LessWrong 2.0 Reader

Some of the underlying evidence, like e.g. Altman's public statements, is relevant to other forms of scaling. Some of the underlying evidence, like e.g. the data wall, is not. That cashes out to differing levels of confidence in different versions of the prediction.

At first look, I like your design a lot. Even though I am required to (because "imitation is the sincerest flattery"), it has its own fairly unique style which doesn't immediately remind me of anything else (aside from mine). I like the playfulness and use of some color. I am also impressed by your design writeup: you have covered far more than I would have expected and definitely thought it through. I may wind up stealing some ideas here.

More miscellaneous observations:

• dark mode selector: you use a 2-state light vs dark selector. This is the obvious thing, but we think that it is ultimately wrong and you need a three-state selector to accommodate "auto". I think this is especially important given how many people now read websites like yours or mine on their smartphones, often at night or in bed, and just assume now that everything will use dark-mode as appropriate. (I'm sure you've seen many screenshots of Gwern.net on Twitter at this point, and noticed that they are almost always smartphones, and then much of the time, dark-mode. I don't expect auto to become any less common.)

• Smallcaps acronyms: I did the same thing originally but ultimately removed them. They wound up adding a lot to the page, and while they initially (ahem) looked cool and fancy, they alienated readers and over time I just kept noticing them and feeling more and more alienated by them. Smallcaps may be "proper" typographically, but I think that ship has sailed: we read so little material with acronyms typeset in true smallcaps, that it now achieves the opposite of the intended effect - it's the 'NASA' which is smallcapsed which is bizarre and alien looking, not the regular old 'NASA'. Is it worth spending "weirdness points" on? I ultimately felt not.

• Color: You mention that link-icons can be chaotic if colored. I agree, but in your case, I think you have a lot of scope to be playful with color.

  For example, you went to a lot of trouble to separate the dropcaps and enable the fun colored dropcaps. Why not make the dropcaps colored... on hover?

  In fact, why not make 'on hover' a core design principle? (This would be a good time to try to write down a few catchphrases or design principles to sum up your goals here. Why dropcaps or the animated pond logo? etc) When I look at your pond, I feel like it would be wonderful if the pond was animated on hover - if when I hovered, then it was animated.

  Right now, it feels a bit awkward. It's animated just enough to bother me in the corner of my eye, but not enough to consciously notice it. It is also too small, IMO. The detail is illegible at this size. Also, in the long run, I think you are better off looking into generative pixel art for adding more images/video in that style. You may think you are willing to pay $270 each time, but you're not. It'll deter you and inherently create a scarcity mindset. This is a website design which would benefit from fun little pixel art motifs all over the place, and you want to be able to flip over as soon as an idea for a trout element hits you and start creating it. You don't have to go all Yamauchi No.10 Family Office on the reader, but for this sort of cozy playful design, I think the more the better, so there's a feeling of always something cute around the corner.

  You have a nice fleuron footer. But wouldn't it be so much niftier if that fish were cheerfully animated once I hover over it, and it does little trout flips around my cursor? And if the fleurons became brighter blue or richer texture?

  And wouldn't it be nice if all of the trout link-icons also turned blue on hover? (I think the trout link-icon spacing is a bit off, incidentally. The Youtube link icon is also definitely bad with the "YouTube's logo is definitely red" example - way too close to the 's'.) We have recently implemented link-icon colors on Gwern.net (some background), and while I'm still not sure how appropriate it is for Gwern.net or if it needs to be rethought, I feel it's very appropriate for your design.

  Lots of things you could do with it. For example, you could have a gentle "breathing" cycle of all of the colors, similar to some of Apple's light icons - the page could use JS to very slowly cycle through the default color-less version to the hover versions and back. (Perhaps just for the first minute, or perhaps instead after a few minutes, whatever feels more esthetic.)

• Might note "callouts" are also called "admonitions".

• Visual regression testing: you can also check snapshots of the raw HTML too. Since you are trying to bake a lot into the HTML, this should work well for you and complement the image approach. This can be as just downloading some URLs and running diff against a directory of older downloads. I implemented this a few months ago and it was easy to implement and has given me more confidence when I review the lorem unit-test pages to check that any changes in the final HTML make sense.

