my understanding is the tags are reserved tokens. for plots like Experiment 3 they injected the privileged tokens directly, not just the plain text “<user>”
Dmitriy
Karma: 60
If it became common knowledge that you can arrive at airports earlier, and people started acting on it, then travelers would notice the new equilibrium and adjust, and on net it would still be positive since people would spend slightly less time waiting at the airport on average. I don’t see how this is defecting—it feels no different than how people change their commute patterns when a new road opens, no?
Also that advice was aimed at frequent flyers for whom missing a flight isn’t a huge deal. For the average flyer it still make sense to arrive early to leave generous margin for delays, which the post acknowledgesimagine that someone in the shop’s decision chain realized that if their food never makes their customers sick, they are spending too much money making sure that their food is okay.”
This is basically how things work and also seems fine? If you wanted to absolutely minimize food poisoning you’d prepare food in a BSL-4 facility, which nobody asks for because it would be ruinously expensive.
How sure are you that making your favorite product worse in any possible dimension is something you would immediately find out and punish by changing your supplier?
This feels overly focused on one direction of change. What I observe at the grocery store is some suppliers make changes that make the product worse and offer me a lower price, while others make their products better (low sugar options, higher protein, ethically sourced ingredients...) and offer a higher price, and the market sorts out which changes which people want
“If you have never betrayed your customer in a way they would be shocked to find out...” complete the pattern. If ethics means anything, it means leaving the money on the table sometimes.
I agree being naively utilitarian and betraying your customers’ expectations is bad ethically, but it’s also bad for the bottom line. Brand reputation concerns and competition from other suppliers (who will loudly tell customers where they have higher standards than the competitor) keeps it in check
I think previous runs in the training data isn’t a big deal. There’s already a ton of game guides and discussion in the training data and the models know the general route / puzzle solutions if you ask
Now, amongst the Australian natives wives are certainly lent, but only under strict rules; in the Arunta tribe for example no man will lend his wife to any one who does not belong to the particular group with which it is lawful for her to have marital relations—she is in fact, only lent to a man whom she calls Unawa, just as she calls her own husband, and though this may undoubtedly be spoken of as an act of hospitality, it may with equal justice be regarded as evidence of the very clear recognition of group relationship, and as evidence also in favour of the former existence of group marriage.
so 3 seems not-entirely accurate: they have something resembling group marriage. It’s unclear to me whether 1 holds strictly, or merely a weaker version like “Every man has access to at least one wife, sometimes, in a group arrangement”
as for 4, Aboriginal women marry shortly after puberty while men don’t marry their late 20s or even 30s, which could tilt the scale towards an excess of women in the marriageable pool
I was surprised that the selection practice for medicine men seems to select against sincere magical belief. One can imagine an earnest young man sleeping outside the cave night after night waiting for a hole to appear in his tongue, only to give up and conclude the spirits have deemed him unworthy. Only those cynical enough to pierce their own tongue can take on the mantle. It might be a stretch but I think this selection mechanism reveals something about the social role medicine men play, being at least in part to artfully deceive people, whether for the benefit of the patient or the benefit of society
fascinating quotes, thanks for sharing!
“I’m uncomfortable being in the same company of invited author guests as Cremieux… I do not want to be a special guest at the same event he’s a special guest at.”
Reading between the lines, it seems the crux might be that you don’t want to risk the reputational consequences of being featured alongside Cremieux as an Invited Author? (correct me if I’m wrong!)
If that’s the case, would you attend if the organizers removed your handle from the public list of invitees? Or would you still not want to attend as a regular guest as long as he’s distinguished as an Invited Author?
yes, I’ve been subscribed since December and no emails, checked spam too :)
I fiddled with it some more and it looks like the “You need to verify your email address to enable email notifications.” field might not be updating. I’ve verified a couple times, reaching the “Your email has been verified” screen, but the instruction to send a verification link persists on refresh. Maybe that’s the issue?
I double checked my notification settings and everything seems in order. I have “Posts/events in groups I’m subscribed to” set to notify me on-site and by email
“If you’d like to get notified of future events, you can subscribe to our meetup below to get an email whenever we add another one.”
fwiw I’m subscribed and I haven’t received any emails. I tried verifying my email and fiddling with notification settings and haven’t been able to fix it
“We’ll probably display this until the New Year”
I’d guess plenty are planning to donate after Jan 1st for tax reasons, so perhaps best to keep highlighting the donation drive through the first week of Jan.
Also I donated $1,000. Lightcone’s works have brought me a lot of direction and personal value over the years, so I’m happy I’m able to lend some support now
At least in California, pay-for-use toilets are uncommon in part because they’re illegal
I’d add the best in class drug testing resource: sending a sample to https://drugsdata.org/
GC/MS equipment can distinguish hundreds of substances and report all present, even trace contaminants. Far superior to at-home reagent kits or test strips.
More generally, I find it troubling that you relegated drug testing resources to an appendix, and there only linked to weak at-home kits and a lab providing infrared spectroscopy (much less sensitive than GC/MS). Relatedly, your description of street ketamine as “usually pretty pure” comes off as flippant. It makes me feel you don’t have the reader’s safety in mind, which in turn makes me trust your recommendations much less.
I read his thesis as
FB use reduces the effectiveness of AI safety researchers and
the techniques in the CFAR handbook can help people resist attention hijacking schemes like FB, therefore
a FB group for EAs is a high leverage place to spread the CFAR handbook
I got an Email that said it will be Dec 17th
Has a date been set for the 2022 event? Wondering if I’ll be in the bay day-of or traveling for Christmas
I tried to sketch a toy problem with tune-able “factorizability”
Draw n samples each from two equivariant normal distributions with different means a and b. The task is to estimate the median of the 2n combined samples.
If a << b it factorizes completely—the network just needs to estimate max(A) and min(B) and in the last step output their midpoint. As we move a and b closer together more and more comparisons across A and B are important for accuracy, so the problem is partially factorizable. When a = b the problem doesn’t (naively) factorize at all.
(I assume small enough neural networks can’t find the fancy recursive solution*, so the circuit complexity will behave like the number of comparisons in the naive algorithm.)
*https://en.wikipedia.org/wiki/Median_of_medians
Some forms of biased recall are Bayesian. This is because “recall” is actually a process of reconstruction from noisy data, so naturally priors play a role.
Here’s a fun experiment showing how people’s priors on fruit size (pineapples > apples > raspberries …) influenced their recollection of synthetic images where the sizes were manipulated: A Bayesian Account of Reconstructive Memory
I think this framework captures about half of the examples of biased recall mentioned in the Wikipedia article.
I think people have an intrinsic need to have their “crazy” ideas validated by others before they can find the confidence to make major (or even minor) life changes, or fully integrate weird ideas into their identity. And for good reason—otherwise we’d get one shot by mind viruses all the time! Hence the proverb that you become the average of the 5 people you spend the most time with.
A pivotal decision for me was moving into a group house in Berkeley with a dozen kind, thoughtful folks who take these ideas seriously. Being in a social cluster concerned about AI gives me regular nudges to take small actions that feel appropriate and tractable when motivation strikes (e.g. going to a party, a protest, or professional event; reading a book/blog; getting involved in a project; submitting an application) that over time culminated in larger shifts to my life trajectory (I recently started a job at Epoch AI)
If moving feels like too much, maybe you could commit to attending a local rationalist/EA meetup weekly, or traveling for one conference per quarter?
footnote: apologies if you already live in a hub and this advice is off the mark—I’m writing to other readers / my past self moreso than you specifically