All of EliasHasle's Comments + Replies

Conditional on being in a billion-human universe, your probability of having an index between 1 and 1 billion is 1 in 1 billion, and your probability of having any other index is 0. Conditional on being in a trillion-human universe, your probability of having an index between 1 and 1 trillion is 1 in 1 trillion, and your probability of having any other index is 0.

Ehm.. Huh? I would say that:

Conditional on being in a billion-human universe, your probability of having an index between 1 and 1 billion is 1, and your probability of having any other index is 0.... (read more)

2jessicata
Read "having an index" as "having some specific index", e.g. 42. The probabilities add up because typed 1,2,3,4 are each probability 1/4 in total.

If repetitions arise from sampling merely due to high conditional probability given an initial "misstep", they should be avoidable in an MCTS that sought to maximize unconditional probability of the output sequence (or rather conditional upon its input but not upon its own prior output). After entering the "trap" once or a few times, it would simply avoid the unfortunate misstep in subsequent "playouts". From my understanding, that is.

I kind of fell over board when V popped up in the equations without being introduced first.

It seems like the method is sensitive to the ranges of the game reward and the auxiliary penalty. In real life, I suppose one would have to clamp the "game" reward to allow the impact penalty to dominate even when massive gains are foreseen from a big-impact course?

What if the encoding difference penalty were applied after a counterfactual rollout of no-ops after the candidate action or no-op? Couldn't that detect "butterfly effects" of small impactful actions, avoiding "salami slicing" exploits?

Building upon this thought, how about comparing mutated policies to a base policy by sampling possible futures to generate distributions of the encodings up to the farthest step and penalize divergence from the base policy?

Or just train a sampling policy by GD, using a Monte Carlo Tree Search that penalizes actions which alter the future encodings when compared to a pure no-op policy.

I must admit that I did not understand everything in the paper, but I think this excerpt summarizes a crucial point:

"The key issue here is proper conditioning. The unbiasedness of the value estimates V_i discussed in §1 is unbiasedness conditional on mu. In contrast, we might think of the revised estimates ^v_i as being unbiased conditional on V. At the time we optimize and make the decision, we know V but we do not know mu, so proper conditioning dictates that we work with distributions and estimates conditional on V."

The proposed "solution" converts n in... (read more)

The big problem arises when the number of choices is huge and sparsely explored, such as when optimizing a neural network.

But restricting ourselves to n superficially evaluated choices with known estimate variance in each evaluation and with independent errors/noise, then if – as in realistic cases like Monte Carlo Tree Search – we are allowed to perform some additional "measurements" to narrow down the uncertainty, it will be wise to scrutinize the high-expectance choices most – in a way trying to "falsify" their greatness, while increasing the certainty ... (read more)

I think "Metaculus" is a pun on "meta", "calculus" and "meticulous".

"WW3" and "28 years passing" are similarly dangerous "events" for the individual gambler. Why invest with a long-term perspective if there is a significant probability that you eventually cannot harvest... Crucially, the probability of not harvesting the reward may be a lot higher in a "force majeure" situation like WW3, even if one stays alive. But on the other hand, an early WW3 would chop off a lot of the individual existential risk associated with 28 years passing. 🤔 I think there cou... (read more)

