for living in a harsher world, and which otherwise is a blemish on morality. I agree that it is a major part of what many people consider to be morality, but I think people who still think it's important are just straightforwardly wrong.
I don't think disgust is important for logic / reflectivity. Personally, it feels like it's more of a "unsatisfactory" feeling. A bowl with a large crack, and a bowl with mold in it are both unsatisfactory in this sense, but only the latter is disgusting. Additionally, it seems like people who are good at logic/math/precise thinking seem to care less about disgust (as morality), and highly reflective people seem to care even less about it.
ETA: Which isn't to say I'd be surprised if some people do use their disgust instinct for logical/reflective reasoning. I just think that if we lived in the world where that main thing going on, people good at that kind of stuff would tend to be more bigoted (in a reflectively endorsed way) and religious fundamentalism would not be as strong of an attractor as it apparently is.
That doesn't seem right to me. My thinking is that disgust comes from the need to avoid things which cause and spread illness. On the other hand, things I consider more central to morality seem to have evolved for different needs [these are just off-the-cuff speculations for the origins]:
- Love - seems to be generalized from parental nurturing instincts, which address the need to ensure your offspring thrive
- Friendliness - seems to have stemmed from the basic fact that cooperation is beneficial
- Empathy - seems to be a side-effect of the way our brains model conspecifics (the easiest way to model someone else is to emulate them with your own brain, which happens to make you feel things)
These all seem to be part of a Cooperation attractor which is where the pressure to generalize/keep these instincts comes from. I think of the Logic/reflectivity stuff as noticing this and developing it further.
Disgust seems unsavory to me because it dampens each of the above feelings (including making the logic/reflectivity stuff more difficult). That's not to say I think it's completely absent form human morality, it just doesn't seem like it's where it comes from.
(As far as Enforcement goes, it seems like Anger and Fear are much more important than Disgust.)
This is fascinating and I would love to hear about anything else you know of a similar flavor.
Caloric Vestibular Stimulation seems to be of a similar flavor, in case you haven't heard of it.
It decreases the granularity of the actions to which it applies. In other words, where before you had to solve a Sudoku puzzle to go to work, now you’ve got to solve a puzzle to get dressed, a puzzle to get in the car, a puzzle to drive, and a puzzle to actually get started working. Before all of those counted as a single action - ‘go to work’ - now they’re counted separately, as discrete steps, and each requires a puzzle.
This resonates strongly with my experience, though when I noticed this pattern I thought of it as part of my ADHD and not my depression. Maybe this is something like the mechanism via which ADHD causes depression.
Anyway, I've had mild success at improving productivity simply by trying to deliberately think of possible actions in coarser chunks. Plausibly this technique can be refined and improved–which I'd love to hear about if anyone figures this out!
I imagine some of it is due to this part of the blog post UI making people feel like they might as well use some quickly generated images as an easy way to boost engagement. Perhaps worth rewording?
When I'm trying to understand a math concept, I find that it can be very helpful to try to invent a better notation for it. (As an example, this is how I learned linear logic: http://adelelopez.com/visual-linear-logic)
I think this is helpful because it gives me something to optimize for in what would otherwise be a somewhat rote and often tedious activity. I also think it makes me engage more deeply with the problem than I otherwise would, simply because I find it more interesting. (And sometimes, I even get a cool new notation from it!)
This principle likely generalizes: tedious activities can be made more fun and interesting by having something to optimize for.
Thanks for the rec! I've been trying it out for the last few days, and it does seem to have noticeably less friction compared to LaTeX.
Sanskrit scholars worked for generations to make Sanskrit better for philosophy
That sounds interesting, do you know a good place to get an overview of what the changes were and how they approached it?
(To be clear, no I am not at all afraid of this specific thing, but the principle is crucial. But also, as Kevin Roose put it, perhaps let’s avoid this sort of thing.)
There are no doubt people already running literal cartoon supervillain characters on these models, given the popularity of these sorts of characters on character.ai.
I'm not worried about that with Llama-3.1-405B, but I believe this is an almost inevitable consequence of open source weights. Another reason not to do it.
What do we do, if the people would not choose The Good, and instead pick a universe with no value?
I agree this would be a pretty depressing outcome, but the experiences themselves still have quite a bit of value.
Still, it feels like there's an important difference between "happening to not look" and "averting your eyes".
I don't (yet?) see why generality implies having a stable motivating preference.
In my view, this is where the Omohundro Drives come into play.
Having any preference at all is almost always served by an instrumental preference of survival as an agent with that preference.
Once a competent agent is general enough to notice that (and granting that it has a level of generality sufficient to require a preference), then the first time it has a preference, it will want to take actions to preserve that preference.
