I don't think the Elimination approach gives P(Heads|Awake) = 1/3 or P(Monday|Awake) = 2/3 in the Single Awakening problem. In that problem, there are 6 possibilities:
P(Heads&Monday) = 0.25
P(Heads&Tuesday) = 0.25
P(Tails&Monday&Woken) = 0.125
P(Tails&Monday&Sleeping) = 0.125
P(Tails&Tuesday&Woken) = 0.125
P(Tails&Tuesday&Sleeping) = 0.125
Therefore:
P(Heads|Awake)
= P(Heads&Monday) / (P(Heads&Monday) + P(Tails&Monday&Woken) + P(Tails&Tuesday&Woken))
= 0.5
And:
P(Monday|Awake)
= (P(Heads&Monday) + P(T...
I also consider myself as someone who had—and still has—high hopes for rationality, and so I think it’s sad that we disagree, not on the object level, but on whether we can trust the community to faithfully report their beliefs. Sure, some of it may be political maneuvering, but I mostly think it’s political maneuvering of the form of—tailoring the words, metaphors, and style to a particular audience, and choosing to engage on particular issues, rather than outright lying about beliefs.
I don’t think I’m using “semantics” in a non-standard sense, but I may ...
I owe you an apology; you’re right that you did not accuse me of violating norms, and I’m sorry for saying that you did. I only intended to draw parallels between your focus on the meta level and Zack’s focus on the meta level, and in my hurry I erred in painting you and him with the same brush.
I additionally want to clarify that I didn’t think you were accusing me of lying, but merely wanted preemptively close off some of the possible directions this conversation could go.
Thank you for providing those links! I did see some of them on his blog and skipped ...
I haven’t read everything Zack has written, so feel free to link me something, but almost everything I’ve read, including this post, includes far more intra-rationalist politicking than discussion of object level matters.
I know other people are interested in those things. I specifically phrased my previous post in an attempt to avoid arguing about what other people care about. I can neither defend nor explain their positions. Neither do I intend to dismiss or malign those preferences by labeling them semantics. That previous sentence is not to be read as a...
I haven’t read everything Zack has written, so feel free to link me something, but almost everything I’ve read, including this post, includes far more intra-rationalist politicking than discussion of object level matters.
Certainly:
http://unrem...
Yeah, what factual question about empirical categories is/was Zack interested in resolving? Tabooing the words “man” and “woman”, since what I mean by semantics is “which categories get which label”. I’m not super interested in discussing which empirical category should be associated with the phonemes /mæn/, and I’m not super interested in the linguistic investigation of the way different groups of English speakers assign meaning to that sequence of phonemes, both of which I lump under the umbrella of semantics.
I think this post could be really good, and perhaps there should be an effort to make this post as good as it can be. Right now I think it has a number of issues.
It's too short. It moves very quickly past the important technical details, trusting the user to pick them up. I think it would be better if it was a bit longer and luxuriated on the important technical bits.
It is very physics-brained. Ideally we could get some math-literate non-physicists to go over this with help from a physicist to do a better job phrasing it in ways that are unfamiliar t
Yeah, this post makes me wonder if there are non-abusive employers in EA who are nevertheless enabling abusers by normalizing behavior that makes abuse popular. Employers who pay their employees months late without clarity on why and what the plan is to get people paid eventually. Employers who employ people without writing things down, like how much people will get paid and when. Employers who try to enforce non-disclosure of work culture and pay.
None of the things above are necessarily dealbreakers in the right context or environment, but when an employ...
Find an area of the thing you want to do where quality matters to you less. Instead of trying to write the next great American novel, write fanfic[1]. Instead of trying to paint a masterpiece, buy a sketchbook and trace a bunch of stuff. Instead of trying to replace your dish-ware with handmade ceramics, see how many mugs you can make in an hour. Instead of trying to invent a new beautiful operating system in a new programming language, hack together a program for a one-off use case and then throw it away.