[This comment is no longer endorsed by its author]Reply
0[anonymous]
Usually, college courses are the way to go. Find out if you can sit in at courses at nearby colleges and use the textbooks they use. Failing that, I see Coursera has a course on information theory.
[This comment is no longer endorsed by its author]Reply
1mbitton24
I think the reason for the downvotes is that people on LW have generally already formulated their ethical views past the point of wanting to speculate about entirely new normative theories. Your post probably would have received a better reaction had you framed it as a question ("What flaws can you guys find in a utilitarian theory that values the maximization of the amount of computation energy causes before dissolving into high entropy?") rather than as some great breakthrough in moral reasoning. As for constructive feedback, I think Creutzer's response was pretty much spot on. There are already mainstream normative theories like preference utilitarianism that don't directly value pain and pleasure and yet seem to make more sense than the alternatives you offered. Also, your post is specifically about ethics in the age of superintelligence, but doesn't mention CEV. If you're going to offer a completely new theory in a field as well-trod as normative ethics, you need to spend more time debunking alternative popular theories and explaining the advantages yours has over them.
0drethelin
You can discuss it here, just do it in comments. People don't like overambitious top level posts. Read up on the relevant posts that already exist on lesswrong, and comment there and on open threads.
[This comment is no longer endorsed by its author]Reply
0Kawoomba
This would be a variant of the "utility monster" thought experiment. The sensible implementations of utilitarianism take care not to fall into such traps.
[This comment is no longer endorsed by its author]Reply
1drethelin
Don't do either. think about it for a good long while, and ask questions and talk about these topics elsewhere.
[This comment is no longer endorsed by its author]Reply
1[anonymous]
Yes, diversity is important, but you have to take into consideration just how much you add to that diversity. When it comes to physics and information theory, there are people here that know everything you know, and there is nothing you know that nobody here knows. Your advantage over other people come from your experiences, which give you a unique and possibly enlightening perspective. Your ideas, no doubt, come from that perspective. What you should do now is to study so that you can find out whether your ideas can lead to anything useful, and if it does, so that you can tell us in a way we can understand. I recommend studying information theory first. You'll be able to tackle the interesting ideas quickly without having to slog through months of Newtonian mechanics. (Not that mechanics isn't interesting, it's just doesn't seem very relevant to your ideas. It's also possible to skip most of the early undergraduate physics curriculum and go straight to your fields of interest, which I expect would include statistical mechanics and such, but I really don't recommend it unless you have an extremely solid foundation in math. Which you should get anyway if you want to get anywhere, but studying years' worth of prerequisites before the stuff you're actually interested in is demotivating. Start with the stuff you like and fill in as you notice what you don't know. This is difficult but doable with information theory, and ridiculous for say, quantum field theory.) And most importantly, Fix this. You will get nowhere without being a proficient reader. But I'm sure you'll do fine on this one. The very fact that you're on Less Wrong means you enjoy reading about abstract ideas. Just read a lot and it'll come to you.
[This comment is no longer endorsed by its author]Reply
0Creutzer
The point of the reference to paperclip-maximisers was that these values are just as alien to me as those of the paperclip-maximiser. "Putting up a fight against nature's descent from order to chaos" is a bizarre terminal value.
[This comment is no longer endorsed by its author]Reply
3Viliam_Bur
Instead of paperclips, the life-maximizer would probably fill the universe with some simple thing that qualifies as life. Maybe it would be a bacteria-maximizer. Maybe some fractal-shaped bacteria or many different kinds of bacteria, depending on how exactly its goal will be specified. Is this the best "a real altruist" can hope for? If UFAI would be inevitable, I would hope for some approximation of a FAI. If no decent approximation is possible, I would wish some UFAI which is likely to destroy itself. Among different kinds of smart paperclip-maximizers... uhm, I guess I prefer red paperclips aesthetically, but that's really not so important. My legacy is important only if there is someone able to enjoy it.
5DaFranker
This seems like it falls face-first, hands-tied-behind-back right in the giant pit of the Repugnant Conclusion and all of its corollaries, including sentience and intelligence and ability-to-enjoy and ability-to-value. For instance, if I'm a life-maximizer and I don't care about whether the life I create even has the ability to care about anything, and just lives, but has no values or desires or anything even remotely like what humans think of (whatever they do think of) when they think about "values" or "utility"... does that still make me more altruistically ideal and worthy of destroying all humanity? What about intelligence? If the universe is filled to the planck with life, but not a single being is intelligent enough to even do anything more than be, is that simply not an issue? What about consciousness? And, as so troubling in the repugnant conclusion, what if the number of lives is inversely proportional to the maximum quality of each?

But among the people who care and who have a favorite before the campaigning period, would not some change their minds if they saw their favorite being intellectually humiliated on TV? (For the first time ever, that is.)

3[anonymous]
Probably not
[This comment is no longer endorsed by its author]Reply
0Stefan_Schubert
Hi Elias, nice to see that you've found your way here. What are your academic interests? Philosophy, it seems, but what kind? And what else are you interested in?