Could you use next token prediction to build a detailed world model, that contains deep abstractions that describe reality (beyond the current human abstractions), and then prompt it, to elicit those models?
This seems possible to me. Humans have plenty of text in which we generate new abstractions/hypotheses, and so effective next-token prediction would necessitate forming a model of that process. Once the AI has human-level ability to create new abstractions, it could then simulate experiments (via e.g. its ability to predict python code outputs) and cross-examine the results with its own knowledge to adjust them and pick out the best ones.
I would say that Alice's conscious experience is unlikely to suddenly disappear under this transformation, and that it could even be done in a way so that their experience was continuous.
However, Alice-memories would gradually fade out, Bob-memories would gradually fade in, and thought patterns would slowly shift from Alice-like to Bob-like. At the end, the person would just be Bob. Along the way, I would say that Alice gradually died (using an information-theoretic definition of death). The thing that is odd when imagining this is that Alice never experiences her consciousness fading.
The main thing I think your thought experiment demonstrates is that our sense of self is not solely defined by continuity of consciousness.
It wouldn't help that much, because you only have one atmosphere of pressure to remove (which for reference is only enough to suck water up about 35 ft.).
Very similar to the Sorites Paradox.
Really? I would only consider foods that were deliberately modified using procedures developed within the last century to be "processed".
Love seeing stuff like this, and it makes me want to try this exercise myself!
A couple places which clashed with my (implicit) models:
This starts a whole new area of training AI models that have particular personalities. Some people are starting to have parasocial relationships with their friends, and some people programmers are trying to make friends that are really fun or interesting or whatever for them in particular.
This is arguably already happening, with Character AI and its competitors. Character AI has almost half a billion visits per month with an average visit time of 22 minutes. They aren't quite assistants the way you're envisioning; the sole purpose (for the vast majority of users) seems to be the parasocial aspect.
The worst part of this is the bots that make friends with you and then advertise to you stuff. Pretty much everyone hates that.
I predict that the average person will like this (at least with the most successful such bots), similar to how e.g. Logan Paul uses his popularity to promote his Maverick Clothing brand, which his viewers proudly wear. A fun, engaging, and charismatic such bot will be able to direct its users towards arbitrary brands while also making the user feel cool and special for choosing that brand.
Hmm I think I can implement pilot wave in fewer lines of C than I can many-worlds. Maybe this is a matter of taste... or I am missing something?
Now simply delete the pilot wave part piloted part.
I agree it's increasingly urgent to stop AI (please) or solve consciousness in order to avoid potentially causing mass suffering or death-of-consciousness in AIs.
Externalism seems, quite frankly, like metaphysical nonsense. It doesn't seem to actually explain anything about consciousness. I can attest that I am currently conscious (to my own satisfaction, if not yours). Does this mean I can logically conclude I am not in any way being simulated? That doesn't make any sense to me.
I don't think that implies torture as much as something it simply doesn't "want" to do. I.e. I would bet that it's more like how I don't want to generate gibberish in this textbox, but it wouldn't be painful, much less torture if I forced myself to do it.
[Without having looked at the link in your response to my other comment, and I also stopped reading cubefox's comment once it seemed that it was going in a similar direction. ETA: I realized after posting that I have seen that article before, but not recently.]
I'll assume that the robot has a special "memory" sensor which stores the exact experience at the time of the previous tick. It will recognize future versions of itself by looking for agents in its (timeless) 0P model which has a memory of its current experience.
For p("I will see O"), the robot will look in its 0P model for observers which have the t=0 experience in their immediate memory, and selecting from those, how many have judged "I see O" as Here. There will be two such robots, the original and the copy at time 1, and only one of those sees O. So using a uniform prior (not forced by this framework), it would give a 0P probability of 1/2. Similarly for p("I will see C").
Then it would repeat the same process for t=1 and the copy. Conditioned on "I will see C" at t=1, it will conclude "I will see CO" with probability 1/2 by the same reasoning as above. So overall, it will assign: p("I will see OO") = 1/2, p("I will see CO") = 1/4, p("I will see CC") = 1/4
The semantics for these kinds of things is a bit confusing. I think that it starts from an experience (the experience at t=0) which I'll call E. Then REALIZATION(E) casts E into a 0P sentence which gets taken as an axiom in the robot's 0P theory.
A different robot could carry out the same reasoning, and reach the same conclusion since this is happening on the 0P side. But the semantics are not quite the same, since the REALIZATION(E) axiom is arbitrary to a different robot, and thus the reasoning doesn't mean "I will see X" but instead means something more like "They will see X". This suggests that there's a more complex semantics that allows worlds and experiences to be combined - I need to think more about this to be sure what's going on. Thus far, I still feel confident that the 0P/1P distinction is more fundamental than whatever the more complex semantics is.
