Trying to become stronger.
hath
Karma: 1,260
Deepmind’s Gopher—more powerful than GPT-3
From Redwood’s application update (rejecting those who didn’t make the cut):
We had many more applicants than I expected, and even though we expanded the program to have space for 30 participants instead of 20, we aren’t able to accept that many of our 500 applicants, including many applicants who seem very promising and competent. I am sad that we don’t have space for more people.
What if we could score on a power-law level of quality, instead of just five stars for fulfilling each of five categories? There could be one order of magnitude for “well written/thought out”, the next order of magnitude higher meaning “has become part of my world model” and one higher of “implemented the recommendations in this post, positively improved my life”. The potential issue I see in the 5 star rating system is that it doesn’t have enough variance; probably 95% of the posts I’ve read on here would be either 4 or 5 stars. Being able to rate posts with a decimal, so you can rate a post 4.5 instead of just 4 or 5, would also help, though it’d clutter the UI and make voting cost more spoons.
A few related thoughts: if we do push farther upon the epistemic hygiene axis, it may be worth writing a Sequence on “Living Epistemically Clean” or something along those lines, so that the standards for discussion are clear and we have a guide to upholding those standards. Specification and implementation, if you will. Such a sequence could potentially also cover, say, implementing TAPs for epistemic hygiene in the outside world. It would probably be useful to have more info on how to address your list of “things your brain tries to do” as a community.
I notice that I value some way of allowing new users to grow accustomed to writing on LW. I imagine that my first post may not be epistemically clean enough to pass the standards you’re pointing towards (which isn’t a point against your standards) and that others new to the site may be discouraged by, say, a ban. The help reviewing drafts that LW currently offers helps, and that’s probably a useful place to focus efforts towards helping people uphold the standards you want to put up. Potentially have multiple levels of users—specific karma amounts or moderator approval would be required to comment in some areas?
Slightly meta: I notice that this post seems to come off as somewhat finger-pointing-towards-moon-like. It works for me, because I see what you’re pointing towards, but it looks like some others don’t, and therefore the post comes off as filled with applause lights and no substance. Or I might just be failing their ITT. I’m not entirely sure how to resolve the missed communication, but I figured it might be worth mentioning.
I’m surprised to see how many people view the Roko’s Basilisk tag. Is that a trend over more than just the last month?
I (weakly) predict that building a shoulder advisor or two out of less-useful-but-more-emulable people might be worth it via giving you the skill of emulating to the available max? Such that, finding emulation in general a little easier and more familiar, you might be able to try again with the actual higher-value targets?
That makes sense, I’ll give it a try.
FWIW, I have been genuinely surprised by advice from shoulder advisors that I could not predict;
Ah, I see.
My own personal experience following this post: I don’t have enough training data for most of the people I’d like to emulate. When I think of the people I know irl that is like to learn from, I’ve spent about ten hours 1-on-1 with each of them; not enough to have a solid mental model of what advice they may give. At the same time, part of why I value their advice is that I can’t predict it; they have wisdom and experience that I don’t. Often, I’ll ask them for advice and be surprised by their answer. When I tried to create a shoulder advisor of one of them, it didn’t work; I just didn’t know enough about them to accurately understand what they were thinking in a certain situation.
Still a great post, though; just didn’t work for a specific use case of mine.
Same thing happened to me. Might’ve been a bug with page loading? I’ve had similar things happen with other sites.
I might try to parallelize it to some degree. It seems like the biomes might be able to be run separately, and the pairings might as well. No promises, though.
From the code, yes.
Invitation to any potential collaborators who want to design species that will be able to survive in equilibrium with each other; PM me for details.
You can vote for a specific tag on a page, and “most relevant” sorts by the post where that tag has been upvoted the most instead of the karma of the posts with the tag.
Interesting. In what ways?
Thank you for pointing that out—should be fixed now.
Self-Responsibility
[APPRENTICE]
Writing PM now.
Fantastic post, as always. Possible typos:
and