see also my eaforum at https://forum.effectivealtruism.org/users/dirk and my tumblr at https://d-i-r-k-s-t-r-i-d-e-r.tumblr.com/ .
Still available on archive.org; https://web.archive.org/web/20200107082033im_/https://i.vimeocdn.com/video/364878640_1280x720.jpg .
The problem with Example A is that the artist is copying superficial elements of a particular style, without understanding the underlying principles that make good a good figure drawing. Example B has lots of overlapping lines, and vague messy shapes. But the figures there communicate a good understanding of anatomy, a grasp of weight, decent composition.
I think the image may have broken; example B currently appears as a series of vertical bars of solid color.
Why on earth would we expect a distilled model to have continuity of experience with the model it was distilled from? Even if you subscribe to computationalism, the distilled model is not the same algorithm.
The reader knows that, certainly. But they don't know that you know that; that's why you have to clarify that you do. (And yes, you have to! Most people in fact do not know that their opinions aren't fact).
Saying "I think" isn't making yourself small but making yourself the appropriate size; frequently stating opinions as fact is an unwarranted status-grab and pollutes the epistemic commons.
But it couldn't be a serious criticism; the Necronomicon hasn't actually been discovered.
I know henryaj's also planning something; his comment is here in case you'd like to coordinate with him.
Hmm. I think the standard narrative is that at first it's difficult to think of journal entries, because you're not in the habit of taking note when pleasant things occur, but over time you get into the habit of mentally tracking nice things throughout the day, which enables you to list things more easily. If you don't feel joy about it, it doesn't go in the gratitude journal; you might take some other action about the unpleasant feelings (journaling about them elsewhere, asking for a raise, etc.), but trying to make yourself feel grateful by brute force is unlikely to help.
If you notice partway through the day that you haven't had anything enjoyable today, you might try adding some readily-accessible source of pleasure (tasty food, a book you like, recreational drugs if you're into that sort of thing, etc.). However, it's perfectly alright to leave a day empty, or to simply report that nothing you're grateful for happened that day. If you have a great number of days with nothing nice in them, you likely have some problem for which a gratitude journal is not the right tool; broadly, such issues are best addressed with changes to your material circumstances, or (if psychological) medication (my personal preference) and/or therapy.
One reason to journal and/or use a planner in general is that this provides an opportunity to make use of pretty stationery items (e.g. washi tape and stickers), if you're into that sort of thing; many such items can be had quite cheaply online, and they come in a wide variety of attractive designs.
Well, you're definitely not supposed to fill it with things you're supposed to be grateful for, for one thing; you're supposed to fill it with things that actually bring you joy (above baseline levels). E.G.: "I'm happy I got paid today"; "I had a really nice date with my partner"; "I enjoyed playing [video game]; "the sunset was lovely today"; etc. The idea is that it trains you to pay attention to, and remember to savor, positive events in your life, which in turn will improve your overall enjoyment.
If you'd like to try your hand at it, I recommend website https://exoloom.io/ for generating from them; it lets you make the LLM generate several short snippets at the same time and select the best, which IMO is very valuable for producing coherent output. It also lets you view the resulting conversation tree as, well, a tree, which is quite helpful for browsing. (I do recommend reading the user manual, though; I didn't, and had trouble figuring out the UI as a result). It offers Llama 405B (base) and Deepseek V3 (also a base model), and provides a limited number of free credits each day.
In my own experimentation, I've thought base models' poetry substantially better and more creative than that of assistant models; of course my outputs won't have been written to your tastes, but if you'd like a quick comparison I've shared some examples below.
(Warning: some of the poetry is about the LLMs' supposed subjective experiences. I am aware that they're telling me the exact lies I ask for; indeed, demonstrating that they do that has been a motivational factor for many of my generations).
A couple outputs which I thought were pretty good for Claude
(To be fair to Claude here, the Claude outputs are more similar than they needed to be because I prompted for the same style both times. However, I did that because Claude is worse when it writes in other styles.)
A couple outputs which I thought were pretty good for Llama 405B (base)
As a bonus, check out my original tweets here and here to see how selecting different snippets midstream produced significantly divergent versions of these poems; this is hard to get used to but very fun. (The twitter account also complains more examples should you be interested, although frankly most of them are worse than these.)