• I notice way down in the footer a backlinks section, but doesn't seem to be covered in the design page yet? Also, possible bug: the backlinks section of the design page includes... the design page?

• "Text transformers" seems like a risky terminology choice, especially given your profession & site content. Maybe just call them "compilers" or something.

• List indentation: your lists do not indent the contents / outdent the list marker. Is that deliberate?

Overall, best new personal website I've seen in a while: ★★★★☆.

I look forward to it being tidied up some ore, and seeing what clever new touches you put on it as you keep evolving it and presumably can experiment with things like LLM rewrites or integration or add more pixel art, so I can add that last star. :)

If blocks of text could be properly encoded and decoded vae style, this would not only improve compute requirements for transformers (since you would only have to predict latents), but it might also offer a sort of "idea" autoregression, where a thought is slowly formulated, then decoded into a sequence of tokens. Alas, it seems there is a large chunk of research left unresolved here; I doubt that researchers are willing to allocate time to this.

Quantum immortality and gun jammed do not contradict each other: for example, if we survive 10 rounds failures because of QI, we most likely survive only on those timelines where gun is broken. So both QI and gun jamming can be true and support one another and there is no contradiction.

I don't see advantage to remaining agnostic, compared to:

1) Acquire all the private truth one can.

Plus:

2) Tell all the public truth one is willing to incur the costs of, with priority for telling public truths about what one would and wouldn't share (e.g. prioritizing to not pose as more truth-telling than one is).

--

The reason I prefer this policy to the OP's "don't seek truth on low-import highly-politicized matters" is that I fear not-seeking-truth begets bad habits.  Also I fear I may misunderstand how important things are if I allow politics to influence which topics-that-interest-my-brain I do/don't pursue, compared to my current policy of having some attentional budget for "anything that interests me, whether or not it seems useful/virtuous."

Yeah but a bunch of people might actually answer how their neigbours will vote, given that that's what the pollster asked - and if the question is phrased as the post assumes, that's going to be a massive issue.

James, thank you for a well-written comment. It was a pleasure to read. Looking forward to Seth's response. Genuinely interested in hearing his thoughts.

One problem here is that quantum immortality and angel immortality eventually merges: for example, if we survive 10 LHC failures because of QI, we most likely survive only on those timelines where some alien stops LHC. So both QI and angel immortality can be true and support one another and there is no contradiction. 

I know this post and have two problems with it: what they call 'anthropic shadow" is not proper term as Bostrom defined anthropic shadow as underestimation of past risks based on the fact of survival in his article this the same name. But it's ok. 

The more serious problem is that quantum immortality and angel immortality eventually merges: for example, if we survive 10 LHC failures because of QI, we most likely survive only on those timeline where some alien stops LHC. So both QI and angel immortality can be true and support one another and there is no contradiction. 

We can discuss anything that exists, that might exist, that did exist, that could exist, and that could not exist. So no matter what form your predict-the-next-token language model takes, if it is trained over the entire corpus of the written word, the representations it forms will be pretty hard to understand, because the representations encode an entire understanding of the entire world.

Perhaps. 

Imagine a huge number of very skilled programmers tried to manually hard code a ChatGPT in python. 

Ask this pyGPT to play chess, and it will play chess. Look under the hood, and you see a chess engine programmed in. Ask it to solve algebra problems, a symbolic algebra package is in there. All in the best neat and well commented code.

Ask it to compose poetry, and you have some algorithm that checks if 2 words rhyme. Some syllable counter. Etc. 

Rot13 is done with a hardcoded rot13 algorithm. 

Somewhere in the algorithm is a giant list of facts, containing "Penguins  Live In Antarctica".  And if you change this fact to say "Penguins Live in Canada", then the AI will believe this. (Or spot it's inconsistency with other facts?) 

And with one simple change, the AI believes this consistently. Penguins appear when this AI is asked for poems about canada, and don't appear in poems about Antarctica. 

When asked about the native canadian diet, it will speculate that this likely included penguin, but say that it doesn't know of any documented examples of this. 

Can you build something with ChatGPT level performance entirely out of human comprehensible programmatic parts?

Obviously having humans program these parts directly would be slow. (We are still talking about a lot of code.) But if some algorithm could generate that code?