[1] not a diss to fanfic—but for me, at least, it's easier to not worry about my writing quality when I do so
I think an important point missing from the discussion on compute is training vs inference: you can totally get a state-of-the-art language model performing inference on a laptop.
This is a slight point in favor of Yudkowsky: thinking is cheap, finding the right algorithm (including weights) is expensive. Right now we're brute-forcing the discovery of this algorithm using a LOT of data, and maybe it's impossible to do any better than brute-forcing. (Well, the human brain can do it, but I'll ignore that.)
Could you run a LLM on a desktop from 2008? No. But, o...
So, like, I remain pretty strongly pro Hanson on this point:
I think LLaMA 7b is very cool, but it's really stretching it to call it a state-of-the-art language model. It's much worse than LLaMA 65b, which much worse than GPT-4, which most people think is > 100b as far as I know. I'm using a 12b model right now while working on an interpretability project... and it is just much, much dumber than these big ones.
Not being able to train isn't a small deal, I think. Learning in a long-term way is a big part of intelligence.
Overall, and not to be too
Hmm, no, I don't believe I use sex and gender interchangeably. Let's taboo those two terms.
I think that most people don't care about a person's chromosomes. When I inspect the way I use the words "sex" and "gender", I don't feel like either of them is a disguised query for that person's chromosomes.
I think that many people care about hormone balances. Testosterone and Estrogen change the way your body behaves, and the type of hormone a person's body naturally produces and whether they're suppressing that and/or augmenting with a different hormone is defini...
I think that one thing you're missing is that lots of people... use gender as a very strong feature of navigating the world. They treat "male" and "female" as natural categories, and make lots of judgements based on whether someone "is" male or female.
You don't seem to do that, which puts you pretty far along the spectrum towards gender abolition, and you're right, from a gender abolition perspective there's no reason to be trans (or to be worried about people using the restroom they prefer or wearing the clothes they prefer or taking hormones to alter the...
I've met many self-identified women (trans and otherwise) that did not prefer female-gendered terms, prompting plenty of inadvertent social gaffes on my end.
I think that if someone self identifies as a woman to you, and you use a gendered term to describe them (she, policewoman, actress) that is not a social gaffe on your part. I think that it is fine for someone to identify as a woman, but advocate for the use of gender neutral language in all cases even applied to them, but they should not put pressure on those who do so differently.
...and the most rel
Some examples of places where knowing someone identifies as a woman vs. as a man vs. as nonbinary would affect your view of their behavior:
I appreciate that you provided specific examples but my immediate reaction to your list was one of bafflement.
While I personally don't care what bathrooms or changing room anyone uses, to the extent someone does prompt a negative reaction for being in the "wrong" room it would be entirely predicated on how that individual is perceived by others (read: pass), not what they personally identify as.
Similarly, I don't ...
The common justification trotted out (that it’s necessary to include the theoretically-possible transman who somehow can get pregnant and apparently suffers no dysphoria from carrying a fetus to term) is completely daft.
This is as far as I can tell completely false. Plenty of trans men carry fetuses to term. Plenty of trans men carried fetuses to term before they came out as trans men. Plenty of trans men decide to carry fetuses to term after they come out as trans men. A couple of facts I believe about the world that may help you make sense of this:
Hi! I'm not sure where exactly in this thread to jump in, so I'm just doing it here.
I like this thread! It's definitely one of my favorite discussions about gender between people with pretty different perspectives. I also like the OP; I found it to be surprisingly clear and grounded, and to point at some places where I am pretty confused myself.
>Originally you said that my post lacked an "understanding of the experiences of trans people" and I'm still eager to learn more! What am I missing exactly and what sources would you recommend I read?
I'm taking a...
Orthogonality in design states that we can construct an AGI which optimizes for any goal. Orthogonality at runtime would be an AGI design that would consist of an AGI which can switch between arbitrary goals while operating. Here, we are only really talking about the latter orthogonality
This should not be relegated to a footnote. I've always thought that design-time orthogonality is the core of the orthogonality thesis, and I was very confused by this post until I read the footnote.
