https://metaphor.systems has been absolutely incredible for finding new things on the internet. It hands-down beats every other site I’ve tried, including some of the ones you’d think of when I say something like that. I’m going to put some examples in the comments over the next day or so asynchronously, please add your own as well.
please vote up or down the examples you find most useful, and provide an example or two of your own if you come up with anything particularly good. there’s also a discord server where people can share prompts, though those prompts are usually not as effective as mine, and I haven’t been sharing mine there much.
You can also use metaphor to do a similarity search for any url. The results can sometimes be a bit samey, but often it’s good for finding more services like a given service.
my single favorite site I’ve found on metaphor so far is microsolidarity.cc
I used Metaphor to find many entries on the upcoming map of AI safety.
Just for calibration, what are the other things you’ve tried? I’ve tried alternative search engines like https://millionshort.com, link aggregators/curators like reddit/slashdot/hackernews/etc, manually curated lists.
(I’ve been playing around with it for a bit and it seems quite good to me too)
in my workflow, the ranking had been, in order of what I try first to answer a question, text-davinci-2, ddg, kagi, google, millionshort, teclis, with a diversion to youtube if I needed a visual explanation of something. I’d go straight for ddg or google if trying to get a specific result, and that hasn’t changed with introduction of metaphor.systems. text-davinci-3 pushes further up the list; I’m more likely to ask it a question now, but not massively so. meanwhile, metaphor near entirely replaced kagi and teclis and also usually google, I’m back to google for exact-result queries except when google refuses to give multiple word sense interpretations in which case I open ddg, and sometimes I ask text-davinci-3 for help figuring out what I’m looking for. I usually use metaphor first. I have it set as my default search. the different query format it needs sometimes takes multiple tries, but once I find a good search, I get incredible results I’ve never seen another tool match, not even text-davinci-3 or chatgpt.
chatgpt didn’t enter my workflow. the rlhf flavor it has makes it too hard to use and it gives overly verbose responses.
https://metaphor.systems/search?q=this is one of the most cost efficient and effective constructive community focused on the ground organizations I’ve found in any of the leftist%2C anarchist or charity communities worldwide%3A
some results, not closely examined for quality, but all high enough quality to be clearly worth a consideration:
also has interesting results if you substitute “worldwide” for your own location. eg, this same query has interesting results for the bay area in particular. lots of overlap, couple more pages of good results, but here are some non-overlapping results:
I used metaphor to find a bunch of suggestions for how to improve community coprotection in the face of adversarial behavior from community members: https://www.lesswrong.com/posts/WmscKHmyZvuRg48HL/women-and-effective-altruism?commentId=shncEPGboY7s9JXPx
I used metaphor to find the results I gave for this question answer: https://www.lesswrong.com/posts/nZBfpwRnZ63xF67hQ/is-there-some-reason-llms-haven-t-seen-broader-use?commentId=hnP3kyAWrcm3X8Z9K—I don’t remember what query I used there, but paste any link from those into metaphor and you’ll get a really nice similarity search you can use to browse for more.
commentary on a “synthesizing cooperation plans” post, discussing what queries I’d use to find related work (very good post, by the way!) https://www.lesswrong.com/posts/HbPXiB5Aiv3uHPXEh/master-plan-spec-needs-audit-logic-and-cooperative-ai?commentId=9AtA2LkgKzrf4SCqz
another example I gave was in response to a call for reliable link quality tagging; there are many projects to find reliable link tagging. the original comment—some of my queries in that comment are rather long: https://www.lesswrong.com/posts/NLm7QEzFfTRq252My/reestablishing-reliable-sources-a-system-for-tagging-urls?commentId=jN5HJLwcEbFkZf6Be
I used it to make a list a few days ago of existing safety benchmarks, in response to a prompt for such things being posted: https://www.lesswrong.com/posts/KQ6fGiPeMnzzC6p9q/race-to-the-top-benchmarks-for-ai-safety?commentId=A7FMZpfHgjRFHCcCq