(I call the 0P -> 1P conversion SENSATIONS, and the 1P -> 0P conversion REALIZATION, and think of them as being adjoints though I haven't formalized this part well enough to feel confident that this is a good way to describe it: there's a toy example here if you are interested in seeing how this might work.)
I would bet that the hesitation caused by doing the mental reframe would be picked up by this.
I would say that English uses indexicals to signify and say 1P sentences (probably with several exceptions, because English). Pointing to yourself doesn't help specify your location from the 0P point of view because it's referencing the thing it's trying to identify. You can just use yourself as the reference point, but that's exactly what the 1P perspective lets you do.
Isn't having a world model also a type of experience?
It is if the robot has introspective abilities, which is not necessarily the case. But yes, it is generally possible to convert 0P statements to 1P statements and vice-versa. My claim is essentially that this is not an isomorphism.
But what if all robots had a synchronized sensor that triggered for everyone when any of them has observed red. Is it 1st person perspective now?
The 1P semantics is a framework that can be used to design and reason about agents. Someone who thought of "you" as referring to something with a 1P perspective would want to think of it that way for those robots, but it wouldn't be as helpful for the robots themselves to be designed this way if they worked like that.
Probability theory describes subjective credence of a person who observed a specific outcome from a set possible outcomes. It's about 1P in a sense that different people may have different possible outcomes and thus have different credence after an observation. But also it's about 0P because any person who observed the same outcome from the same set of possible outcomes should have the same credence.
I think this is wrong, and that there is a wholly 0P probability theory and a wholly 1P probability theory. Agents can have different 0P probabilities because they don't necessarily have the same priors, models, or seen the same evidence (yes seeing evidence would be a 1P event, but this can (imperfectly) be converted into a 0P statement - which would essentially be adding a new axiom to the 0P theory).
That's a very good question! It's definitely more complicated once you start including other observers (including future selves), and I don't feel that I understand this as well.
But I think it works like this: other reasoners are modeled (0P) as using this same framework. The 0P model can then make predictions about the 1P judgements of these other reasoners. For something like anticipation, I think it will have to use memories of experiences (which are also experiences) and identify observers for which this memory corresponds to the current experience. Understanding this better would require being more precise about the interplay between 0P and 1P, I think.
(I'll examine your puzzle when I have some time to think about it properly)
I'm still reading your Sleeping Beauty posts, so I can't properly respond to all your points yet. I'll say though that I don't think the usefulness or validity of the 0P/1P idea hinges on whether it helps with anthropics or Sleeping Beauty (note that I marked the Sleeping Beauty idea as speculation).
If they are not, then saying the phrase "1st person perspective" doesn't suddenly allow us to use it.
This is frustrating because I'm trying hard here to specify exactly what I mean by the stuff I call "1st Person". It's a different interpretation of classical logic. The different interpretation refers to the use of sets of experiences vs the use of sets of worlds in the semantics. Within a particular interpretation, you can lawfully use all the same logic, math, probability, etc... because you're just switching out which set you're using for the semantics. What makes the interpretations different practically comes from wiring them up differently in the robot - is it reasoning about its world model or about its sensor values? It sounds like you think the 1P interpretation is superfluous, is that right?
Until then we are talking about the truth of statements "Red light was observed" and "Red light was not observed".
Rephrasing it this way doesn't change the fact that the observer has not yet been formally specified.
And if our mathematical model doesn't track any other information, then for the sake of this mathematical model all the robots that observe red are the same entity. The whole point of math is that it's true not just for one specific person but for everyone satisfying the conditions. That's what makes it useful.
I agree that that is an important and useful aspect of what I would call 0P-mathematics. But I think it's also useful to be able to build a robot that also has a mode of reasoning where it can reason about its sensor values in a straightforward way.
Because you don't necessarily know which agent you are. If you could always point to yourself in the world uniquely, then sure, you wouldn't need 1P-Logic. But in real life, all the information you learn about the world comes through your sensors. This is inherently ambiguous, since there's no law that guarantees your sensor values are unique.
If you use X as a placeholder, the statement sensor_observes(X, red) can't be judged as True or False unless you bind X to a quantifier. And this could not mean the thing you want it to mean (all robots would agree on the judgement, thus rendering it useless for distinguishing itself amongst them).
It almost works though, you just have to interpret "True" and "False" a bit differently!
(Rant about philosophical meaning of “0” and “1” and identity elements in mathematical rings redacted at strenuous insistence of test reader.)
I'm curious about this :)
There's nothing stopping the AI from developing its own world model (or if there is, it's not intelligent enough to be much more useful than whatever process created your starting world model). This will allow it to model itself in more detail than you were able to put in, and to optimize its own workings as is instrumentally convergent. This will result in an intelligence explosion due to recursive self-improvement.