There are tactics I have available to me which are not oriented towards truthseeking, but instead oriented towards "raising my status at the expense of yours". I would like to not use those tactics, because I think that they destroy the commons. I view "collaborative truth seeking" as a commitment between interlocutors to avoid those tactics which are good at status games or preaching to the choir, and focus on tactics which are good at convincing.
Additionally,
...I can can just ... listen to the counterarguments and judge them on their merits, without getting
I walked through some examples of Shapley Value here, and I'm not so sure it satisfies exactly what we want on an object level. I don't have a great realistic example here, but Shapley Value assigns counterfactual value to individuals who did in fact not contribute at all, if they would have contributed were your higher-performers not present. So you can easily have "dead weight" on a team which has a high Shapley Value, as long as they could provide value if their better teammates were gone.
my friend responds, "Your beliefs/desires, and ultimately your choice to study English, was determined by the state of your brain, which was determined by the laws of physics. So don't feel bad! You were forced to study English by the laws of physics."
The joke I always told was
"I choose to believe in free will, because if free will does exist, then I've chosen correctly, and if free will doesn't exist, then I was determined to choose this way since the beginning of the universe, and it's not my fault that I'm wrong!"
Reductionism means that "Jones" and "Smith" are not platonic concepts. They're made out of parts, and you can look at the parts and ask how they contribute to moral responsibility.
When you say "Smith has a brain tumor that made him do it", you are conceiving of Smith and the brain tumor as different parts, and concluding that the non-tumor part isn't responsible. If you ask "Is the Smith-and-brain-tumor system responsible for the murder", the answer is yes. If you break the mind of Jones into parts, you could similarly ask "Is the visual cortex of Jones re...
No public estimates, but the difficulty of self driving cars definitely pushed my AGI timelines back. In 2018 I predicted full self driving by 2023; now that’s looking unlikely. Yes, the advance in text and image understanding and generation has improved a lot since 2018, but instead of shortening my estimates that’s merely rotated which capabilities will come online earlier and which will wait until AGI.
However, I expect some crazy TAI in the next few years. I fully expect “solve all the millennium problems” to be doable without AGI, as well as much of coding/design/engineering work. I also think it’s likely that text models will be able to do the work of a paralegal/research assistant/copywriter without AGI.
That’s what people used to say about chess and go. Yes, mathematics requires intuition, but so does chess; the game tree’s too big to be explored fully.
Mathematics requires greater intuition and has a much broader and deeper “game” tree, but once we figure out the analogue to self-play, I think it will quickly surpass human mathematicians.
GPT-4 (Edited because I actually realize I put way more than 5% weight on the original phrasing): SOTA on language translation for every language (not just English/French and whatever else GPT-3 has), without fine-tuning.
Not GPT-4 specifically, assuming they keep the focus on next-token prediction of all human text, but "around the time of GPT-4": Superhuman theorem proving. I expect one of the millennium problems to be solved by an AI sometime in the next 5 years.
"metis" is an ancient greek word/goddess which originally meant "magical cunning", which drifted throughout ancient greek culture to mean something more like "wisdom/prudence/the je ne sais quoi of being able to solve practical problems".
James C. Scott uses it in his book Seeing Like a State to mean the implicit knowledge passed down through a culture.
I think it's not necessarily the case that free-market pairwise bargaining always leads to the Shapley value. 10Y = X -Y has an infinite number of solutions, and the only principled ways I know of for choosing solutions is either Shapley value or the fact that in this scenario, since there are no other jobs, the owner should be able to negotiate X and Y down to epsilon.
My experience:
...High dimensional spaces are unlikely to have local optima, and probably don’t have any optima at all.
Just recall what is necessary for a set of parameters to be at a optimum. All the gradients need to be zero, and the hessian needs to be positive semidefinite. In other words, you need to be surrounded by walls. In 4 dimensions, you can walk through walls. GPT3 has 175 billion parameters. In 175 billion dimensions, walls are so far beneath your notice that if you observe them at all it is like God looking down upon individual protons.
If there’s any randomne
DallE2 is bad at prepositional phrases (above, inside) and negation. It can understand some sentence structure, but not reliably.
In the first example, none of those are paragraphs longer than a single sentence.
In the first example, the images are not stylistically coherent! The bees are illustrated inconsistently from picture to picture. They look like they were drawn by different people working off of similar prompts and with similar materials.
The variational feature is not what I’m talking about; I mean something like “Draw a dragon sleeping o
DallE2 is bad at prepositional phrases (above, inside) and negation. It can understand some sentence structure, but not reliably.
Goalpost moving. DALL-E 2 can generate samples matching lots of complex descriptions which are not 'noun phrases', and GLIDE is even better at it (also covered in the paper). You said it can't. It can. Even narrowly, your claim is poorly supported, and for the broader discussion this is in the context of, misleading. You also have not provided any sources or general reasons for this sweeping assertion to be true, or for the br...
Here's another possible explanation: The models aren't actually as impressive as they're made out to be. For example, take DallE2. Yes, it can create amazingly realistic depictions of noun phrases automatically. But can it give you a stylistically coherent drawing based on a paragraph of text? Probably not. Can it draw the same character in three separate scenarios? No, it cannot.
DallE2 basically lifts the floor of quality for what you can get for free. But anyone who actually wants or needs the things you can get from a human artist cannot yet get it from an AI.
See also, this review of a startup that tries to do data extraction from papers: https://twitter.com/s_r_constantin/status/1518215876201250816
Meta: I disagree with Alex's decision to delete Gwern's comment on this answer. People can reasonably disagree about the optimal balance between 'more dickish' (leaves more room for candor, bluntness, and playfulness in discussions) and 'less dickish' (encourages calm and a focus on content) in an intellectual community. And on LW, relatively high-karma users like Alex are allowed to moderate discussion of their posts, so Alex is free to promote the balance he thinks is best here.
But regardless of where you fall on that spectrum, I think LW should have a s...
As other commenters have said, approximating integer ratios is important.
and it just so happens that these ratios are close to powers of the 12th root of 2.
You can do the math and verify those numbers are relatively close.
It's important to recognize that this correspondence is relatively recently discov...
“Republicans (30%) are approximately three times as likely as Democrats (11%) to agree with the statement, “Because things have gotten so far off track, true American patriots may have to resort to violence in order to save our country.” Agreement with this statement rises to 40% among Republicans who most trust far-right news sources and is 32% among those who most trust Fox News. One in four white evangelicals (26%) also agree that political violence may be necessary to save the country”
https://www.prri.org/press-release/competing-visions-of-america-an-e...
There are no reader comments -- this is a "glowfic", the whole thread is the story. Each "comment" is a portion of the story. You need read it all.
You do need pay attention to the usernames because they often identify which character is speaking (e.g. "Keltham", or "Carissa Sevar"), and an approximation of their facial expression via the graphics.
(The post times and history of each comment are not important in-story, the "history" are just edits made to that story portion, the post times are just when it was posted)
One of the authors is "Iarwain" (who is EY) the other author is "lintamande".
...There was a critical moment in 2006(?) where Hinton and Salakhutdinov(?) proposed training Restricted Boltzmann machines unsupervised in layers, and then 'unrolling' the RBMs to initialize the weights in the network, and then you could do further gradient descent updates from there, because the activations and gradients wouldn't explode or die out given that initialization. That got people to, I dunno, 6 layers instead of 3 layers or something? But it focused attention on the problem of exploding gradients as the reason why deeply layered neural nets
Perhaps there is a different scheme for dividing gains from coöperation which satisfies some of the things we want but not superadditivity, but I’m unfamiliar with one. Please let me know if you find anything in that vein, I’d love to read about some alternatives to Shapley Value.