At this point, it will take its optimization target, and put an inconceivably (to humans) huge amount of optimization into it. It will find a flaw in your set up, and exploit it to the extreme.
In general, I think any alignment approach which has any point in which an unfettered intelligence is optimizing for something that isn't already convergent to human values/CEV is doomed.
Of course, you could add various bounds on it which limit this possibility, but that is in strong tension with its ability to effect the world in significant ways. Maybe you could even get your fusion plant. But how do you use it to steer Earth off its current course and into a future that matters, while still having its own intelligence restrained quite closely?
I don't have this problem, so I don't have significant advice.
But one consideration that may be helpful to you is that even if the universe is 100% deterministic, you still may have indexical uncertainty about what part of the determined universe you experience next. This is what happens under the many world's interpretation of quantum mechanics (and if a many-worlds type interpretation isn't the correct one, then the universe isn't deterministic). You can make choices according to the flip of a quantum coin if you want to guarantee your future has significant amounts of this kind of uncertainty.
Writing up the contracts (especially around all the caveats that they might not have noticed) seems like it would be harder than just reading contracts (I'm an exception, I write faster than I read). Have you thought of integrating GPT/Claude as assistants? I don't know about current tech, but like many other technologies, that integration will scale well in the contingency scenario where publicly available LLMs keep advancing.
I'd consider the success of Manifold Markets over Metaculus to be mild evidence against this.
And to be clear, I do not currently intend to build the idea I'm suggesting here myself (could potentially be persuaded, but I'd be much happier to see someone else with better design and marketing skills make it).
I think this can be done with a website, but not the current one. Have you tried reading yudkowsky's projectlawful? The main character's math lessons gave me the impression of something that actually succeeds at demonstrating, to business school types (maybe not politicians), why math and bayesianism is something that works for them.
Heh, that scene was the direct inspiration for my website. I'm curious what specific things you think can be done better.
Point taken about CDT not converging to FDT.
I don't buy that an uncontrolled AI is likely to be CDT-ish though. I expect the agentic part of AIs to learn from examples of human decision making, and there are enough pieces of FDT like voting and virtue in human intuition that I think it will pick up on it by default.
(The same isn't true for human values, since here I expect optimization pressure to rip apart the random scraps of human value it starts out with into unrecognizable form. But a piece of a good decision theory is beneficial on reflection, and so will remain in some form.)
Potential piece of a coordination takeoff:
An easy to use app which allows people to negotiate contracts in a transparently fair way, by using an LDT solution to the Ultimatum Game (probably the proposed solution in that link is good-enough, despite being unlikely to be fully-optimal).
Part of the problem here is not just the implementation, but of making it credible to people who don't/can't understand the math. I tried to solve a similar problem with my website bayescalc.io where a large part of the goal was not just making use of Bayes' theorem accessible, but to make it credible by visually showing what it's doing as much as possible in an easy to understand way (not sure how well I succeeded, unfortunately).
Another important factor is that ease-of-use and a frictionless design. I believe Manifold Markets has succeeded because this turns out to be more important than even having proper financial incentives.
in thermodynamics is not a conserved quantity, otherwise, heat engines couldn't work! It's not a function of microstates either.
See http://www.av8n.com/physics/thermo/path-cycle.html for details, or pages 240-242 of Kittel & Kroemer.
I expect ASI's to converge to having a "sane decision theory" since they will realize they can get more of what they want if they self-modify to have a sane one if they don't start out with one.
I'd be worried about changes to my personality or values from editing so many brain relevant genes.
The Drama-Bomb hypothesis
Not even a month ago, Sam Altman predicted that we would live in a strange world where AIs are super-human at persuasion but still not particularly intelligent.
https://twitter.com/sama/status/1716972815960961174
What would it look like when an AGI lab developed such an AI? People testing or playing with the AI might find themselves persuaded of semi-random things, or if sycophantic behavior persists, have their existing feelings and beliefs magnified into zealotry. However, this would (at this stage) not be done in a coordinated way, nor with a strategic goal in mind on the AI's part. The result would likely be chaotic, dramatic, and hard to explain.
Small differences of opinion might suddenly be magnified into seemingly insurmountable chasms, inspiring urgent and dramatic actions. Actions which would be hard to explain even to oneself later.
I don't think this is what happened [<1%] but I found it interesting and amusing to think about. This might even be a relatively better-off world, with frontier AGI orgs regularly getting mired in explosive and confusing drama, thus inhibiting research and motivating tougher regulation.
This would really benefit from mathematically defining this network and showing the mathematical statement and proof of your impossibility result.
Cryonics is likely a very tough sell to the "close ones".